WEBVTT 00:00:06.830 --> 00:00:08.000 Welcome, everybody. NOTE CONF {"raw":[100,100]} 00:00:08.120 --> 00:00:11.000 Our speaker tonight is Dr. Vijay Sharma. NOTE CONF {"raw":[100,100,100,100,100,100,100]} 00:00:11.420 --> 00:00:14.840 He is currently an associate professor of medicine and biomedicine NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:00:14.840 --> 00:00:17.150 at the Boston University School of Medicine. NOTE CONF {"raw":[100,100,100,100,100,100,100]} 00:00:17.630 --> 00:00:20.630 He earned his undergraduate degree from the Indian Institute of NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:00:20.630 --> 00:00:23.570 Technology and then pursued a PhD at the University of NOTE CONF {"raw":[100,100,100,100,100,97,100,100,100,100]} 00:00:23.570 --> 00:00:25.850 Southampton in the United Kingdom. NOTE CONF {"raw":[100,100,100,100,100]} 00:00:26.090 --> 00:00:30.380 His current research focuses on neurodegenerative diseases and digital pathology, NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:00:30.380 --> 00:00:32.220 and he's been a collaborator with Unmc. NOTE CONF {"raw":[100,100,100,100,100,100,100]} 00:00:32.240 --> 00:00:34.130 S own Dr. Dan Mirman. NOTE CONF {"raw":[71,100,85,100,52]} 00:00:34.370 --> 00:00:37.520 Tonight, he'll be speaking on Multimodal Machine Learning for Dementia NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:00:37.520 --> 00:00:40.970 assessment, and I'm very excited to hear about this topic. NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:00:41.360 --> 00:00:42.390 Welcome, Dr. Koch. NOTE CONF {"raw":[100,97,33]} 00:00:42.770 --> 00:00:44.240 Thank you so much for joining us. NOTE CONF {"raw":[100,100,100,100,100,100,100]} 00:00:44.360 --> 00:00:44.880 Thank you. NOTE CONF {"raw":[100,100]} 00:00:44.900 --> 00:00:49.700 Thank you for giving me this opportunity to talk about NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:00:49.700 --> 00:00:52.130 some of the work that we've been doing over the NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:00:52.130 --> 00:00:56.900 past few years that kind of related to both deep NOTE CONF {"raw":[100,100,100,75,100,100,100,100,100,100]} 00:00:56.900 --> 00:01:00.740 learning model development, more focused on the methods and they're NOTE CONF {"raw":[100,100,100,100,52,100,100,100,100,32]} 00:01:00.740 --> 00:01:05.330 actually application of that in the context of assessing dementia. NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:01:06.650 --> 00:01:12.050 Special thanks to both Dr. Tarasenko and Dr. Mermin, who NOTE CONF {"raw":[100,100,100,100,98,100,100,100,97,100]} 00:01:12.050 --> 00:01:15.440 are the clinicians at Braska who continue to work with NOTE CONF {"raw":[100,100,100,89,89,100,100,100,100,100]} 00:01:15.440 --> 00:01:21.170 us, and also to Dr. Swaminathan and Dr. Kidder, who NOTE CONF {"raw":[100,100,100,100,99,100,100,99,51,100]} 00:01:21.170 --> 00:01:25.440 were, I think, former members of your center and still NOTE CONF {"raw":[48,100,100,100,100,100,100,100,100,83]} 00:01:25.490 --> 00:01:28.160 tell me that before you talk about science, Vijay, in NOTE CONF {"raw":[100,100,100,100,100,100,100,100,47,97]} 00:01:28.160 --> 00:01:30.470 any of these conferences, talk about your team. NOTE CONF {"raw":[100,100,100,100,100,100,100,100]} 00:01:30.560 --> 00:01:33.140 So I'm kind of following his lead and wanted to NOTE CONF {"raw":[100,100,100,100,100,100,100,100,98,100]} 00:01:33.140 --> 00:01:35.420 sort of talk a little bit about my team here. NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:01:35.420 --> 00:01:38.270 And as a quick introduction, you know, our lab is NOTE CONF {"raw":[100,100,100,100,100,56,56,100,100,100]} 00:01:38.270 --> 00:01:42.020 a mix of both students in the computer science department NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:01:42.020 --> 00:01:43.730 as well as MD students. NOTE CONF {"raw":[100,100,100,62,100]} 00:01:44.630 --> 00:01:48.380 We collaborate very closely with and it's actually really nice NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:01:48.380 --> 00:01:51.080 to see that we are kind of interacting together in NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:01:51.080 --> 00:01:53.330 this day and age because the kind of questions that NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:01:53.330 --> 00:01:58.250 we are attempting to address honestly requires a multidisciplinary team. NOTE CONF {"raw":[100,100,100,100,100,100,100,99,98,100]} 00:01:59.180 --> 00:02:01.280 And I think in this talk you will see that NOTE CONF {"raw":[74,100,100,100,100,100,65,65,100,100]} 00:02:01.280 --> 00:02:02.960 I'm not actually going to spend too much time to NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:02:02.960 --> 00:02:05.240 discuss why I'm doing it, because I think you probably NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:02:05.240 --> 00:02:07.640 know this better, but I would rather want to focus NOTE CONF {"raw":[100,100,100,100,68,100,100,100,100,100]} 00:02:07.640 --> 00:02:10.220 on how we are trying to solve some of the NOTE CONF {"raw":[100,100,81,81,100,100,100,100,100,100]} 00:02:10.220 --> 00:02:13.250 clinically relevant questions using some cutting edge methods. NOTE CONF {"raw":[100,100,100,100,100,100,100,100]} 00:02:13.250 --> 00:02:15.080 And in fact, we are coming up with some novel NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:02:15.080 --> 00:02:18.830 methods related to machine learning applied to mainly looking at NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:02:19.460 --> 00:02:21.160 neurodegenerative diseases, right? NOTE CONF {"raw":[100,100,78]} 00:02:21.170 --> 00:02:23.240 So that's kind of really the talk. NOTE CONF {"raw":[100,100,100,100,100,100,100]} 00:02:24.530 --> 00:02:28.580 And this slide kind of actually summarizes our Labs primary NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:02:28.580 --> 00:02:29.120 goal. NOTE CONF {"raw":[100]} 00:02:30.200 --> 00:02:34.640 Back in 2018, I read this paper published in Lancet NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:02:34.640 --> 00:02:38.000 Neurology that kind of really talked about the supply and NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:02:38.000 --> 00:02:42.080 demand issues related to neurology practices in the US primarily. NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:02:42.950 --> 00:02:46.430 The figure on the right actually shows how apparently fragmented NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:02:46.430 --> 00:02:48.260 the neurology practices are in the country. NOTE CONF {"raw":[100,100,100,100,100,100,100]} 00:02:48.260 --> 00:02:50.150 This was a slightly older version. NOTE CONF {"raw":[100,91,100,100,100,100]} 00:02:50.150 --> 00:02:52.880 I think it's 2019, but still, I think it kind NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:02:52.880 --> 00:02:56.000 of really speaks to the supply and demand issues. NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100]} 00:02:56.870 --> 00:03:00.260 I then actually spoke with a lot of practitioners both NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:03:00.260 --> 00:03:03.620 in Boston and outside, and most of them kind of NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:03:03.620 --> 00:03:04.760 agree to this statement. NOTE CONF {"raw":[94,100,56,100]} 00:03:05.090 --> 00:03:08.930 And I think they kind of talk about the the NOTE CONF {"raw":[100,95,100,100,100,100,98,100,100,100]} 00:03:08.930 --> 00:03:10.650 the need for more neurologists. NOTE CONF {"raw":[100,100,100,100,100]} 00:03:10.670 --> 00:03:13.160 On the other hand, I think people who are also NOTE CONF {"raw":[100,100,100,100,99,100,100,94,99,100]} 00:03:13.160 --> 00:03:15.230 practicing today, they are very, very busy. NOTE CONF {"raw":[100,100,91,91,100,100,100]} 00:03:15.680 --> 00:03:19.490 So so I thought, you know, maybe maybe we should NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:03:19.490 --> 00:03:21.410 kind of really think about what we want to do NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:03:21.410 --> 00:03:22.610 in machine learning in medicine. NOTE CONF {"raw":[100,100,100,100,100]} 00:03:22.610 --> 00:03:25.100 So we began to pursue this development of these novel NOTE CONF {"raw":[100,100,89,100,100,100,100,100,100,100]} 00:03:25.100 --> 00:03:30.410 machine learning frameworks to ultimately build a clinically clinical, great NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,81]} 00:03:30.410 --> 00:03:32.390 tool for dementia assessment. NOTE CONF {"raw":[100,100,100,100]} 00:03:32.390 --> 00:03:34.340 I think the goal is to sort of assist these NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:03:34.340 --> 00:03:38.120 practices and it's an ambitious goal and I think we NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:03:38.120 --> 00:03:39.890 made some important strides. NOTE CONF {"raw":[100,100,100,100]} 00:03:40.070 --> 00:03:42.650 Um, simply put, what we want to do is we NOTE CONF {"raw":[91,100,100,100,100,100,100,100,100,100]} 00:03:42.650 --> 00:03:44.900 want to mimic the neurologist who is seeing a patient NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:03:44.900 --> 00:03:46.190 with memory complaints. NOTE CONF {"raw":[100,100,100]} 00:03:46.700 --> 00:03:48.290 I don't think I have to elaborate too much on NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:03:48.290 --> 00:03:50.750 this point, but briefly, we are talking about a tool NOTE CONF {"raw":[100,100,100,100,62,62,100,100,100,100]} 00:03:50.750 --> 00:03:53.690 that can ingest routinely collected data, and I think that's NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:03:53.690 --> 00:03:54.610 an important point. NOTE CONF {"raw":[100,100,100]} 00:03:54.620 --> 00:04:00.440 Routinely collected clinical data that includes demographics, patient history, neuropsychology NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:04:00.440 --> 00:04:06.530 and neuropsychiatric evaluations, other cognitive tests and imaging also, and NOTE CONF {"raw":[100,100,100,96,100,100,100,100,100,100]} 00:04:06.530 --> 00:04:09.590 kind of provide an assessment of the patient's cognitive status. NOTE CONF {"raw":[100,100,100,100,100,100,100,90,100,100]} 00:04:09.590 --> 00:04:12.200 And so to make sure that our model is doing NOTE CONF {"raw":[48,100,100,100,100,100,100,100,100,100]} 00:04:12.200 --> 00:04:13.280 things in the right way. NOTE CONF {"raw":[100,100,100,100,100]} 00:04:14.430 --> 00:04:16.620 What we also do is we spend a significant amount NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:04:16.620 --> 00:04:20.400 of time to validate our model using the data obtained NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:04:20.400 --> 00:04:24.600 from various cohorts, both clinical cohorts and population cohorts. NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100]} 00:04:25.140 --> 00:04:29.010 We bring experts such as Dr. Mermin and Dr. Tarasenko NOTE CONF {"raw":[100,100,100,100,100,100,60,100,100,100]} 00:04:29.010 --> 00:04:32.010 and perform head to head validation of our models. NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100]} 00:04:32.010 --> 00:04:36.090 And finally, on a few select cases, we also. NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100]} 00:04:36.840 --> 00:04:41.130 We also try and attempt to evaluate if the model NOTE CONF {"raw":[100,100,100,100,69,100,100,100,100,100]} 00:04:41.130 --> 00:04:45.480 is indicating any signs that can be confirmed using neuropathology NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:04:45.540 --> 00:04:46.710 or postmortem evidence. NOTE CONF {"raw":[100,62,98]} 00:04:46.920 --> 00:04:49.320 So that's kind of really one way to really think NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:04:49.320 --> 00:04:51.960 about how to not just think about model development, but NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:04:51.960 --> 00:04:55.710 also seek ways to validate and confirm whatever the model NOTE CONF {"raw":[100,100,100,100,100,100,87,100,100,100]} 00:04:55.710 --> 00:04:57.750 is trying to do and with other evidence that is NOTE CONF {"raw":[100,100,100,100,94,100,100,100,100,100]} 00:04:57.750 --> 00:04:58.890 available in the literature. NOTE CONF {"raw":[100,100,100,100]} 00:04:59.430 --> 00:04:59.820 Right. NOTE CONF {"raw":[64]} 00:04:59.850 --> 00:05:01.920 So I'm going to spend a little bit of time NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:05:01.920 --> 00:05:03.480 to talk about some technical background. NOTE CONF {"raw":[100,100,100,100,100,100]} 00:05:03.510 --> 00:05:05.460 I'm happy to take any questions, but I think it's NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:05:05.460 --> 00:05:06.090 important. NOTE CONF {"raw":[100]} 00:05:06.230 --> 00:05:06.500 Right. NOTE CONF {"raw":[94]} 00:05:06.510 --> 00:05:11.010 So let's say we kind of really broadly, let's say NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:05:11.010 --> 00:05:13.890 we consider the problem of deciding whether a patient has NOTE CONF {"raw":[100,93,100,100,100,100,100,100,100,100]} 00:05:13.980 --> 00:05:15.240 Alzheimer's disease. NOTE CONF {"raw":[100,100]} 00:05:15.780 --> 00:05:19.200 For example, when given a large number of numeric input NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:05:19.200 --> 00:05:23.400 variables, let's say MSC, score, age, demographics, these kinds of NOTE CONF {"raw":[100,100,100,43,100,100,100,100,100,100]} 00:05:23.940 --> 00:05:27.210 values that you routinely collect in a clinic that kind NOTE CONF {"raw":[100,100,100,100,100,100,62,100,100,100]} 00:05:27.210 --> 00:05:29.010 of represent the characteristics of the patient. NOTE CONF {"raw":[100,97,100,100,100,100,100]} 00:05:29.040 --> 00:05:30.960 Of course, there is a broad spectrum of data you NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:05:30.960 --> 00:05:34.260 collect, but just for the sake of discussion, let's assume NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:05:34.260 --> 00:05:38.100 there are some numeric values, numeric input variables that represent NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:05:38.100 --> 00:05:39.540 the characteristics of the patient. NOTE CONF {"raw":[100,100,100,100,100]} 00:05:40.300 --> 00:05:42.490 You know, one standard approach that I think has been NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:05:42.490 --> 00:05:45.910 used is to use something that is very simple logistic NOTE CONF {"raw":[100,100,100,100,83,81,98,100,100,100]} 00:05:45.910 --> 00:05:49.810 regression, that kind of estimates how to weight each of NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:05:49.810 --> 00:05:53.380 those input variables so that they are weighted sum is NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:05:53.380 --> 00:05:55.510 a good indicator of Alzheimer's disease. NOTE CONF {"raw":[100,100,100,100,100,100]} 00:05:55.510 --> 00:05:55.870 Right. NOTE CONF {"raw":[75]} 00:05:56.050 --> 00:05:58.700 But as you all know more than I do, all NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:05:58.990 --> 00:06:02.080 disease is very complex to diagnose and often involves complex NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:06:02.080 --> 00:06:02.920 interactions. NOTE CONF {"raw":[100]} 00:06:03.010 --> 00:06:05.830 So if you want to model this correctly, then we NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:06:05.830 --> 00:06:07.540 can add extra inputs, right? NOTE CONF {"raw":[100,100,100,100,100]} 00:06:07.540 --> 00:06:09.910 So and these extra inputs are what we call as NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:06:09.910 --> 00:06:10.930 interaction terms. NOTE CONF {"raw":[100,100]} 00:06:10.930 --> 00:06:13.930 They each represent the product of two or more input NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:06:13.930 --> 00:06:14.680 variables. NOTE CONF {"raw":[100]} 00:06:15.340 --> 00:06:19.270 But if this multi-way interactions need to be modeled, the NOTE CONF {"raw":[100,100,100,79,100,100,100,100,100,100]} 00:06:19.270 --> 00:06:22.480 number of interaction terms in fact increases exponentially. NOTE CONF {"raw":[100,100,100,100,100,100,100,100]} 00:06:23.170 --> 00:06:25.270 So the more number of features you want to sort NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,92]} 00:06:25.270 --> 00:06:27.850 of include in the model, the number of interaction terms NOTE CONF {"raw":[92,100,100,100,100,100,100,100,100,100]} 00:06:27.850 --> 00:06:29.230 increases exponentially. NOTE CONF {"raw":[100,100]} 00:06:29.290 --> 00:06:33.520 So the neural network alternative is to add a layer NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:06:33.520 --> 00:06:36.300 of these hidden factors or hidden layers. NOTE CONF {"raw":[100,100,100,100,100,100,100]} 00:06:36.310 --> 00:06:39.640 And the first step is to actually determine which hidden NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:06:39.640 --> 00:06:40.870 factors are active. NOTE CONF {"raw":[100,100,100]} 00:06:40.870 --> 00:06:43.420 And then the active ones are used to determine whether NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:06:43.420 --> 00:06:44.250 the disease is present. NOTE CONF {"raw":[100,100,100,100]} 00:06:44.260 --> 00:06:47.950 So this is kind of really the idea behind thinking NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:06:47.950 --> 00:06:51.640 about neural networks to process this kind of multimodal data. NOTE CONF {"raw":[100,100,100,100,100,100,100,100,97,100]} 00:06:51.710 --> 00:06:56.320 I think statistical statisticians and or statistical learning techniques, I NOTE CONF {"raw":[86,100,100,100,100,100,100,100,100,100]} 00:06:56.320 --> 00:06:59.470 think most of you probably are familiar with logistic regression NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:06:59.470 --> 00:07:01.720 and think of this as sort of a more sort NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:07:01.720 --> 00:07:03.730 of an extended version of logistic regression. NOTE CONF {"raw":[100,100,100,100,100,100,100]} 00:07:04.300 --> 00:07:08.140 So now the challenge is to learn a good set NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:07:08.140 --> 00:07:09.070 of hidden factors, right? NOTE CONF {"raw":[100,100,100,98]} 00:07:09.070 --> 00:07:11.260 So how do you sort of learn those hidden factors? NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:07:11.350 --> 00:07:14.050 And what we do is by repeatedly modifying the weights NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:07:14.050 --> 00:07:15.010 on these connections, right? NOTE CONF {"raw":[100,100,100,100]} 00:07:15.010 --> 00:07:17.020 Because like I said, on the left side here, we NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:07:17.020 --> 00:07:19.390 showed this kind of a hidden layer, which means now NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:07:19.390 --> 00:07:21.610 you're able to sort of understand the hidden factors and NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:07:21.610 --> 00:07:22.560 the interactions. NOTE CONF {"raw":[100,100]} 00:07:22.570 --> 00:07:26.080 So now if you extend it further, what we want NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:07:26.080 --> 00:07:28.510 to think about is to trying to understand this kind NOTE CONF {"raw":[100,100,100,100,100,96,100,100,100,100]} 00:07:28.510 --> 00:07:31.570 of a good set of hidden factors by repeatedly modifying NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:07:31.570 --> 00:07:34.750 the weights on those connections from the input variables to NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:07:34.750 --> 00:07:37.390 the hidden factors, and then the weights on connections from NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:07:37.390 --> 00:07:39.160 the hidden factors to the output variables. NOTE CONF {"raw":[100,100,100,100,100,100,52]} 00:07:39.190 --> 00:07:42.130 I know this is a bit of a terminology, but NOTE CONF {"raw":[100,100,100,100,100,100,100,86,100,100]} 00:07:42.130 --> 00:07:44.830 I think if more and more hidden factors are added NOTE CONF {"raw":[100,100,45,100,100,100,100,100,100,100]} 00:07:44.980 --> 00:07:48.160 to account for these complex terms, then obviously, you know, NOTE CONF {"raw":[100,100,100,100,100,100,100,100,80,80]} 00:07:48.190 --> 00:07:51.850 making the modifications on these weights one at a time NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:07:51.850 --> 00:07:53.530 becomes very, very time consuming. NOTE CONF {"raw":[100,100,100,100,100]} 00:07:53.890 --> 00:07:54.160 Right? NOTE CONF {"raw":[58]} 00:07:54.220 --> 00:07:58.840 So so a technique called as Backpropagation was introduced a NOTE CONF {"raw":[100,100,100,100,100,100,62,100,100,100]} 00:07:58.840 --> 00:08:03.100 few years ago that essentially brought over this 1000 fold NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,93]} 00:08:03.100 --> 00:08:06.280 efficiency in terms of addressing this weight modification problem. NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100]} 00:08:06.280 --> 00:08:08.080 Essentially, if you are able to adjust the weights in NOTE CONF {"raw":[100,100,97,100,100,100,100,100,100,100]} 00:08:08.080 --> 00:08:10.600 an efficient way, then you're solving the problem, right? NOTE CONF {"raw":[82,100,100,100,100,100,100,100,100]} 00:08:10.600 --> 00:08:14.050 So that major breakthrough essentially opened up the floodgates and NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:08:14.050 --> 00:08:17.020 people then started building various deep learning approaches and so NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:08:17.020 --> 00:08:18.010 on and so forth. NOTE CONF {"raw":[100,100,100,100]} 00:08:18.250 --> 00:08:18.550 Right. NOTE CONF {"raw":[100]} 00:08:18.550 --> 00:08:21.220 So in summary, I think the the by creating this NOTE CONF {"raw":[100,100,100,100,100,96,100,100,100,100]} 00:08:21.220 --> 00:08:25.000 hierarchical framework, we were able to sort of learn or NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:08:25.000 --> 00:08:28.210 model these several interactions between the inputs and how they NOTE CONF {"raw":[100,100,100,100,100,69,100,100,100,100]} 00:08:28.210 --> 00:08:29.910 relate to an output of interest, right? NOTE CONF {"raw":[100,100,100,100,100,100,78]} 00:08:29.920 --> 00:08:32.740 So like predict if someone has, let's say, Alzheimer's, right? NOTE CONF {"raw":[100,100,93,100,100,100,100,100,96,100]} 00:08:33.039 --> 00:08:38.080 So over the last decade or so, this radically different NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:08:38.080 --> 00:08:42.250 approach to AI, we call this now I it kind NOTE CONF {"raw":[100,100,97,100,100,100,100,55,50,100]} 00:08:42.250 --> 00:08:45.130 of really produced major breakthroughs and now it's kind of NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:08:45.130 --> 00:08:49.030 used on, you know, millions of digital devices for complex NOTE CONF {"raw":[100,100,91,91,100,100,100,100,100,100]} 00:08:49.030 --> 00:08:53.080 tasks such as speech recognition, image interpretation, like you see NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:08:53.080 --> 00:08:55.810 here, and in fact even language translation, right? NOTE CONF {"raw":[100,100,100,100,100,100,100,98]} 00:08:55.810 --> 00:08:59.980 So in the context of images, right, there are certain NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:08:59.980 --> 00:09:02.770 approaches that sort of exploit the the structure of the NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:09:02.770 --> 00:09:03.200 image, right. NOTE CONF {"raw":[100,94]} 00:09:03.220 --> 00:09:05.590 Which is organized in this kind of an XY grid NOTE CONF {"raw":[100,100,93,95,100,100,100,100,98,100]} 00:09:05.590 --> 00:09:06.280 format. NOTE CONF {"raw":[100]} 00:09:07.060 --> 00:09:10.420 Um, and in addition to the hierarchies of or the NOTE CONF {"raw":[100,100,100,100,80,100,94,96,100,100]} 00:09:10.420 --> 00:09:13.690 intermediate layers of the neural network are created by performing NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:09:13.690 --> 00:09:17.380 some operations called as convolutions, which is a term people NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:09:17.380 --> 00:09:20.200 use to sort of really describe this process of simply NOTE CONF {"raw":[94,100,100,100,100,100,100,100,100,100]} 00:09:20.230 --> 00:09:23.260 an arithmetic operation that is done at a very regional NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,97]} 00:09:23.260 --> 00:09:23.710 level. NOTE CONF {"raw":[100]} 00:09:24.580 --> 00:09:26.830 It's kind of what it means is essentially, is that NOTE CONF {"raw":[95,100,100,100,100,100,100,100,100,100]} 00:09:26.830 --> 00:09:28.150 the convolutional operator? NOTE CONF {"raw":[100,93,100]} 00:09:28.420 --> 00:09:30.790 It's kind of a very generic filter that can be NOTE CONF {"raw":[52,100,100,100,100,100,100,100,100,100]} 00:09:30.790 --> 00:09:33.250 applied on these images when you apply these filters. NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100]} 00:09:33.370 --> 00:09:35.620 The convolutional neural network is a sort of a deep NOTE CONF {"raw":[100,64,100,100,100,78,100,100,98,100]} 00:09:35.620 --> 00:09:38.770 learning architecture that can be used or developed when these NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:09:38.770 --> 00:09:40.480 images are often the inputs, right? NOTE CONF {"raw":[100,100,100,100,100,84]} 00:09:40.480 --> 00:09:43.090 So, for example, if I want to learn from a NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:09:43.090 --> 00:09:46.360 brain MRI scans of hundreds of individuals to predict who NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:09:46.360 --> 00:09:49.330 would have signs of atrophy, for instance, in the brain, NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:09:49.330 --> 00:09:52.270 that corresponds to, let's say, Alzheimer's, then what I do NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:09:52.270 --> 00:09:54.850 is I go back to these deep learning frameworks to NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:09:54.850 --> 00:09:58.180 process them, using these convolution approaches where you can hierarchically NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:09:58.180 --> 00:10:02.320 learn the information in those regions and then process that NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:10:02.320 --> 00:10:05.740 information and the volumetric sense to get that output of NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:10:05.740 --> 00:10:06.250 interest, right? NOTE CONF {"raw":[100,96]} 00:10:06.250 --> 00:10:09.430 So that's kind of really the idea behind using these NOTE CONF {"raw":[93,74,100,100,100,100,100,100,100,100]} 00:10:09.430 --> 00:10:12.250 deep learning approaches to look at imaging data as well NOTE CONF {"raw":[100,100,100,100,97,100,100,100,100,100]} 00:10:12.250 --> 00:10:15.250 as, as I previously showed, even on imaging data. NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100]} 00:10:15.980 --> 00:10:16.380 Right. NOTE CONF {"raw":[100]} 00:10:16.400 --> 00:10:19.220 So, you know, most of us actually would prefer to NOTE CONF {"raw":[100,99,99,100,100,100,100,100,100,100]} 00:10:19.220 --> 00:10:22.130 have a very simple explanation for how a neural network NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:10:22.130 --> 00:10:26.120 arrives at its classification, especially for a particular case. NOTE CONF {"raw":[100,97,100,100,100,100,100,100,100]} 00:10:26.660 --> 00:10:29.330 Um, in the in the example of predicting, let's say, NOTE CONF {"raw":[78,100,100,100,100,100,100,100,87,100]} 00:10:29.330 --> 00:10:32.990 whether a person has Alzheimer's or some other etiology, one NOTE CONF {"raw":[100,100,100,100,98,100,100,100,100,100]} 00:10:32.990 --> 00:10:35.060 would like to know what actually are the hidden factors NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:10:35.060 --> 00:10:36.650 they the network is using. NOTE CONF {"raw":[86,100,100,100,100]} 00:10:36.650 --> 00:10:36.830 Right? NOTE CONF {"raw":[69]} 00:10:36.830 --> 00:10:39.440 So but what happens is when you when you apply NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:10:39.440 --> 00:10:42.050 this deep learning network or a deep neural network, when NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:10:42.050 --> 00:10:44.000 it is trained on making predictions on a very large NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:10:44.000 --> 00:10:47.360 data set, it typically uses layers of many, many, many NOTE CONF {"raw":[68,68,100,100,100,100,100,100,100,100]} 00:10:47.360 --> 00:10:50.480 layers of nonlinear features to model a huge number of NOTE CONF {"raw":[100,100,81,100,54,100,100,100,100,100]} 00:10:50.480 --> 00:10:53.480 complicated but kind of weak regularities in the data. NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100]} 00:10:54.750 --> 00:10:58.080 So it's kind of really infeasible to interpret these features NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:10:58.080 --> 00:11:01.860 because they're meaning kind of depends on these complex interactions NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:11:02.610 --> 00:11:05.550 with all these non interpretable features in other layers. NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100]} 00:11:05.550 --> 00:11:07.830 So if the same neural network is to sort of NOTE CONF {"raw":[100,84,100,100,100,100,100,100,100,100]} 00:11:07.830 --> 00:11:10.170 refit the same data, but with changes in the initial NOTE CONF {"raw":[100,100,100,100,100,96,100,100,100,100]} 00:11:10.170 --> 00:11:13.740 random values, there will be different features in intermediate layers, NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:11:13.740 --> 00:11:13.940 right? NOTE CONF {"raw":[88]} 00:11:13.950 --> 00:11:16.530 So in a sense, what the neural network is not NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:11:16.530 --> 00:11:18.600 trying to do is identify the correct hidden factors. NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100]} 00:11:18.600 --> 00:11:21.300 It's kind of merely using these student factors to model NOTE CONF {"raw":[100,100,100,100,100,96,97,100,100,100]} 00:11:21.300 --> 00:11:24.780 the complex relationship between the input variables and the output NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:11:24.780 --> 00:11:25.200 variables. NOTE CONF {"raw":[100]} 00:11:25.200 --> 00:11:25.380 Right? NOTE CONF {"raw":[100]} 00:11:25.380 --> 00:11:29.280 So what our goal is, we are trying to open NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:11:29.280 --> 00:11:30.120 up the black box. NOTE CONF {"raw":[100,100,100,100]} 00:11:30.120 --> 00:11:33.180 So whenever we are trying to use this data, we NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:11:33.180 --> 00:11:35.100 want to make sure we are able to go back NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:11:35.100 --> 00:11:38.640 into those neural networks and identify which factors are actually NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:11:38.640 --> 00:11:40.920 influencing the output of interest, right? NOTE CONF {"raw":[100,100,100,100,100,100]} 00:11:40.920 --> 00:11:44.390 So that's kind of really is the goal because especially NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,97]} 00:11:44.400 --> 00:11:46.770 we are thinking about questions like, you know, can we NOTE CONF {"raw":[100,100,100,100,100,100,78,78,100,100]} 00:11:46.770 --> 00:11:51.060 build models that are interpretable, Can we use these deep NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:11:51.060 --> 00:11:54.570 learning approaches to sort of routinely process clinical data and NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:11:54.660 --> 00:11:58.530 can we build methods that can process all this multimodal NOTE CONF {"raw":[100,100,100,100,100,100,100,100,97,95]} 00:11:58.530 --> 00:12:01.380 data, including imaging, clinical data and so on and so NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:12:01.380 --> 00:12:02.050 forth? NOTE CONF {"raw":[100]} 00:12:02.070 --> 00:12:04.320 Can we make sure these things are, you know, we NOTE CONF {"raw":[100,100,100,100,100,100,100,62,62,100]} 00:12:04.320 --> 00:12:07.080 can go back into those neural networks and explain what's NOTE CONF {"raw":[100,100,100,95,100,100,100,100,100,100]} 00:12:07.080 --> 00:12:08.790 actually going on within those networks? NOTE CONF {"raw":[100,100,100,100,100,100]} 00:12:08.790 --> 00:12:08.940 Right. NOTE CONF {"raw":[97]} 00:12:08.940 --> 00:12:10.590 That's kind of really one of the important things we NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:12:10.590 --> 00:12:11.100 do. NOTE CONF {"raw":[100]} 00:12:12.510 --> 00:12:15.960 So and I think it's important to also realize that NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:12:16.200 --> 00:12:18.090 it's not just about model development. NOTE CONF {"raw":[100,100,100,100,100,100]} 00:12:18.090 --> 00:12:20.850 It's also very, very important to think about how do NOTE CONF {"raw":[100,100,100,100,100,95,100,100,100,100]} 00:12:20.850 --> 00:12:22.290 you validate these methods, right? NOTE CONF {"raw":[100,100,100,100,100]} 00:12:22.290 --> 00:12:24.420 So validation, I think is very, very important. NOTE CONF {"raw":[100,100,100,100,100,100,100,100]} 00:12:24.420 --> 00:12:26.850 And we sort of, you know, seek help from all NOTE CONF {"raw":[100,100,100,100,90,90,100,100,100,100]} 00:12:26.850 --> 00:12:28.980 the clinicians as well as other ways of sort of NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:12:28.980 --> 00:12:30.780 thinking about how to validate these methods. NOTE CONF {"raw":[100,100,100,100,100,100,100]} 00:12:30.780 --> 00:12:30.990 Right. NOTE CONF {"raw":[100]} 00:12:30.990 --> 00:12:31.530 So. NOTE CONF {"raw":[100]} 00:12:33.670 --> 00:12:36.100 So clearly for us to do any of this stuff. NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:12:36.100 --> 00:12:37.150 We need a lot of data. NOTE CONF {"raw":[100,100,100,100,100,100]} 00:12:37.240 --> 00:12:39.360 We need data from many different cohorts. NOTE CONF {"raw":[100,100,100,100,100,100,100]} 00:12:39.370 --> 00:12:42.730 And we've been very fortunate to collaborate with some senior NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:12:42.730 --> 00:12:44.040 investigators at BU. NOTE CONF {"raw":[100,100,100]} 00:12:44.050 --> 00:12:46.270 I think they are part of the Framingham Heart Study, NOTE CONF {"raw":[95,100,100,100,100,100,100,100,100,100]} 00:12:46.270 --> 00:12:50.710 which is, you know, a study that started in 1948 NOTE CONF {"raw":[100,100,81,81,100,100,100,100,100,100]} 00:12:50.980 --> 00:12:52.630 basically focused on the heart. NOTE CONF {"raw":[100,100,100,100,100]} 00:12:52.780 --> 00:12:56.470 But I think it's since 1999, they started collecting also NOTE CONF {"raw":[100,98,100,67,100,100,100,100,100,100]} 00:12:56.470 --> 00:12:59.800 a lot of stuff related to brain health, including imaging, NOTE CONF {"raw":[100,100,100,100,97,100,100,100,100,100]} 00:12:59.800 --> 00:13:03.460 including all the neuropsychological evaluations and everything. NOTE CONF {"raw":[100,100,100,100,100,100,100]} 00:13:03.460 --> 00:13:05.680 There is a dementia review that happens every month on NOTE CONF {"raw":[100,100,100,100,96,100,100,100,100,100]} 00:13:05.680 --> 00:13:09.850 those participants who are basically residents of the town of NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:13:09.850 --> 00:13:12.370 Framingham, which is part of western Massachusetts. NOTE CONF {"raw":[100,100,100,100,100,100,100]} 00:13:12.970 --> 00:13:16.300 And the advantage of that data set is that it's NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:13:16.300 --> 00:13:19.960 actually a community cohort as opposed to a clinical cohort NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:13:19.960 --> 00:13:22.140 that you you would get from, let's say, in a NOTE CONF {"raw":[100,82,100,99,100,100,90,100,98,95]} 00:13:22.150 --> 00:13:23.830 Boston medical center or some other places. NOTE CONF {"raw":[100,100,100,100,100,100,100]} 00:13:23.830 --> 00:13:24.030 Right? NOTE CONF {"raw":[72]} 00:13:24.040 --> 00:13:24.550 So. NOTE CONF {"raw":[98]} 00:13:24.690 --> 00:13:27.550 So we've been very fortunate that we have had access NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:13:27.550 --> 00:13:31.240 to this Framingham study, as well as a few other NOTE CONF {"raw":[100,100,100,100,100,100,100,96,100,100]} 00:13:31.240 --> 00:13:33.040 data sets that I've outlined here. NOTE CONF {"raw":[100,100,100,100,100,100]} 00:13:33.040 --> 00:13:35.420 I think some of you probably know about the ADN NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,39]} 00:13:35.620 --> 00:13:36.190 study. NOTE CONF {"raw":[92]} 00:13:36.370 --> 00:13:39.730 That's also, again, a multi-site study, but again, focused on NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:13:39.730 --> 00:13:40.510 Alzheimer's. NOTE CONF {"raw":[99]} 00:13:41.050 --> 00:13:44.460 There is this National Alzheimer's coordinating center, which is the NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,96]} 00:13:44.470 --> 00:13:47.260 data that is it's sort of really a coordinating center. NOTE CONF {"raw":[100,100,100,64,100,100,100,100,100,100]} 00:13:47.260 --> 00:13:51.620 So it's collect data from 37 or 30 8 or NOTE CONF {"raw":[100,92,100,100,100,95,95,95,100,100]} 00:13:51.620 --> 00:13:54.840 39 Alzheimer's disease research centers around the country. NOTE CONF {"raw":[100,90,100,100,100,100,100,100]} 00:13:54.850 --> 00:13:56.920 So these are, again, clinical cohorts. NOTE CONF {"raw":[100,100,100,100,100,100]} 00:13:56.920 --> 00:13:59.320 But I think these are very useful data set that NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:13:59.320 --> 00:14:00.340 we've been able to access. NOTE CONF {"raw":[100,100,100,100,100]} 00:14:00.340 --> 00:14:02.920 Some of them are publicly available, so we have been NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:14:02.920 --> 00:14:05.350 able to access all this data because we want to NOTE CONF {"raw":[100,100,100,100,97,100,100,100,100,100]} 00:14:05.350 --> 00:14:08.830 make sure whatever models we construct are generalizable. NOTE CONF {"raw":[100,100,100,100,100,100,100,100]} 00:14:09.040 --> 00:14:12.040 So we don't want to make sure we've overfit our NOTE CONF {"raw":[100,100,100,100,100,100,100,93,100,100]} 00:14:12.040 --> 00:14:14.410 models on one cohort, but also make sure they work NOTE CONF {"raw":[100,100,100,100,100,100,100,100,95,91]} 00:14:14.410 --> 00:14:16.030 on other cohorts as well. NOTE CONF {"raw":[100,100,100,100,100]} 00:14:16.030 --> 00:14:16.450 Right? NOTE CONF {"raw":[100]} 00:14:16.870 --> 00:14:21.040 So given all this data collected from multiple cohorts, and NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:14:21.070 --> 00:14:23.560 again, because we were fortunate to have some computer scientists NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,93]} 00:14:23.560 --> 00:14:26.200 work with us, one of the first things that we NOTE CONF {"raw":[97,100,100,100,100,100,100,100,100,100]} 00:14:26.200 --> 00:14:30.130 started doing was to actually develop computational frameworks that allowed NOTE CONF {"raw":[100,100,100,100,100,100,96,100,100,100]} 00:14:30.130 --> 00:14:31.950 us to automatically process the data. NOTE CONF {"raw":[100,100,100,100,100,100]} 00:14:31.960 --> 00:14:33.220 This is the biggest challenge, right? NOTE CONF {"raw":[100,100,100,100,100,100]} 00:14:33.220 --> 00:14:35.530 Like you can see on the right side, we don't NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:14:35.530 --> 00:14:37.080 do machine learning every day. NOTE CONF {"raw":[100,100,100,99,99]} 00:14:37.090 --> 00:14:40.720 We actually spend about 75% of our time to process NOTE CONF {"raw":[100,100,96,100,100,100,100,100,100,100]} 00:14:40.720 --> 00:14:41.410 all this data. NOTE CONF {"raw":[100,98,100]} 00:14:41.860 --> 00:14:44.230 Actually, when it comes to imaging, things become a lot NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:14:44.230 --> 00:14:45.160 more tedious, right? NOTE CONF {"raw":[100,100,61]} 00:14:45.190 --> 00:14:46.630 So you need to do a lot of things to NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:14:46.630 --> 00:14:49.390 preprocess the data, such as register all these images to NOTE CONF {"raw":[57,100,100,100,100,92,100,100,100,100]} 00:14:49.390 --> 00:14:53.170 a predefined template, you know, perform simple quality checks to NOTE CONF {"raw":[100,84,100,81,81,100,100,100,100,100]} 00:14:53.170 --> 00:14:55.530 just to make sure the scans have the data that NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,93]} 00:14:55.540 --> 00:14:59.050 without artifacts, for instance, and if you find artifacts or NOTE CONF {"raw":[100,100,100,100,100,100,52,100,100,100]} 00:14:59.050 --> 00:15:01.960 some other simply unwanted elements on those images, then we NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,97]} 00:15:01.960 --> 00:15:04.120 have to figure out a way to remove them and NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:15:04.120 --> 00:15:07.000 maybe select certain regions of interest within the scans that NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:15:07.000 --> 00:15:11.290 we want to focus in some specific analysis, if you NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:15:11.290 --> 00:15:13.570 want to process data from multiple cohorts, then we need NOTE CONF {"raw":[100,100,100,100,100,100,100,100,72,100]} 00:15:13.570 --> 00:15:15.820 to also figure out a way to normalize or harmonize NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:15:15.820 --> 00:15:16.240 the data. NOTE CONF {"raw":[97,100]} 00:15:16.630 --> 00:15:18.520 So I think that's a very important element, right? NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100]} 00:15:18.520 --> 00:15:20.530 So these aspects are not trivial. NOTE CONF {"raw":[100,100,100,100,100,100]} 00:15:21.010 --> 00:15:23.410 And also when we are dealing with different imaging data NOTE CONF {"raw":[62,100,100,99,100,100,100,100,100,100]} 00:15:23.410 --> 00:15:25.870 such as PET scans and MRI scans, then we need NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:15:25.870 --> 00:15:28.570 to also make sure we think about correcting the elements NOTE CONF {"raw":[100,100,100,100,100,100,100,95,100,100]} 00:15:28.570 --> 00:15:32.770 and correcting the bias between those different different imaging modalities. NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:15:33.910 --> 00:15:36.100 To be honest, you know, these things may not sound NOTE CONF {"raw":[100,100,100,85,85,100,100,100,100,100]} 00:15:36.100 --> 00:15:36.730 important. NOTE CONF {"raw":[100]} 00:15:36.730 --> 00:15:38.800 That important, at least if you are dealing with, let's NOTE CONF {"raw":[100,100,100,100,100,88,88,100,100,100]} 00:15:38.800 --> 00:15:41.170 say, 50 or, you know, 100 cases, you know, you NOTE CONF {"raw":[100,100,100,71,71,100,100,81,81,100]} 00:15:41.170 --> 00:15:42.760 can just get away with it. NOTE CONF {"raw":[100,100,100,100,100,100]} 00:15:42.760 --> 00:15:45.940 But if you're dealing with 50,000 cases, then we need NOTE CONF {"raw":[100,100,92,100,100,100,100,100,100,100]} 00:15:45.940 --> 00:15:48.370 to start paying attention to these aspects more carefully. NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100]} 00:15:48.640 --> 00:15:50.620 So so in our lab, we spend like, like I NOTE CONF {"raw":[100,100,100,100,100,100,98,91,100,100]} 00:15:50.620 --> 00:15:53.740 said, or 75% of our time working on all these NOTE CONF {"raw":[100,65,100,100,100,100,100,95,100,100]} 00:15:53.740 --> 00:15:57.070 image processing routines, which are, you know, significant, at least NOTE CONF {"raw":[100,100,100,100,100,90,90,100,100,100]} 00:15:57.070 --> 00:15:59.800 obviously important for us because you're dealing with a very NOTE CONF {"raw":[100,100,100,100,100,69,100,100,100,100]} 00:15:59.800 --> 00:16:01.030 large data set, right? NOTE CONF {"raw":[100,90,90,100]} 00:16:01.030 --> 00:16:03.910 And then finally, we think about machine learning as such. NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:16:04.650 --> 00:16:05.030 Right. NOTE CONF {"raw":[100]} 00:16:05.040 --> 00:16:08.940 So again, if we have created a very big database NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:16:08.940 --> 00:16:10.380 of registered MRI scans. NOTE CONF {"raw":[100,100,93,81]} 00:16:10.380 --> 00:16:12.480 So here what you're seeing here on the top is NOTE CONF {"raw":[100,92,100,100,100,100,100,100,100,100]} 00:16:12.480 --> 00:16:17.160 this a simple example of a slice by slide registration NOTE CONF {"raw":[100,100,100,100,100,100,100,100,58,100]} 00:16:17.160 --> 00:16:20.340 of the MRI to a specific template that we got NOTE CONF {"raw":[100,100,54,100,100,100,100,100,100,100]} 00:16:20.340 --> 00:16:23.610 from McGill University from in Canada. NOTE CONF {"raw":[100,95,100,100,100,100]} 00:16:24.450 --> 00:16:27.720 So there is a tool called Flirt, which is publicly NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:16:27.720 --> 00:16:30.390 available within a package called as FSL. NOTE CONF {"raw":[100,100,100,100,100,100,100]} 00:16:31.110 --> 00:16:34.410 That FSL package was, I think from Oxford University. NOTE CONF {"raw":[96,96,86,34,100,100,100,100,100]} 00:16:34.410 --> 00:16:36.450 And what they allow us to do is they allow NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:16:36.450 --> 00:16:38.910 us to align these scans with respect to a template. NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:16:39.120 --> 00:16:40.980 So one thing we realize is that none of these NOTE CONF {"raw":[100,100,100,100,96,100,100,100,100,100]} 00:16:40.980 --> 00:16:44.890 automated registration techniques actually work for all cases, right? NOTE CONF {"raw":[100,100,100,100,100,100,100,100,87]} 00:16:44.910 --> 00:16:47.940 So especially if you're looking at a community cohort data, NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:16:48.330 --> 00:16:49.460 it's very messy. NOTE CONF {"raw":[100,100,100]} 00:16:49.470 --> 00:16:52.260 So we have to somehow and clinical scans especially are NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:16:52.260 --> 00:16:53.240 also very messy, right? NOTE CONF {"raw":[100,100,100,87]} 00:16:53.250 --> 00:16:57.240 So we have to perform lots of steps to make NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:16:57.240 --> 00:16:59.490 sure all these things are registered in the right way. NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:16:59.490 --> 00:17:01.620 You need to figure out what landmarks exist in these NOTE CONF {"raw":[78,100,100,100,100,100,100,100,100,100]} 00:17:01.620 --> 00:17:03.990 images and then make sure these things are automated at NOTE CONF {"raw":[100,86,100,100,100,100,100,100,100,100]} 00:17:03.990 --> 00:17:04.620 the same time. NOTE CONF {"raw":[100,100,100]} 00:17:04.620 --> 00:17:06.360 Also quality checked over time. NOTE CONF {"raw":[100,100,100,100,100]} 00:17:07.199 --> 00:17:10.829 Um, what we also do is, you know, what is NOTE CONF {"raw":[54,100,100,100,100,100,82,82,100,54]} 00:17:10.829 --> 00:17:11.430 called stripping. NOTE CONF {"raw":[100,100]} 00:17:11.430 --> 00:17:14.040 It's a term people use to just remove the, the NOTE CONF {"raw":[100,99,95,100,96,100,100,100,100,100]} 00:17:14.040 --> 00:17:16.920 bright spot around the brain because we think that sort NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:17:16.920 --> 00:17:21.030 of interferes with the signal to noise of the images. NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:17:21.030 --> 00:17:23.819 So we don't feel like it's sort of really adding NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:17:23.819 --> 00:17:25.860 any value to our assessment. NOTE CONF {"raw":[100,100,100,100,100]} 00:17:25.860 --> 00:17:28.800 So we sort of do these operations a priori to NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:17:28.800 --> 00:17:31.320 remove those regions on those images. NOTE CONF {"raw":[100,100,100,100,100,100]} 00:17:31.320 --> 00:17:34.860 And we do that slice by slice all three plains NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,72]} 00:17:34.860 --> 00:17:37.950 on hundreds of them and on all those 50,000 cases. NOTE CONF {"raw":[73,100,100,100,100,100,100,100,100,100]} 00:17:38.100 --> 00:17:41.340 So that's the significant amount of time that we spend. NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,68]} 00:17:41.340 --> 00:17:43.020 And I think it's valuable at the end of the NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:17:43.020 --> 00:17:43.500 day. NOTE CONF {"raw":[100]} 00:17:43.680 --> 00:17:46.740 Like I said, we have now this database of registered NOTE CONF {"raw":[100,96,100,100,100,100,100,100,100,100]} 00:17:46.740 --> 00:17:50.310 segmented MRI scans, which would be happy to share if NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,95]} 00:17:50.310 --> 00:17:52.920 anybody else is interested because we have them. NOTE CONF {"raw":[100,99,99,100,100,100,100,91]} 00:17:53.810 --> 00:17:55.400 We have done all the work right. NOTE CONF {"raw":[99,99,100,100,100,100,70]} 00:17:56.660 --> 00:17:59.570 And finally, after all this, we were able to, you NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:17:59.570 --> 00:18:01.100 know, luckily publish some work. NOTE CONF {"raw":[100,79,93,100,100]} 00:18:01.100 --> 00:18:03.230 And these are a couple of papers that were recently NOTE CONF {"raw":[100,100,100,85,100,100,100,100,100,100]} 00:18:03.230 --> 00:18:03.800 published. NOTE CONF {"raw":[100]} 00:18:03.800 --> 00:18:06.440 You know, the one on the top was kind of NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:18:06.440 --> 00:18:11.000 primarily focused on a novel method that was interpretable, and NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:18:11.000 --> 00:18:14.030 the other one focused on creating a more robust framework NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:18:14.030 --> 00:18:17.030 for sort of combining all these different modalities of data. NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:18:17.630 --> 00:18:18.020 Right. NOTE CONF {"raw":[94]} 00:18:18.950 --> 00:18:21.800 Just to expand further, you know, the model that we NOTE CONF {"raw":[79,100,100,100,90,90,100,100,100,100]} 00:18:22.010 --> 00:18:24.800 created, I think we use one cohort, which is this NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:18:24.800 --> 00:18:29.240 NACC, the National Alzheimer's Coordinating Center data to train the NOTE CONF {"raw":[52,100,100,100,100,94,100,100,100,100]} 00:18:29.240 --> 00:18:29.780 model. NOTE CONF {"raw":[100]} 00:18:30.470 --> 00:18:33.530 And then once we did the training, we took a NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:18:33.530 --> 00:18:36.380 lot of time and process data from other cohorts to NOTE CONF {"raw":[100,100,100,100,90,100,100,100,100,100]} 00:18:36.380 --> 00:18:38.750 make sure that the model that is trained on one NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:18:38.750 --> 00:18:42.170 cohort actually did well on the other cohorts independently. NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100]} 00:18:42.170 --> 00:18:42.370 Right? NOTE CONF {"raw":[78]} 00:18:42.380 --> 00:18:45.860 So the question that we asked was very simple, which NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:18:45.860 --> 00:18:49.280 is given all this multimodal data, can the model place NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,96]} 00:18:49.280 --> 00:18:51.350 the person along the dementia spectrum? NOTE CONF {"raw":[100,100,100,100,100,100]} 00:18:51.350 --> 00:18:51.540 Right. NOTE CONF {"raw":[84]} 00:18:51.560 --> 00:18:54.260 So that is, can we first predict if the person NOTE CONF {"raw":[100,100,100,100,97,100,100,100,100,100]} 00:18:54.260 --> 00:18:58.670 has healthy cognition, MCI or mild cognitive impairment or dementia, NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:18:59.390 --> 00:19:01.850 and if the model predicts the person to have dementia? NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:19:02.180 --> 00:19:04.610 Can you further understand if the dementia was due to NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:19:04.610 --> 00:19:07.010 Alzheimer's disease or due to other etiologies? NOTE CONF {"raw":[100,100,100,100,100,100,100]} 00:19:07.250 --> 00:19:11.300 So while we combine all the non Alzheimer's etiologies in NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:19:11.300 --> 00:19:15.050 one group, this was just the first study and we NOTE CONF {"raw":[100,100,100,85,100,100,100,100,100,100]} 00:19:15.050 --> 00:19:17.840 kind of expanded since then to perform a much more NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:19:17.840 --> 00:19:21.410 broader task of, you know, differential diagnosis of dementia. NOTE CONF {"raw":[100,100,100,87,87,100,100,100,100]} 00:19:21.420 --> 00:19:23.540 That's something that we are very excited about. NOTE CONF {"raw":[100,100,100,100,100,100,100,100]} 00:19:24.110 --> 00:19:28.300 Um, to again, to process this multimodal data. NOTE CONF {"raw":[76,100,100,100,100,100,97,100]} 00:19:28.310 --> 00:19:32.000 We kind of created this two tier neural network architecture NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:19:32.000 --> 00:19:35.140 because imaging imaging data is kind of a different format. NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:19:35.150 --> 00:19:39.470 All the other non imaging information, including demographics, patient history, NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:19:39.470 --> 00:19:41.180 functional assessments, neuropsychology. NOTE CONF {"raw":[100,100,100]} 00:19:41.480 --> 00:19:43.780 So all these things can be tabulated in a in NOTE CONF {"raw":[89,100,100,100,100,100,100,100,100,100]} 00:19:43.780 --> 00:19:45.050 a nice table, right? NOTE CONF {"raw":[100,100,100,100]} 00:19:45.050 --> 00:19:47.630 So, so we want to make sure the images are NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:19:47.630 --> 00:19:48.500 processed correctly. NOTE CONF {"raw":[100,100]} 00:19:48.500 --> 00:19:51.650 And then we also want to combine that information with NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:19:51.650 --> 00:19:53.150 the other data sets. NOTE CONF {"raw":[100,100,87,87]} 00:19:53.190 --> 00:19:53.450 Right. NOTE CONF {"raw":[100]} 00:19:53.450 --> 00:19:56.540 So the first is basically this convolutional neural network that NOTE CONF {"raw":[100,100,100,100,100,100,95,100,100,100]} 00:19:56.540 --> 00:20:00.950 I talked earlier, which allows us to process the volumetric NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:20:00.950 --> 00:20:01.580 images. NOTE CONF {"raw":[97]} 00:20:02.570 --> 00:20:03.710 With each voxel. NOTE CONF {"raw":[90,100,100]} 00:20:03.950 --> 00:20:06.380 It's actually we process each box at things at the NOTE CONF {"raw":[98,100,80,90,100,100,86,100,100,100]} 00:20:06.380 --> 00:20:10.220 voxel level, which voxel in the MRI is processed in NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:20:10.220 --> 00:20:14.000 a very hierarchical manner and the information learned from the NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:20:14.000 --> 00:20:18.080 entire brain MRI is then carried forward in the form NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:20:18.080 --> 00:20:22.790 of a very simplified vector representation, which is then combined NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:20:22.940 --> 00:20:26.240 with the non imaging data to then assess the person's NOTE CONF {"raw":[100,100,71,100,100,100,100,100,100,100]} 00:20:26.240 --> 00:20:26.880 cognitive state. NOTE CONF {"raw":[100,70]} 00:20:26.900 --> 00:20:28.970 So this is kind of really the essence of this NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:20:28.970 --> 00:20:30.380 multimodal deep learning, right? NOTE CONF {"raw":[100,100,100,97]} 00:20:30.380 --> 00:20:33.800 So we need to appreciate what kind of modalities, modalities NOTE CONF {"raw":[100,100,100,100,100,100,100,100,94,100]} 00:20:33.800 --> 00:20:36.320 of data exist in that multimodal framework. NOTE CONF {"raw":[100,100,100,100,100,100,100]} 00:20:36.320 --> 00:20:39.350 And we need to sort of create frameworks that can NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:20:39.350 --> 00:20:40.490 process those data. NOTE CONF {"raw":[100,100,100]} 00:20:40.490 --> 00:20:42.620 And then at the same time, also at the end NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:20:42.770 --> 00:20:45.320 figured out another way to combine all that information, right? NOTE CONF {"raw":[91,100,100,100,100,100,100,100,100,97]} 00:20:45.320 --> 00:20:47.210 So I think that's the key point here. NOTE CONF {"raw":[100,100,100,100,100,100,100,100]} 00:20:47.750 --> 00:20:51.440 Uh, and the advantage of doing all this is at NOTE CONF {"raw":[58,100,100,100,100,100,100,100,100,100]} 00:20:51.440 --> 00:20:53.960 least one major advantage that I think is we are NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:20:53.960 --> 00:20:56.930 now able to create these interpretable results, right? NOTE CONF {"raw":[100,100,100,100,100,100,100,98]} 00:20:56.930 --> 00:20:59.660 So we can in fact go back through those neural NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:20:59.660 --> 00:21:04.340 networks to identify features or factors that are most predictive NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:21:04.340 --> 00:21:05.450 of the output of interest. NOTE CONF {"raw":[100,100,100,100,100]} 00:21:05.450 --> 00:21:07.640 So what you can see here on this slide is NOTE CONF {"raw":[94,100,100,100,100,100,100,72,100,100]} 00:21:07.640 --> 00:21:09.290 on the left side is the first task that I NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:21:09.290 --> 00:21:12.500 described, which is, you know, given all this multimodal information, NOTE CONF {"raw":[100,100,100,88,88,100,100,100,100,100]} 00:21:12.500 --> 00:21:16.280 can you assess whether the person has healthy cognition, MCI NOTE CONF {"raw":[100,100,100,100,100,100,100,65,100,100]} 00:21:16.280 --> 00:21:17.030 or dementia? NOTE CONF {"raw":[100,100]} 00:21:17.450 --> 00:21:20.900 And then if you for that specific task, you actually NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:21:20.900 --> 00:21:25.010 are seeing now this list of features that the model NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:21:25.010 --> 00:21:27.410 thought were important in terms of making an assessment. NOTE CONF {"raw":[100,100,100,100,100,100,100,61,100]} 00:21:27.410 --> 00:21:30.020 And clearly it turned out to be number one for NOTE CONF {"raw":[100,100,92,100,100,100,100,100,100,100]} 00:21:30.050 --> 00:21:30.530 that case. NOTE CONF {"raw":[100,100]} 00:21:30.530 --> 00:21:32.180 It's not surprise to this audience. NOTE CONF {"raw":[100,100,100,100,100,100]} 00:21:32.750 --> 00:21:36.440 But on the second question, which is once the model NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:21:36.440 --> 00:21:40.400 identifies the person to have dementia, can you then go NOTE CONF {"raw":[100,100,100,73,99,100,100,100,100,100]} 00:21:40.400 --> 00:21:43.340 in to see if the dementia was due to Alzheimer's NOTE CONF {"raw":[85,85,100,100,100,100,100,100,100,100]} 00:21:43.340 --> 00:21:45.050 disease or due to some other etiologies? NOTE CONF {"raw":[100,100,100,100,100,100,100]} 00:21:45.050 --> 00:21:48.770 And it interestingly, it turned out for that specific question, NOTE CONF {"raw":[100,83,100,100,100,100,100,100,100,100]} 00:21:48.770 --> 00:21:52.040 the MRI turned out to be the most important factor, NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:21:52.640 --> 00:21:55.970 which was which allowed for those persons to already who NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:21:55.970 --> 00:21:59.570 are already having dementia to differentiate between them, between Alzheimer's NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:21:59.570 --> 00:22:00.770 and non Alzheimer's etiology. NOTE CONF {"raw":[100,98,100,100]} 00:22:00.770 --> 00:22:00.940 Right. NOTE CONF {"raw":[89]} 00:22:00.950 --> 00:22:03.200 So that I think was an interesting finding. NOTE CONF {"raw":[100,100,67,67,100,100,100,100]} 00:22:03.890 --> 00:22:07.130 Uh, and now with this kind of framework, again, one NOTE CONF {"raw":[87,98,56,100,100,100,100,100,100,100]} 00:22:07.130 --> 00:22:09.500 can also visualize these high risk regions, right? NOTE CONF {"raw":[100,100,100,100,100,100,99,100]} 00:22:09.500 --> 00:22:12.200 So and this was actually an innovation that was appreciated NOTE CONF {"raw":[100,100,100,56,100,100,100,100,100,100]} 00:22:12.200 --> 00:22:13.250 by the reviewers. NOTE CONF {"raw":[100,100,100]} 00:22:13.880 --> 00:22:17.510 Um, from the computation standpoint, this framework can work very NOTE CONF {"raw":[92,100,96,100,100,100,100,100,100,100]} 00:22:17.510 --> 00:22:20.900 efficiently because, you know, it can process these volumetric scans NOTE CONF {"raw":[97,100,80,80,100,100,100,100,100,100]} 00:22:20.900 --> 00:22:23.600 very quickly as the model was trained to infer these NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:22:23.600 --> 00:22:28.250 local patterns of the cerebral structures that suggested the overall NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:22:28.460 --> 00:22:29.420 disease state. NOTE CONF {"raw":[100,100]} 00:22:29.420 --> 00:22:32.420 So and you can then lay these inferences directly on NOTE CONF {"raw":[100,98,58,100,100,100,100,100,100,100]} 00:22:32.420 --> 00:22:35.810 the MRIs so that one can easily observe which regions NOTE CONF {"raw":[96,90,100,100,100,100,100,100,100,100]} 00:22:35.810 --> 00:22:38.990 are actually hot or cold on the MRIs where these NOTE CONF {"raw":[100,100,100,100,100,100,100,100,98,100]} 00:22:38.990 --> 00:22:42.500 hot voxels actually indicate higher probability of disease risk. NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100]} 00:22:42.680 --> 00:22:44.930 And that was, I think, a very interesting way to NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:22:44.930 --> 00:22:48.140 interpret how the model is trying to learn from those NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:22:48.510 --> 00:22:48.980 images. NOTE CONF {"raw":[81]} 00:22:48.980 --> 00:22:49.400 Right. NOTE CONF {"raw":[100]} 00:22:49.940 --> 00:22:53.270 Um, what we then did was we basically identified a NOTE CONF {"raw":[81,100,100,100,100,100,100,100,100,100]} 00:22:53.270 --> 00:22:57.450 few cases where some neuropathology reports were available, right? NOTE CONF {"raw":[100,100,100,100,100,100,100,100,73]} 00:22:57.500 --> 00:23:00.080 So and I think we got data from three different NOTE CONF {"raw":[100,100,87,100,100,100,100,100,100,100]} 00:23:00.080 --> 00:23:03.800 cohorts and, and these reports were pretty detailed. NOTE CONF {"raw":[100,100,100,100,100,100,100,100]} 00:23:03.890 --> 00:23:07.940 They included region specific as well as global scores like NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:23:07.940 --> 00:23:11.330 ABC, the amyloid beta phase, as well as the block NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,44]} 00:23:11.360 --> 00:23:12.890 staging, as well as see that. NOTE CONF {"raw":[100,100,100,100,63,88]} 00:23:13.760 --> 00:23:16.250 And I think all these scores of different proteins is NOTE CONF {"raw":[100,79,100,100,100,90,100,100,89,100]} 00:23:16.250 --> 00:23:18.080 indicative of Alzheimer's. NOTE CONF {"raw":[100,100,100]} 00:23:18.320 --> 00:23:21.200 And as you can see from all the three plots, NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:23:21.350 --> 00:23:23.660 uh, what you're seeing on the y axis is this NOTE CONF {"raw":[85,100,100,100,100,100,100,100,100,100]} 00:23:23.660 --> 00:23:25.370 model prediction of disease risk. NOTE CONF {"raw":[100,100,100,100,100]} 00:23:25.640 --> 00:23:28.880 And statistically it turned out to be as the model NOTE CONF {"raw":[100,100,83,100,100,100,100,100,100,100]} 00:23:28.880 --> 00:23:30.680 was able to sort of predict the disease risk. NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100]} 00:23:30.680 --> 00:23:33.140 And at the same time, the disease risk was proportional NOTE CONF {"raw":[100,93,100,100,100,100,100,100,100,100]} 00:23:33.140 --> 00:23:35.870 to the degree of severity of pathology that you had NOTE CONF {"raw":[100,100,100,100,100,100,100,100,87,83]} 00:23:35.870 --> 00:23:39.380 observed on these cases, which I think was very, very NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:23:39.590 --> 00:23:42.020 informative for us because I think this was very important. NOTE CONF {"raw":[100,100,100,100,84,100,100,100,100,100]} 00:23:42.170 --> 00:23:44.870 So essentially what this model is trying to do is NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:23:44.870 --> 00:23:47.120 it gave us confidence, give us confidence that the model NOTE CONF {"raw":[100,100,100,100,95,100,100,100,100,100]} 00:23:47.120 --> 00:23:50.180 was not just predicting an output, but also provided a NOTE CONF {"raw":[100,100,100,100,96,100,100,100,100,100]} 00:23:50.180 --> 00:23:54.110 score which sort of increased proportionally with the severity of NOTE CONF {"raw":[100,100,100,100,100,95,100,100,100,100]} 00:23:54.110 --> 00:23:57.410 the disease, which I think was very, very important for NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:23:57.410 --> 00:23:57.980 us to do. NOTE CONF {"raw":[100,100,100]} 00:23:59.120 --> 00:24:02.540 And then after we did this kind of a global NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:24:02.810 --> 00:24:06.020 association, we we then created visualizations. NOTE CONF {"raw":[100,100,100,100,100,100]} 00:24:06.230 --> 00:24:08.900 So we wanted to make sure whatever these model is NOTE CONF {"raw":[100,100,100,100,100,100,100,58,100,100]} 00:24:08.900 --> 00:24:11.840 trying to do, Can you sort of combine those interpretations NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:24:11.840 --> 00:24:13.760 with the regions where you see pathology? NOTE CONF {"raw":[100,100,100,100,100,100,96]} 00:24:13.760 --> 00:24:16.520 So what we did was we just simply use these NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:24:16.790 --> 00:24:20.270 created these heatmaps derived from the model and overlaid them NOTE CONF {"raw":[65,100,100,100,100,100,100,100,100,100]} 00:24:20.270 --> 00:24:23.750 on the regions where you actually where we had the NOTE CONF {"raw":[100,100,100,100,100,100,100,100,72,100]} 00:24:24.680 --> 00:24:28.820 neuropathology scores available on those different regions. NOTE CONF {"raw":[100,100,100,100,100,100,100]} 00:24:29.330 --> 00:24:32.240 So, so what you're seeing on the slide is basically NOTE CONF {"raw":[100,100,100,99,100,100,80,100,100,100]} 00:24:32.240 --> 00:24:33.500 the overlay of. NOTE CONF {"raw":[100,100,100]} 00:24:34.670 --> 00:24:38.750 The the model predictive regions of high disease risk with NOTE CONF {"raw":[100,100,100,54,100,100,100,100,100,100]} 00:24:38.750 --> 00:24:42.710 postmortem findings of disease pathology on a single person. NOTE CONF {"raw":[94,100,100,98,100,100,100,100,100]} 00:24:43.760 --> 00:24:48.080 And this this person actually had clinically confirmed Alzheimer's disease, NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:24:48.080 --> 00:24:51.470 which affected basically, I think, regions which are including the, NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:24:51.470 --> 00:24:55.940 you know, the the bilateral asymmetrical temporal lobes and the NOTE CONF {"raw":[94,94,100,100,100,100,100,100,100,100]} 00:24:55.940 --> 00:24:59.390 right side hippocampus, the cingulate cortex is the other region NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:24:59.390 --> 00:25:02.270 in the corpus callosum and part of the parietal lobe NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:25:02.270 --> 00:25:04.220 and the frontal lobe in this case. NOTE CONF {"raw":[100,100,100,100,100,100,100]} 00:25:05.150 --> 00:25:07.850 And then the first column actually just shows the slices NOTE CONF {"raw":[100,91,100,100,100,100,100,100,100,100]} 00:25:07.850 --> 00:25:12.140 in all the three different planes, followed by the column. NOTE CONF {"raw":[86,100,100,100,100,100,100,100,100,100]} 00:25:12.140 --> 00:25:16.880 Second column basically shows, um, you know, the model predicted NOTE CONF {"raw":[100,100,100,100,88,100,100,100,100,100]} 00:25:16.880 --> 00:25:18.380 disease risk maps. NOTE CONF {"raw":[100,100,100]} 00:25:19.100 --> 00:25:21.470 And then we just simply created a cutoff value to NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:25:21.470 --> 00:25:23.060 show the regions of high risk. NOTE CONF {"raw":[100,100,100,100,100,100]} 00:25:23.060 --> 00:25:25.580 And we overlap that with the MRI scan in the NOTE CONF {"raw":[100,100,60,100,100,100,100,100,100,100]} 00:25:25.580 --> 00:25:26.210 next column. NOTE CONF {"raw":[100,100]} 00:25:26.980 --> 00:25:29.980 And then the final the fourth column basically shows those NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:25:29.980 --> 00:25:33.940 segmented regions within that brain, especially the cortical and the NOTE CONF {"raw":[100,100,100,79,100,100,100,100,100,100]} 00:25:33.940 --> 00:25:35.560 subcortical structures of the brain. NOTE CONF {"raw":[100,100,100,100,100]} 00:25:35.560 --> 00:25:39.400 So this we obtain from, I think maybe folks who NOTE CONF {"raw":[100,100,100,96,100,100,100,100,100,100]} 00:25:39.400 --> 00:25:41.950 are familiar with free server, which is actually a tool NOTE CONF {"raw":[100,100,100,91,52,100,100,100,100,100]} 00:25:41.950 --> 00:25:44.110 that is used to process these scans. NOTE CONF {"raw":[100,100,100,100,100,100,100]} 00:25:44.110 --> 00:25:46.030 So we used free software to sort of segment all NOTE CONF {"raw":[100,100,87,100,74,100,100,100,100,100]} 00:25:46.030 --> 00:25:47.200 these different regions. NOTE CONF {"raw":[100,100,100]} 00:25:47.290 --> 00:25:49.870 And then finally we sort of really overlaid the MRI NOTE CONF {"raw":[84,100,100,100,100,100,100,85,100,100]} 00:25:49.870 --> 00:25:53.170 scan, the disease probability maps of high risk and the NOTE CONF {"raw":[100,100,100,100,100,100,94,94,100,100]} 00:25:53.170 --> 00:25:56.560 color coded regions to just show that spatially, at least NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:25:56.560 --> 00:25:58.720 the model was able to not only just predict who NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:25:58.720 --> 00:26:02.200 has the disease, but also identified regions of high risk. NOTE CONF {"raw":[100,100,100,100,100,82,100,100,100,100]} 00:26:02.200 --> 00:26:05.140 And that kind of corresponded to the regions of neuropathology NOTE CONF {"raw":[100,100,100,100,58,100,100,100,100,100]} 00:26:05.530 --> 00:26:08.290 that I think was very useful for us because this NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:26:08.290 --> 00:26:10.870 kind of really served as a very important mode of NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:26:10.870 --> 00:26:14.020 validation of model of the models that we created. NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100]} 00:26:15.200 --> 00:26:15.470 Right. NOTE CONF {"raw":[100]} 00:26:15.920 --> 00:26:19.190 Um, and then finally, icing on the cake. NOTE CONF {"raw":[100,100,100,100,100,100,100,100]} 00:26:19.460 --> 00:26:23.900 This is where wonderful neurologists and neuro radiologist collaborated with NOTE CONF {"raw":[100,100,100,100,93,100,100,81,100,100]} 00:26:23.900 --> 00:26:25.280 us over the years. NOTE CONF {"raw":[100,100,100,100]} 00:26:26.060 --> 00:26:29.540 So basically we teamed up with various practicing neurologists, some NOTE CONF {"raw":[100,100,100,100,100,100,100,98,100,100]} 00:26:29.540 --> 00:26:33.170 of them behavioral neurologists, some of them general neurologists, and NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:26:33.170 --> 00:26:36.350 also some of them who are neuro radiologists from many NOTE CONF {"raw":[100,100,100,100,89,92,100,79,100,100]} 00:26:36.350 --> 00:26:41.810 different places, including Nebraska Medical Center, Boston Medical Center. NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100]} 00:26:41.810 --> 00:26:43.760 And in fact, some of them are international. NOTE CONF {"raw":[100,100,100,100,100,100,100,100]} 00:26:44.220 --> 00:26:45.900 They are practicing physicians. NOTE CONF {"raw":[58,58,100,100]} 00:26:45.920 --> 00:26:49.910 So what we did was we basically gave them some NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:26:49.910 --> 00:26:53.030 randomly selected cases, about 100 of them, and then we NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:26:53.030 --> 00:26:56.120 asked them to independently provide an impression on which cases NOTE CONF {"raw":[97,100,100,100,100,100,100,100,100,100]} 00:26:56.120 --> 00:27:01.100 had healthy cognition MCI, Alzheimer's, dementia, and also even on NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,72]} 00:27:01.100 --> 00:27:02.120 Alzheimer's dementia. NOTE CONF {"raw":[100,100]} 00:27:03.310 --> 00:27:06.390 Um, interestingly, the task was slightly different for the radiologists. NOTE CONF {"raw":[58,100,100,100,100,100,100,98,100,70]} 00:27:06.410 --> 00:27:08.600 They didn't want to predict who had MCI. NOTE CONF {"raw":[100,100,100,100,100,100,100,100]} 00:27:08.720 --> 00:27:11.810 They hated that their goal was, you know what, give NOTE CONF {"raw":[100,100,100,100,100,100,100,100,91,56]} 00:27:11.840 --> 00:27:14.180 me give me an MRI and I'm going to tell NOTE CONF {"raw":[56,84,100,100,100,100,100,100,100,100]} 00:27:14.180 --> 00:27:16.190 you what kind of regions are affected in the brain, NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:27:16.190 --> 00:27:16.330 Right. NOTE CONF {"raw":[71]} 00:27:16.340 --> 00:27:17.660 So which I think is fair. NOTE CONF {"raw":[97,100,99,100,100,100]} 00:27:17.880 --> 00:27:18.200 Right. NOTE CONF {"raw":[97]} 00:27:18.200 --> 00:27:20.990 So the question was slightly different for the radiologist versus NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:27:20.990 --> 00:27:21.920 the neurologist. NOTE CONF {"raw":[100,100]} 00:27:22.570 --> 00:27:26.660 Um, and, and we then and I think it's important NOTE CONF {"raw":[80,100,100,100,100,100,100,100,100,100]} 00:27:26.660 --> 00:27:30.560 to note that the information that we provided to the NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:27:30.560 --> 00:27:34.190 neurologists was exactly the same information that was fed into NOTE CONF {"raw":[96,100,100,100,100,100,100,100,100,100]} 00:27:34.190 --> 00:27:34.760 the model. NOTE CONF {"raw":[100,100]} 00:27:35.120 --> 00:27:37.040 So exactly the same information. NOTE CONF {"raw":[100,100,100,100,100]} 00:27:37.700 --> 00:27:41.060 Um, and then we, we basically got their assessments. NOTE CONF {"raw":[100,100,100,97,100,100,56,100,100]} 00:27:41.060 --> 00:27:43.850 And I think as you can see here, the overall, NOTE CONF {"raw":[100,96,100,100,100,100,100,100,100,100]} 00:27:43.850 --> 00:27:47.300 I think the, the, the assessment of who has healthy NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:27:47.300 --> 00:27:51.440 cognition, MCI dementia was fairly done by the by the NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:27:51.440 --> 00:27:55.790 neurologist and favorably also the model also was favorably in NOTE CONF {"raw":[96,100,80,100,100,100,100,100,98,100]} 00:27:55.790 --> 00:27:57.200 terms of their assessments. NOTE CONF {"raw":[100,100,100,100]} 00:27:57.870 --> 00:28:00.990 But in terms of the new radiologists trying to assess NOTE CONF {"raw":[100,100,100,100,100,39,97,100,100,100]} 00:28:00.990 --> 00:28:06.060 their task of understanding the differences between Alzheimer's dementia versus NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:28:06.060 --> 00:28:07.980 not on symbols, that was, I think, a very tricky NOTE CONF {"raw":[94,94,73,91,81,100,100,100,100,100]} 00:28:07.980 --> 00:28:11.100 task because most of them, I think, were saying that NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:28:11.100 --> 00:28:14.490 it's very hard for anybody to sort of concede those NOTE CONF {"raw":[100,100,100,100,100,100,100,100,50,100]} 00:28:14.490 --> 00:28:17.580 two because most of them have or may have mixed NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:28:17.580 --> 00:28:18.270 etiologies. NOTE CONF {"raw":[96]} 00:28:18.420 --> 00:28:20.400 So the thing that was kind of a learning experience NOTE CONF {"raw":[100,84,100,100,100,100,100,100,100,100]} 00:28:20.400 --> 00:28:24.600 for us because, you know, things that probably not that NOTE CONF {"raw":[100,100,100,100,100,100,46,100,100,100]} 00:28:24.600 --> 00:28:26.510 kind of a simple binary question, right? NOTE CONF {"raw":[100,100,100,100,100,100,76]} 00:28:26.520 --> 00:28:28.260 So maybe there is always that kind of a mixed NOTE CONF {"raw":[100,100,88,88,100,100,100,100,100,100]} 00:28:28.260 --> 00:28:30.180 etiology that's happening in real world. NOTE CONF {"raw":[100,100,100,100,100,100]} 00:28:30.480 --> 00:28:33.450 So that was kind of a very interesting and learning NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:28:33.450 --> 00:28:33.930 experience. NOTE CONF {"raw":[100]} 00:28:33.930 --> 00:28:35.940 But still, I think we were able to make that NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:28:35.940 --> 00:28:38.940 progress of making sure that the model has some kind NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:28:38.940 --> 00:28:43.110 of a clinical or an expert level validation done both NOTE CONF {"raw":[100,100,100,100,52,100,100,100,100,100]} 00:28:43.110 --> 00:28:45.720 from the standpoint of a neurology practice as well as NOTE CONF {"raw":[100,100,100,100,52,100,100,100,100,100]} 00:28:45.720 --> 00:28:49.050 from the standpoint of a neuro radiology practice. NOTE CONF {"raw":[100,100,100,100,68,100,100,100]} 00:28:49.050 --> 00:28:52.040 I think these are two different practices that use all NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:28:52.050 --> 00:28:54.150 this kind of data, and I think that was very NOTE CONF {"raw":[99,100,100,100,100,100,100,100,100,100]} 00:28:54.150 --> 00:28:55.350 important for us to do. NOTE CONF {"raw":[100,100,100,100,100]} 00:28:56.040 --> 00:28:58.590 And what I think what we are trying to do NOTE CONF {"raw":[98,100,100,100,100,100,100,100,100,100]} 00:28:58.590 --> 00:29:01.680 is making sure that these tools, I think, can go NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:29:01.680 --> 00:29:02.730 to the next level. NOTE CONF {"raw":[100,100,100,100]} 00:29:02.910 --> 00:29:05.760 So I think we just don't want to continue doing NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:29:05.760 --> 00:29:06.300 research. NOTE CONF {"raw":[100]} 00:29:06.300 --> 00:29:09.210 What we began doing since then is we began to NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:29:09.210 --> 00:29:14.100 create this these kinds of software that allows one to NOTE CONF {"raw":[100,94,100,100,100,100,100,100,100,100]} 00:29:14.730 --> 00:29:16.230 process all this information, right? NOTE CONF {"raw":[100,100,100,100,72]} 00:29:16.230 --> 00:29:19.290 So so we have built these tools and we are NOTE CONF {"raw":[100,100,100,72,91,100,100,100,100,100]} 00:29:19.290 --> 00:29:20.610 making them available. NOTE CONF {"raw":[100,100,100]} 00:29:20.700 --> 00:29:22.260 And I think that's going to be really the next NOTE CONF {"raw":[100,100,100,100,85,85,74,98,100,100]} 00:29:22.260 --> 00:29:24.120 step for us to think about, which is, you know, NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:29:24.120 --> 00:29:28.170 can can can we sort of create assistive tools that NOTE CONF {"raw":[100,100,100,97,100,100,100,100,100,100]} 00:29:28.170 --> 00:29:31.680 can help, you know, some practices, maybe not all the NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:29:31.680 --> 00:29:33.790 practices, maybe some practices? NOTE CONF {"raw":[100,100,100,100]} 00:29:33.810 --> 00:29:35.820 I think that's kind of really where we are getting NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:29:35.820 --> 00:29:36.750 at this point. NOTE CONF {"raw":[100,100,100]} 00:29:37.080 --> 00:29:41.220 So I really want to take this opportunity to actually NOTE CONF {"raw":[97,100,100,100,100,100,100,100,100,100]} 00:29:41.220 --> 00:29:44.460 have a discussion and maybe seek your help and understand NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:29:44.460 --> 00:29:46.920 where things stand at this point. NOTE CONF {"raw":[100,100,100,100,100,100]} 00:29:46.920 --> 00:29:48.780 And if there are if there is actually a need NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:29:48.780 --> 00:29:49.980 for such kind of a tool. NOTE CONF {"raw":[100,100,100,100,100,100]} 00:29:49.980 --> 00:29:53.280 So so that's all I got in terms of my NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:29:53.280 --> 00:29:57.600 slides right and can quickly summarize what I basically have NOTE CONF {"raw":[100,90,100,100,100,100,100,100,100,100]} 00:29:57.600 --> 00:30:00.450 described today, which is, you know, all these multimodal deep NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:30:00.450 --> 00:30:04.350 learning frameworks can be used and they can allow us NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:30:04.350 --> 00:30:07.890 to process routinely collected data that can be useful for NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:30:07.890 --> 00:30:08.970 dementia assessment. NOTE CONF {"raw":[100,100]} 00:30:09.910 --> 00:30:12.790 You know, especially those people who are interested in deep NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:30:12.790 --> 00:30:13.690 learning applications. NOTE CONF {"raw":[100,100]} 00:30:13.760 --> 00:30:15.700 I think one thing which I would like to say NOTE CONF {"raw":[74,100,100,100,100,100,100,100,100,100]} 00:30:15.700 --> 00:30:18.910 is that this is only possible because, you know, we NOTE CONF {"raw":[100,100,100,100,100,100,100,87,87,100]} 00:30:18.910 --> 00:30:22.630 as computer scientists were working with health care providers. NOTE CONF {"raw":[100,100,100,100,100,100,73,73,100]} 00:30:22.670 --> 00:30:24.460 I think those are the ones who are teaching us NOTE CONF {"raw":[87,100,100,100,100,100,100,100,100,100]} 00:30:24.460 --> 00:30:26.770 what to do, what not to do and what kind NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:30:26.770 --> 00:30:30.010 of questions are clinically relevant and what kind of data NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:30:30.010 --> 00:30:32.870 is actually supposed to be used to build such models. NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:30:32.890 --> 00:30:34.510 I think that was very, very important. NOTE CONF {"raw":[97,100,100,100,100,100,100]} 00:30:35.170 --> 00:30:38.650 Um, and what we think is that we need to NOTE CONF {"raw":[86,100,100,100,100,100,100,100,100,100]} 00:30:38.650 --> 00:30:41.530 focus on mainly the methods development, but at the same NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:30:41.530 --> 00:30:44.080 time, any method that we develop, we want to make NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:30:44.080 --> 00:30:47.050 sure it's comprehensively validated in the best way we can. NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:30:47.380 --> 00:30:52.210 That could be clinical clinicians, help from the clinicians, neuropathology NOTE CONF {"raw":[100,100,100,100,100,98,100,100,100,100]} 00:30:52.210 --> 00:30:55.480 or data coming from different cohorts and making sure that NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:30:55.480 --> 00:30:58.420 models that are trained on one cohort actually work on NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:30:58.420 --> 00:30:59.380 the other cohort as well. NOTE CONF {"raw":[100,100,100,100,100]} 00:30:59.380 --> 00:30:59.830 Right. NOTE CONF {"raw":[100]} 00:31:00.270 --> 00:31:02.860 Um, but I think fundamentally we believe in this point, NOTE CONF {"raw":[96,100,96,100,100,100,100,100,100,100]} 00:31:02.860 --> 00:31:04.960 which is, you know, the promise of AI, not just NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:31:04.960 --> 00:31:05.740 in medical imaging. NOTE CONF {"raw":[100,100,100]} 00:31:05.740 --> 00:31:08.140 I think I just wrote medical imaging, but generally in NOTE CONF {"raw":[100,100,87,100,100,100,100,100,100,100]} 00:31:08.140 --> 00:31:12.100 terms of thinking about clinical data only. NOTE CONF {"raw":[100,100,100,100,100,100,100]} 00:31:12.100 --> 00:31:15.190 That can happen if clinicians or healthcare providers actually begin NOTE CONF {"raw":[100,100,100,100,100,100,63,100,100,100]} 00:31:15.190 --> 00:31:16.300 to adapt these tools, right? NOTE CONF {"raw":[100,64,100,100,90]} 00:31:16.300 --> 00:31:18.940 So so we are in that process of making or NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:31:18.940 --> 00:31:22.480 at least learning from our side as to how there NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:31:22.480 --> 00:31:25.810 is any value proposition for an AI based tool to NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:31:25.810 --> 00:31:28.990 actually sit within a clinical setting, whether it's coming in NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:31:28.990 --> 00:31:32.200 from a general urology practice or a or a much NOTE CONF {"raw":[100,100,100,100,100,80,67,97,96,100]} 00:31:32.200 --> 00:31:35.620 more specialized practice such as a behavioral neurology setting, or NOTE CONF {"raw":[100,93,100,100,100,97,92,95,100,100]} 00:31:35.620 --> 00:31:38.110 even much more upstream like a primary care setting, right? NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,87]} 00:31:38.110 --> 00:31:40.510 So so we are trying to understand where these tools NOTE CONF {"raw":[91,100,100,100,100,100,100,100,100,100]} 00:31:40.510 --> 00:31:41.800 can actually fit. NOTE CONF {"raw":[100,100,100]} 00:31:42.280 --> 00:31:45.760 And for that we need your help, right? NOTE CONF {"raw":[100,100,100,100,100,100,100,100]} 00:31:45.760 --> 00:31:48.010 So we want to use you want you to use NOTE CONF {"raw":[100,100,100,100,100,63,100,100,100,100]} 00:31:48.010 --> 00:31:50.350 this tool if you if you're interested. NOTE CONF {"raw":[100,100,100,100,100,98,100]} 00:31:50.500 --> 00:31:53.020 So we already created kind of this web based framework, NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:31:53.020 --> 00:31:55.540 which means you can simply log in to your. NOTE CONF {"raw":[100,100,100,100,100,100,96,96,100]} 00:31:56.440 --> 00:31:59.020 Like the way you log in to any email account, NOTE CONF {"raw":[100,100,100,100,100,70,70,100,100,100]} 00:31:59.020 --> 00:32:01.930 you can just log in and then you can simply NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:32:02.200 --> 00:32:05.890 upload information on this to sort of get your assessments NOTE CONF {"raw":[100,100,100,100,96,100,100,100,100,100]} 00:32:05.890 --> 00:32:06.400 on this. NOTE CONF {"raw":[100,100]} 00:32:06.430 --> 00:32:08.470 I mean, it's at this point the research tool. NOTE CONF {"raw":[98,100,100,100,100,100,100,100,100]} 00:32:08.470 --> 00:32:12.670 But we also feel like we we began some discussions NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:32:12.670 --> 00:32:15.550 with the FDA to see if this tool actually has NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:32:15.550 --> 00:32:19.150 any deep needs to go through certain kinds of approvals. NOTE CONF {"raw":[100,45,100,100,100,100,100,100,100,100]} 00:32:19.150 --> 00:32:22.300 And we realize that as long as this serves as NOTE CONF {"raw":[100,100,94,100,100,100,100,100,100,100]} 00:32:22.300 --> 00:32:26.620 an assistive tool in the practice, the regulatory burden is NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:32:26.620 --> 00:32:29.380 actually much less and we are working on that. NOTE CONF {"raw":[100,100,100,100,95,95,100,100,100]} 00:32:30.310 --> 00:32:33.610 But I think I'm going to want to hear your NOTE CONF {"raw":[100,97,100,98,98,100,100,100,100,100]} 00:32:33.610 --> 00:32:36.190 thoughts and see if this can be useful in your NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:32:36.190 --> 00:32:38.140 practice and within your environment. NOTE CONF {"raw":[100,100,100,100,100]} 00:32:38.440 --> 00:32:41.320 And with that, I'll stop sharing. NOTE CONF {"raw":[100,100,100,100,100,100]} 00:32:41.320 --> 00:32:43.270 And these are just my running agency. NOTE CONF {"raw":[100,100,100,100,100,75,97]} 00:32:43.270 --> 00:32:44.220 So thank you. NOTE CONF {"raw":[100,100,100]} 00:32:44.230 --> 00:32:45.100 Thank you for your time. NOTE CONF {"raw":[100,100,100,100,100]} 00:33:00.430 --> 00:33:01.840 See, Justin, you have your hand up. NOTE CONF {"raw":[100,100,100,100,100,100,100]} 00:33:06.830 --> 00:33:08.120 No, I'm an educator. NOTE CONF {"raw":[98,100,100,100]} 00:33:08.120 --> 00:33:09.050 That was applause. NOTE CONF {"raw":[100,100,100]} 00:33:12.060 --> 00:33:12.960 Thank you. NOTE CONF {"raw":[100,100]} 00:33:14.820 --> 00:33:16.900 This, this John Wendel. NOTE CONF {"raw":[100,100,98,38]} 00:33:17.370 --> 00:33:20.310 I don't think we have any neurologist. NOTE CONF {"raw":[100,100,100,100,100,100,80]} 00:33:20.310 --> 00:33:23.400 I know that David Ellis does a lot of work NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:33:23.400 --> 00:33:26.730 on the imaging side and neuroscience side. NOTE CONF {"raw":[100,100,100,100,100,100,100]} 00:33:27.090 --> 00:33:29.910 I'm sure Dr. Mermin is still seeing patients. NOTE CONF {"raw":[100,100,95,74,100,100,100,100]} 00:33:30.690 --> 00:33:34.860 This is very impressive, especially when we see the area NOTE CONF {"raw":[100,96,100,100,100,100,100,100,100,100]} 00:33:34.860 --> 00:33:37.230 under the curve so tightly. NOTE CONF {"raw":[100,100,100,100,100]} 00:33:37.980 --> 00:33:40.740 I want to see if David had any thoughts from NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:33:40.740 --> 00:33:42.420 the methods perspective. NOTE CONF {"raw":[100,98,100]} 00:33:45.430 --> 00:33:45.970 Thank you. NOTE CONF {"raw":[100,100]} 00:33:48.240 --> 00:33:48.780 Hi. NOTE CONF {"raw":[98]} 00:33:49.250 --> 00:33:49.520 Yeah. NOTE CONF {"raw":[50]} 00:33:50.550 --> 00:33:51.270 Had a few questions. NOTE CONF {"raw":[100,100,100,100]} 00:33:51.270 --> 00:33:52.500 Like how did you. NOTE CONF {"raw":[98,100,100,100]} 00:33:53.440 --> 00:33:53.800 I don't know. NOTE CONF {"raw":[89,100,100]} 00:33:54.100 --> 00:33:56.620 There was an overview graph of how you like you NOTE CONF {"raw":[93,70,55,62,100,100,100,100,100,100]} 00:33:56.620 --> 00:33:59.650 integrated the imaging data with the. NOTE CONF {"raw":[100,100,100,100,100,100]} 00:34:01.280 --> 00:34:02.960 I guess not imaging data essentially. NOTE CONF {"raw":[86,100,100,100,100,100]} 00:34:02.960 --> 00:34:05.480 And so how do you have any more details about NOTE CONF {"raw":[90,100,100,100,100,100,100,100,100,100]} 00:34:05.480 --> 00:34:07.610 how you integrated those two? NOTE CONF {"raw":[100,100,100,100,94]} 00:34:09.280 --> 00:34:09.669 Yeah. NOTE CONF {"raw":[100]} 00:34:09.669 --> 00:34:10.330 So. NOTE CONF {"raw":[100]} 00:34:11.230 --> 00:34:15.220 Um, so clearly imaging data has to be processed differently. NOTE CONF {"raw":[55,100,100,100,100,100,100,100,100,100]} 00:34:15.370 --> 00:34:20.649 So yeah, we have leveraged some of these convolutional approaches NOTE CONF {"raw":[100,92,100,100,100,100,100,100,100,100]} 00:34:20.649 --> 00:34:24.909 to process the MRI scans in these 3D volumetric. NOTE CONF {"raw":[100,100,100,91,100,100,100,92,100]} 00:34:24.909 --> 00:34:27.669 So we basically have these DICOM plans or the nifty NOTE CONF {"raw":[100,100,100,100,100,98,60,100,100,100]} 00:34:27.669 --> 00:34:31.690 files that we got from all these cohorts and they NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:34:31.690 --> 00:34:34.929 are converted to these Python arrays, these three dimensional python NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:34:34.929 --> 00:34:36.879 arrays in the NumPy formats. NOTE CONF {"raw":[100,100,100,100,87]} 00:34:37.389 --> 00:34:39.610 And then we apply these convolutional approaches. NOTE CONF {"raw":[100,100,98,88,100,94,100]} 00:34:39.610 --> 00:34:41.590 And again, when I say convolution, it's a very generic NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:34:41.590 --> 00:34:44.050 term these days because there are so many specific ways NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:34:44.050 --> 00:34:45.760 of thinking about convolutions. NOTE CONF {"raw":[100,100,100,100]} 00:34:45.760 --> 00:34:48.820 So we applied a novel method which is based on NOTE CONF {"raw":[100,100,100,87,100,100,100,100,100,100]} 00:34:48.820 --> 00:34:51.639 our previous work, which is thinking about this something called NOTE CONF {"raw":[100,100,100,100,100,100,100,85,100,100]} 00:34:51.639 --> 00:34:54.580 as a fully convolutional approach, which allows us to sort NOTE CONF {"raw":[87,100,100,100,100,100,100,100,100,100]} 00:34:54.580 --> 00:34:58.270 of process volumetric scans in a way without having to NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:34:58.270 --> 00:35:02.260 think about, um, the location information. NOTE CONF {"raw":[100,100,75,100,100,100]} 00:35:02.260 --> 00:35:05.560 So there is some kind of a understanding that's happening NOTE CONF {"raw":[100,100,100,100,100,100,48,100,100,100]} 00:35:05.560 --> 00:35:09.190 or learning that's happening as you process these volumetric images. NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:35:10.400 --> 00:35:15.950 And so 3D MRI gets processed using these convolutions and NOTE CONF {"raw":[100,100,92,97,100,100,100,100,100,100]} 00:35:15.950 --> 00:35:20.030 finally you end up with a single one dimensional vector. NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:35:20.270 --> 00:35:23.840 And that vector is essentially the juice that comes out NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:35:23.840 --> 00:35:25.850 of this 3D volume. NOTE CONF {"raw":[100,100,88,100]} 00:35:25.940 --> 00:35:29.630 And that can be then integrate that or concatenate that NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:35:29.630 --> 00:35:34.910 information with the non imaging data and then run through NOTE CONF {"raw":[100,100,100,88,100,100,100,100,100,100]} 00:35:34.910 --> 00:35:35.750 the neural network. NOTE CONF {"raw":[100,100,100]} 00:35:35.960 --> 00:35:39.290 So it's basically one neural network, but sort of CNN, NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:35:39.290 --> 00:35:41.650 which is the first part combined with something called as NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,94]} 00:35:41.660 --> 00:35:45.290 a multi-layer perceptron, which is then used to together learn NOTE CONF {"raw":[100,68,100,100,100,100,100,100,100,100]} 00:35:45.440 --> 00:35:46.730 and make a prediction. NOTE CONF {"raw":[100,100,100,100]} 00:35:46.730 --> 00:35:49.040 So the back propagation that I was talking about earlier NOTE CONF {"raw":[100,100,95,95,94,100,100,100,100,100]} 00:35:49.040 --> 00:35:51.380 is actually happening on both at the same time. NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100]} 00:35:52.950 --> 00:35:54.990 And so do you just have one output for the NOTE CONF {"raw":[91,100,90,100,100,100,100,100,100,100]} 00:35:54.990 --> 00:35:55.350 network? NOTE CONF {"raw":[100]} 00:35:55.350 --> 00:35:59.370 Then after concatenating both and predicting your output, you just NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:35:59.370 --> 00:36:00.300 have one output. NOTE CONF {"raw":[100,100,100]} 00:36:01.560 --> 00:36:03.570 It's a mutual learning that's happening. NOTE CONF {"raw":[99,99,100,100,100,100]} 00:36:03.860 --> 00:36:07.050 And so just to clarify one output in the sense NOTE CONF {"raw":[85,100,100,100,100,100,100,100,100,100]} 00:36:07.050 --> 00:36:08.820 that we kind of have this kind of a two NOTE CONF {"raw":[100,100,100,100,100,96,100,100,100,100]} 00:36:08.850 --> 00:36:09.810 tier prediction, right? NOTE CONF {"raw":[52,100,97]} 00:36:09.810 --> 00:36:13.740 So we don't want to predict if somebody has non NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:36:13.740 --> 00:36:16.140 Alzheimer's dementia if the person has healthy cognition. NOTE CONF {"raw":[100,100,100,100,100,100,100,100]} 00:36:16.140 --> 00:36:16.350 Right. NOTE CONF {"raw":[100]} 00:36:16.350 --> 00:36:19.710 So so the first tier of prediction is healthy cognition, NOTE CONF {"raw":[77,100,100,100,60,100,100,100,100,100]} 00:36:19.710 --> 00:36:21.360 MCI dementia. NOTE CONF {"raw":[100,100]} 00:36:21.390 --> 00:36:24.060 And if the model has high probability that it's actually NOTE CONF {"raw":[100,100,100,100,100,100,100,100,95,100]} 00:36:24.060 --> 00:36:26.610 dementia, then it does the second task. NOTE CONF {"raw":[100,100,100,100,100,100,100]} 00:36:26.790 --> 00:36:27.930 Otherwise it doesn't, Right? NOTE CONF {"raw":[100,100,100,82]} 00:36:27.930 --> 00:36:30.330 So so we made that kind of a rule to NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:36:30.330 --> 00:36:32.010 make sure the model is doing it in the right NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:36:32.010 --> 00:36:32.310 way. NOTE CONF {"raw":[100]} 00:36:34.190 --> 00:36:34.760 Thanks. NOTE CONF {"raw":[100]} 00:36:35.390 --> 00:36:37.040 This is John Wendel again. NOTE CONF {"raw":[75,65,100,78,100]} 00:36:37.520 --> 00:36:40.190 This kind of led to a follow up question and NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:36:40.490 --> 00:36:44.000 may expose a fair amount of my ignorance. NOTE CONF {"raw":[100,100,100,100,100,100,100,100]} 00:36:44.210 --> 00:36:47.540 But one of the things that we've had discussions here NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:36:47.540 --> 00:36:52.150 is image resolution and how much does that matter? NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100]} 00:36:52.160 --> 00:36:56.870 And so I saw you using 256 by 256. NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100]} 00:36:58.070 --> 00:37:01.400 We've had discussions with other groups who want to use NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:37:01.400 --> 00:37:04.250 JPEG versus DICOM images. NOTE CONF {"raw":[100,100,100,100]} 00:37:05.720 --> 00:37:10.340 How do you you see that there is a sweet NOTE CONF {"raw":[100,100,99,100,100,100,100,100,100,100]} 00:37:10.340 --> 00:37:16.340 spot where you will lose power if the resolution gets NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:37:16.340 --> 00:37:17.780 to a certain level? NOTE CONF {"raw":[100,100,100,100]} 00:37:19.310 --> 00:37:20.480 It's a great question. NOTE CONF {"raw":[100,100,100,100]} 00:37:20.870 --> 00:37:25.040 I don't know if I have a like a perfect NOTE CONF {"raw":[100,100,100,100,100,100,77,100,100,100]} 00:37:25.040 --> 00:37:29.630 answer to this, but in my experience, I think we NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:37:29.630 --> 00:37:32.870 have used the seven Tesla scans before for certain different NOTE CONF {"raw":[100,100,70,100,100,100,100,100,100,100]} 00:37:32.870 --> 00:37:34.100 kinds of questions. NOTE CONF {"raw":[100,100,100]} 00:37:34.190 --> 00:37:37.430 And we have used also one Tesla MRI scans before. NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:37:37.610 --> 00:37:39.560 We have not used JPEGs before. NOTE CONF {"raw":[100,100,100,100,100,100]} 00:37:39.980 --> 00:37:43.130 We always try to resort to taking the DICOM or NOTE CONF {"raw":[100,100,98,100,100,100,100,100,100,100]} 00:37:43.130 --> 00:37:45.530 the Nifty files and then process them directly. NOTE CONF {"raw":[100,100,100,100,100,100,100,100]} 00:37:46.130 --> 00:37:48.830 So we have never done JPEG or we didn't convert NOTE CONF {"raw":[100,84,84,100,100,100,100,100,100,100]} 00:37:48.830 --> 00:37:51.200 those things into some other imaging formats. NOTE CONF {"raw":[100,100,100,100,100,100,100]} 00:37:51.350 --> 00:37:54.740 But what I can tell you is that depending on NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:37:54.740 --> 00:37:57.710 the question you have, for instance, if you are simply NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:37:57.710 --> 00:38:00.380 trying to assess if somebody has atrophy in the brain, NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:38:00.890 --> 00:38:04.880 I don't think a seven Tesla scanner is needed or NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,97]} 00:38:04.910 --> 00:38:07.250 image based on a seven Tesla scanner is needed. NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100]} 00:38:08.180 --> 00:38:10.040 But on the other hand, if we want to somehow NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:38:10.040 --> 00:38:14.450 understand some hippocampal changes that are happening where you want NOTE CONF {"raw":[100,73,100,100,100,100,100,100,100,100]} 00:38:14.450 --> 00:38:18.170 to go in deeper and understand at the voxel level NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:38:18.170 --> 00:38:23.090 what regions are getting changed due to, for example, disease, NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:38:23.570 --> 00:38:25.550 then I think you would need those kinds of a NOTE CONF {"raw":[100,100,100,91,71,100,100,100,100,99]} 00:38:25.550 --> 00:38:27.800 very high end imaging. NOTE CONF {"raw":[100,100,100,100]} 00:38:28.310 --> 00:38:30.890 But in the kind of questions that we are asking NOTE CONF {"raw":[100,100,100,100,100,100,100,99,99,100]} 00:38:30.890 --> 00:38:35.840 here, we have used three Tesla imaging most of the NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:38:35.840 --> 00:38:38.750 time, and I think it seems to have given us NOTE CONF {"raw":[100,100,91,99,100,58,100,100,100,100]} 00:38:38.760 --> 00:38:41.570 reasonable resolution for us to make some progress. NOTE CONF {"raw":[100,100,98,98,100,100,100,100]} 00:38:41.570 --> 00:38:43.700 But I don't know if I'm answering the question, but NOTE CONF {"raw":[100,78,100,100,100,100,100,50,98,100]} 00:38:43.700 --> 00:38:45.770 it's I don't think we have done that kind of NOTE CONF {"raw":[94,100,100,100,100,100,100,100,100,100]} 00:38:45.770 --> 00:38:49.170 a sensitivity analysis to look at different imaging resolutions to NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,78]} 00:38:49.180 --> 00:38:49.820 to make sure. NOTE CONF {"raw":[100,100,100]} 00:38:49.820 --> 00:38:52.220 But at least there is some variability that I think NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:38:52.220 --> 00:38:54.920 the model accounted for while this training was happening. NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100]} 00:38:59.870 --> 00:39:03.320 I think I see two other hands on the screen NOTE CONF {"raw":[67,68,100,100,100,100,100,100,100,100]} 00:39:03.770 --> 00:39:04.220 read. NOTE CONF {"raw":[97]} 00:39:08.720 --> 00:39:09.700 Yes. NOTE CONF {"raw":[100]} 00:39:09.710 --> 00:39:10.180 Hi. NOTE CONF {"raw":[100]} 00:39:10.190 --> 00:39:11.090 Thank you very much. NOTE CONF {"raw":[100,100,100,100]} 00:39:11.090 --> 00:39:12.710 This is riveting. NOTE CONF {"raw":[100,100,100]} 00:39:13.370 --> 00:39:16.220 And appreciate your time today. NOTE CONF {"raw":[100,100,100,100,100]} 00:39:16.220 --> 00:39:23.540 So I'm also one who doesn't normally work in these NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:39:23.540 --> 00:39:24.050 areas. NOTE CONF {"raw":[100]} 00:39:24.050 --> 00:39:27.380 So some of these questions may seem a bit odd. NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:39:27.380 --> 00:39:33.710 I was particularly intrigued with your use of the term NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:39:33.710 --> 00:39:40.700 black box and the importance of understanding and and always NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:39:41.000 --> 00:39:47.540 validating and validating what that black box actually is producing. NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:39:47.540 --> 00:39:53.180 So would you consider that to be a critical factor NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:39:53.180 --> 00:39:58.730 in the design and implementation and production use of these NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:39:58.730 --> 00:39:59.540 systems? NOTE CONF {"raw":[100]} 00:39:59.990 --> 00:40:04.390 Yeah, I absolutely fully agree with that statement because I NOTE CONF {"raw":[98,98,100,100,100,100,100,100,100,85]} 00:40:04.400 --> 00:40:09.140 think, um, the notion of, you know, or relying on NOTE CONF {"raw":[100,79,100,100,100,72,72,100,100,100]} 00:40:09.140 --> 00:40:15.290 a computer to do certain things I think can only NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:40:15.290 --> 00:40:18.080 be trusted if you know what's going on. NOTE CONF {"raw":[100,100,100,100,100,100,100,100]} 00:40:18.470 --> 00:40:23.900 So I think making sure that the model is able NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:40:23.900 --> 00:40:28.370 to learn things and those things are actually relevant, I NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:40:28.370 --> 00:40:31.310 think is one of the as a scientist, I'm interested NOTE CONF {"raw":[100,93,100,100,100,100,100,100,100,100]} 00:40:31.310 --> 00:40:32.270 to know what's happening. NOTE CONF {"raw":[100,100,100,100]} 00:40:32.270 --> 00:40:32.480 Right? NOTE CONF {"raw":[100]} 00:40:32.480 --> 00:40:35.030 So but on the other hand, I think by showing NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:40:35.030 --> 00:40:38.240 these things to the clinicians, by showing these things to NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:40:38.630 --> 00:40:41.090 all the other scientists, I think we were able to NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:40:41.090 --> 00:40:43.970 sort of bring all those people together and be part NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:40:43.970 --> 00:40:44.360 of this, right? NOTE CONF {"raw":[100,100,100]} 00:40:44.360 --> 00:40:46.420 So which is why now we have at least I NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,91]} 00:40:46.430 --> 00:40:51.260 was I fortunately managed to convince some of my colleagues NOTE CONF {"raw":[100,98,100,100,100,100,100,100,100,100]} 00:40:51.260 --> 00:40:52.400 to use the tool. NOTE CONF {"raw":[100,100,100,100]} 00:40:53.030 --> 00:40:56.510 And I think the way they appreciated this was because NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:40:56.510 --> 00:41:00.320 they were seeing when they actually upload an MRI on NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:41:00.320 --> 00:41:04.280 our software, the software basically makes an assessment. NOTE CONF {"raw":[100,100,100,100,100,100,100,100]} 00:41:04.550 --> 00:41:09.110 And then within those images at the slice by slice NOTE CONF {"raw":[100,100,100,100,76,100,100,100,100,100]} 00:41:09.110 --> 00:41:13.070 level, our model tries to identify those hotspots. NOTE CONF {"raw":[100,100,100,100,100,100,100,100]} 00:41:13.730 --> 00:41:16.880 And if those hot spots are somehow corresponding and convincing NOTE CONF {"raw":[100,100,100,72,72,100,100,100,100,100]} 00:41:16.880 --> 00:41:19.760 them about the fact that, okay, yeah, this makes sense, NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:41:19.760 --> 00:41:22.250 then they seem to be very comfortable using the tool, NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:41:22.250 --> 00:41:22.450 right? NOTE CONF {"raw":[80]} 00:41:22.460 --> 00:41:26.060 So this is something that we we heavily rely on NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:41:26.060 --> 00:41:28.370 and we, we think it must be done in order NOTE CONF {"raw":[100,80,100,100,100,100,100,100,100,100]} 00:41:28.370 --> 00:41:31.010 for any kind of a model to to be in NOTE CONF {"raw":[100,100,100,100,94,100,100,100,100,100]} 00:41:31.010 --> 00:41:31.910 the clinical practice. NOTE CONF {"raw":[100,100,100]} 00:41:31.910 --> 00:41:34.520 And I think people who are probably working in the NOTE CONF {"raw":[100,99,100,100,100,100,100,100,100,100]} 00:41:34.520 --> 00:41:37.790 radiology setting may very well know by now that there NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:41:37.790 --> 00:41:40.940 is already AI, whether you want it or not, in NOTE CONF {"raw":[100,100,76,100,100,100,100,100,100,100]} 00:41:40.940 --> 00:41:43.340 the systems and it's happening in the back end. NOTE CONF {"raw":[100,100,100,56,100,100,100,90,90]} 00:41:43.340 --> 00:41:46.100 Some of it actually, you're probably using it even today, NOTE CONF {"raw":[100,100,100,100,84,100,100,100,100,100]} 00:41:46.310 --> 00:41:46.790 right? NOTE CONF {"raw":[100]} 00:41:47.630 --> 00:41:51.860 So so I think as these things evolve, I think NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:41:51.860 --> 00:41:55.670 it's very important for us to think about explainability interpretability NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:41:55.670 --> 00:41:56.510 and validation. NOTE CONF {"raw":[100,100]} 00:41:57.780 --> 00:41:58.110 Okay. NOTE CONF {"raw":[100]} 00:41:58.110 --> 00:41:58.770 Thank you. NOTE CONF {"raw":[100,100]} 00:41:58.810 --> 00:42:01.230 I have a follow on question, but I'll defer to NOTE CONF {"raw":[94,100,100,100,100,100,100,100,100,100]} 00:42:01.230 --> 00:42:02.820 others before ask it. NOTE CONF {"raw":[100,100,100,100]} 00:42:05.200 --> 00:42:05.520 Sure. NOTE CONF {"raw":[66]} 00:42:07.870 --> 00:42:08.470 Please go ahead. NOTE CONF {"raw":[100,100,100]} 00:42:09.160 --> 00:42:10.030 Oh, did I? NOTE CONF {"raw":[100,92,71]} 00:42:10.150 --> 00:42:12.340 Did someone else have a question? NOTE CONF {"raw":[100,100,100,100,100,100]} 00:42:12.460 --> 00:42:13.210 Oh, I see. NOTE CONF {"raw":[100,100,100]} 00:42:13.450 --> 00:42:14.260 Yeah, I see. NOTE CONF {"raw":[100,100,100]} 00:42:14.260 --> 00:42:15.460 One more hand raised. NOTE CONF {"raw":[100,100,100,100]} 00:42:17.100 --> 00:42:17.970 Dr. Wang. NOTE CONF {"raw":[86,75]} 00:42:19.080 --> 00:42:19.440 Um. NOTE CONF {"raw":[100]} 00:42:19.440 --> 00:42:20.010 Hi. NOTE CONF {"raw":[100]} 00:42:20.400 --> 00:42:20.820 Hi. NOTE CONF {"raw":[100]} 00:42:20.820 --> 00:42:22.300 This is John. NOTE CONF {"raw":[100,100,85]} 00:42:22.350 --> 00:42:27.630 So I have several questions about the deep learning techniques. NOTE CONF {"raw":[96,100,100,100,100,100,100,92,100,100]} 00:42:27.630 --> 00:42:31.380 So the first question is, how do you how do NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,58]} 00:42:31.380 --> 00:42:37.360 you integrate multimodal features, especially the imaging data and non NOTE CONF {"raw":[100,100,89,100,100,92,75,100,100,95]} 00:42:37.380 --> 00:42:38.310 image data? NOTE CONF {"raw":[93,100]} 00:42:38.340 --> 00:42:42.550 So if I understand correctly, it seems to your use NOTE CONF {"raw":[100,100,100,100,100,100,100,100,98,87]} 00:42:42.870 --> 00:42:47.160 say to encode imaging feature, I guess you use the NOTE CONF {"raw":[46,100,100,91,82,100,100,100,99,78]} 00:42:47.160 --> 00:42:51.030 same thing to map the image data to a feature NOTE CONF {"raw":[67,73,100,100,100,100,100,100,100,100]} 00:42:51.030 --> 00:42:54.150 space into a feature space and then that will be NOTE CONF {"raw":[100,100,100,100,100,100,93,100,94,100]} 00:42:54.150 --> 00:42:54.750 a vector. NOTE CONF {"raw":[100,100]} 00:42:54.760 --> 00:42:59.020 Then you use that actor and put the put the NOTE CONF {"raw":[97,51,100,95,93,99,97,97,90,88]} 00:42:59.070 --> 00:43:02.520 information like of non image data. NOTE CONF {"raw":[100,100,100,65,84,100]} 00:43:02.910 --> 00:43:04.220 That's also another actress. NOTE CONF {"raw":[65,100,62,80]} 00:43:04.230 --> 00:43:07.770 And just to put it together to do the Amazon NOTE CONF {"raw":[75,100,57,100,100,100,100,100,100,100]} 00:43:08.160 --> 00:43:11.460 continue the deep learning thing, I'm not sure about this. NOTE CONF {"raw":[99,100,100,100,60,100,100,100,100,100]} 00:43:11.460 --> 00:43:12.870 So this is the first thing. NOTE CONF {"raw":[100,100,100,100,100,100]} 00:43:12.870 --> 00:43:16.560 And so technology is so if you you say to NOTE CONF {"raw":[46,100,61,100,100,100,63,64,51,100]} 00:43:16.830 --> 00:43:19.530 encode the image data to feature. NOTE CONF {"raw":[100,100,88,100,100,100]} 00:43:19.530 --> 00:43:21.420 So how do you do that? NOTE CONF {"raw":[100,100,100,100,100,100]} 00:43:21.420 --> 00:43:24.540 So I mean, if it's a supervised learning or unsupervised NOTE CONF {"raw":[100,100,100,90,100,100,100,100,100,100]} 00:43:24.540 --> 00:43:27.540 learning, Uh, so this is a second question. NOTE CONF {"raw":[100,65,100,100,100,100,100,100]} 00:43:27.660 --> 00:43:30.810 So the question is about so you use different kind NOTE CONF {"raw":[51,100,100,100,100,100,100,100,100,96]} 00:43:30.810 --> 00:43:31.560 of data. NOTE CONF {"raw":[100,100]} 00:43:31.560 --> 00:43:35.280 So I guess everyone can have an MRI scan. NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100]} 00:43:35.280 --> 00:43:38.490 So how about the other known imaging information? NOTE CONF {"raw":[100,100,100,100,100,80,92,100]} 00:43:38.490 --> 00:43:41.640 So if one kind of the data is missing, how NOTE CONF {"raw":[100,99,100,98,100,86,100,100,100,100]} 00:43:41.640 --> 00:43:42.900 did you deal with that? NOTE CONF {"raw":[100,100,100,100,100]} 00:43:42.900 --> 00:43:43.230 Yeah. NOTE CONF {"raw":[100]} 00:43:43.590 --> 00:43:44.310 It's a great question. NOTE CONF {"raw":[95,100,100,100]} 00:43:44.310 --> 00:43:48.120 So the first question is, if I understand correctly. NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100]} 00:43:48.120 --> 00:43:50.850 So how are you trying to encode this information? NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100]} 00:43:50.850 --> 00:43:53.340 Yes, we are using CNN and we are kind of NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:43:53.340 --> 00:43:55.560 jointly training the information together. NOTE CONF {"raw":[100,100,89,100,100]} 00:43:55.560 --> 00:43:56.310 So that is right. NOTE CONF {"raw":[100,100,100,100]} 00:43:56.310 --> 00:44:00.000 So we combine the or concatenate the information from the NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:44:00.000 --> 00:44:03.960 CNN or from the MRI and the other one, which NOTE CONF {"raw":[86,100,100,100,100,100,100,100,100,100]} 00:44:03.960 --> 00:44:07.260 is non imaging data, and then we learn them together. NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:44:07.260 --> 00:44:09.810 So the back propagation that is happening is happening on NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:44:09.810 --> 00:44:11.070 both at the same time. NOTE CONF {"raw":[100,100,100,100,100]} 00:44:11.850 --> 00:44:14.910 And the second question is how do we do this? NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:44:14.910 --> 00:44:18.300 Basically, this is a supervised learning task because clearly we NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:44:18.300 --> 00:44:21.480 are trying to predict who has these labels. NOTE CONF {"raw":[100,100,100,100,100,100,100,100]} 00:44:21.480 --> 00:44:24.750 So when we we have built this framework, which, by NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:44:24.750 --> 00:44:28.200 the way, we are happy to share, we have publicly NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:44:28.200 --> 00:44:29.400 released all these codes. NOTE CONF {"raw":[100,100,100,100]} 00:44:29.400 --> 00:44:32.160 And in fact, most people reached out to us and NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:44:32.160 --> 00:44:33.690 we have provided all this information. NOTE CONF {"raw":[100,100,100,100,100,100]} 00:44:33.690 --> 00:44:36.780 So you're welcome to take a look at what we NOTE CONF {"raw":[100,86,100,100,100,100,100,100,100,100]} 00:44:36.780 --> 00:44:37.320 have done. NOTE CONF {"raw":[100,100]} 00:44:38.010 --> 00:44:40.800 The entire computer scripts and everything we made make it NOTE CONF {"raw":[100,100,100,84,100,100,100,94,100,100]} 00:44:40.800 --> 00:44:41.880 publicly available. NOTE CONF {"raw":[100,100]} 00:44:42.720 --> 00:44:44.580 And this is a supervised learning task. NOTE CONF {"raw":[100,100,100,100,100,100,100]} 00:44:44.580 --> 00:44:48.360 So the multimodal learning is happening as a supervised learning NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:44:48.360 --> 00:44:48.810 task. NOTE CONF {"raw":[100]} 00:44:48.810 --> 00:44:51.750 And the way we do is just to clarify the NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:44:51.750 --> 00:44:54.420 first task, which is NC, MCI dementia. NOTE CONF {"raw":[100,100,100,100,77,100,100]} 00:44:54.420 --> 00:44:57.480 We we consider that as a regression task, which means NOTE CONF {"raw":[100,100,86,100,100,100,100,100,100,100]} 00:44:57.480 --> 00:45:01.800 it's basically, you know, healthy means like a zero value NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:45:01.800 --> 00:45:04.590 or a, you know, cases like a two value. NOTE CONF {"raw":[100,92,100,100,100,100,100,100,100]} 00:45:04.590 --> 00:45:07.350 So we sort of do the regression task first and NOTE CONF {"raw":[100,100,100,100,100,94,100,100,100,100]} 00:45:07.350 --> 00:45:09.780 then the second task is more sort of a classification NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:45:09.780 --> 00:45:13.020 task, which is the differentiating between those who have Alzheimer's NOTE CONF {"raw":[100,100,100,56,100,100,100,100,100,100]} 00:45:13.020 --> 00:45:16.770 disease, dementia versus those who have other etiologies of dementia. NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:45:17.860 --> 00:45:19.920 And the third question is missing data, Right. NOTE CONF {"raw":[100,100,100,100,100,100,100,70]} 00:45:19.930 --> 00:45:22.450 So that's a very good question. NOTE CONF {"raw":[100,100,100,100,100,100]} 00:45:22.450 --> 00:45:25.300 So right now, we were fortunate to get access to NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:45:25.300 --> 00:45:29.020 these multimodal data cohorts from like NAC and other places. NOTE CONF {"raw":[100,100,100,100,100,92,75,100,100,100]} 00:45:29.020 --> 00:45:32.740 So we were able to use all this information to NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:45:32.740 --> 00:45:33.910 train the models. NOTE CONF {"raw":[100,100,100]} 00:45:33.910 --> 00:45:36.430 But I do agree with you that some of these NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:45:36.430 --> 00:45:39.460 cohorts may not have all the information that you need NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:45:39.460 --> 00:45:41.740 to then test or do something right. NOTE CONF {"raw":[100,100,100,100,100,100,100]} 00:45:41.740 --> 00:45:45.120 So but again, our goal is not to use MRIs NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,88]} 00:45:45.130 --> 00:45:48.370 alone, because whenever I talk to a neurologist, they keep NOTE CONF {"raw":[100,100,100,100,98,100,100,100,100,100]} 00:45:48.370 --> 00:45:51.730 telling me again and again that MRI is only used NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:45:51.730 --> 00:45:54.760 to either rule in or rule out indications of dementia. NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:45:54.760 --> 00:45:57.190 That's not their primary modality of assessment. NOTE CONF {"raw":[100,100,100,100,100,100,100]} 00:45:57.460 --> 00:46:01.780 And although I'm biased because I'm a computer scientist with NOTE CONF {"raw":[74,100,100,100,100,100,100,100,100,100]} 00:46:01.780 --> 00:46:04.060 a lot of experience in imaging, I was so desperate NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:46:04.060 --> 00:46:06.670 to use MRIs only to do everything. NOTE CONF {"raw":[100,100,100,100,95,95,100]} 00:46:06.820 --> 00:46:09.550 But that's probably not what they do in a clinical NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:46:09.550 --> 00:46:10.180 practice, right? NOTE CONF {"raw":[100,100]} 00:46:10.180 --> 00:46:13.750 So so they have taught me to stay away from NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:46:13.750 --> 00:46:16.240 MRIs, only use MRIs when needed, Right? NOTE CONF {"raw":[100,100,100,100,100,100,100]} 00:46:16.240 --> 00:46:18.190 So so which is why I think we spent a NOTE CONF {"raw":[100,100,100,100,100,100,100,100,86,100]} 00:46:18.190 --> 00:46:20.530 lot of effort to think about what is actually what NOTE CONF {"raw":[100,100,100,100,100,100,100,100,62,100]} 00:46:20.530 --> 00:46:24.160 does it mean by the data that is collected routinely NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:46:24.160 --> 00:46:26.500 in a setting, in a clinical setting, when I say NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:46:26.500 --> 00:46:28.860 clinical setting, a clinical neurology setting, right. NOTE CONF {"raw":[100,100,100,100,100,100,97]} 00:46:28.870 --> 00:46:31.720 So so because our goal is to build the tool NOTE CONF {"raw":[100,100,100,100,100,100,100,100,52,100]} 00:46:31.870 --> 00:46:33.970 and we had to go back and then see what's NOTE CONF {"raw":[100,100,97,100,100,100,100,100,100,100]} 00:46:33.970 --> 00:46:36.730 that that's been collected and then use that information. NOTE CONF {"raw":[100,100,97,100,100,100,100,100,100]} 00:46:37.360 --> 00:46:39.640 And I think the current approaches that we are developing NOTE CONF {"raw":[100,100,100,100,100,100,100,98,98,100]} 00:46:40.000 --> 00:46:42.870 have the capability of also thinking about missing data. NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100]} 00:46:42.880 --> 00:46:47.170 So it's an interesting question for many reasons, because this NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:46:47.170 --> 00:46:49.480 is a thing that we have been trying to address NOTE CONF {"raw":[100,79,100,100,100,100,100,100,100,100]} 00:46:49.480 --> 00:46:52.570 in a different way, which is, you know, if a NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:46:52.570 --> 00:46:55.420 neurologist sees a patient, this is again, a I mean, NOTE CONF {"raw":[100,100,100,100,100,100,100,63,83,100]} 00:46:55.420 --> 00:46:57.580 please correct me on the MDS here. NOTE CONF {"raw":[100,100,100,84,100,91,100]} 00:46:57.580 --> 00:47:00.700 Please correct me if I'm wrong, but a neurologist sees NOTE CONF {"raw":[100,100,100,100,100,100,100,89,98,100]} 00:47:00.700 --> 00:47:03.040 a patient, the patient walks into the clinic. NOTE CONF {"raw":[98,100,100,100,100,100,100,100]} 00:47:03.040 --> 00:47:05.740 They may have some information that is coming from, let's NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:47:05.740 --> 00:47:07.780 say, the primary care provider or some other provider. NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100]} 00:47:07.780 --> 00:47:11.380 In the past, the at that point of care, the NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:47:11.380 --> 00:47:13.990 neurologist has to make a decision, right? NOTE CONF {"raw":[100,100,100,100,100,100,63]} 00:47:14.080 --> 00:47:17.740 Which means whatever data that is available, whether it's history, NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:47:18.310 --> 00:47:21.460 information coming from their care provider, from all the bedside NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:47:21.460 --> 00:47:24.130 testing that is happening at the point of care and NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:47:24.130 --> 00:47:26.860 maybe am I may or may not be MRI, the NOTE CONF {"raw":[100,92,92,100,100,100,100,100,100,100]} 00:47:26.860 --> 00:47:30.730 neurologist has to somehow make the decision of understanding what's NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:47:30.730 --> 00:47:34.570 happening in the patient at that point, which means most NOTE CONF {"raw":[100,76,100,100,100,100,100,100,100,100]} 00:47:34.570 --> 00:47:36.760 often than not, there will be missing data. NOTE CONF {"raw":[100,58,100,100,100,100,100,100]} 00:47:37.000 --> 00:47:39.750 And yet the neurologist is making a decision, right? NOTE CONF {"raw":[100,100,100,100,100,100,100,100,93]} 00:47:39.760 --> 00:47:41.800 So if you really want to build an AI tool, NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:47:42.010 --> 00:47:44.260 I think that's what we need to do, which is NOTE CONF {"raw":[100,100,100,100,100,100,100,52,100,100]} 00:47:44.260 --> 00:47:47.500 take whatever information is available and make a decision. NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100]} 00:47:47.500 --> 00:47:47.740 Right. NOTE CONF {"raw":[100]} 00:47:47.740 --> 00:47:49.870 Which means our models have to be robust enough to NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:47:49.870 --> 00:47:50.470 do that. NOTE CONF {"raw":[100,100]} 00:47:50.470 --> 00:47:55.180 Unfortunately, most of the methods that have been developed kind NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:47:55.180 --> 00:47:58.390 of impute the data to to make those assessments. NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100]} 00:47:58.570 --> 00:47:58.870 Yes. NOTE CONF {"raw":[49]} 00:47:58.990 --> 00:48:01.510 Which I think is one approximate way of doing things, NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:48:01.510 --> 00:48:03.220 but not the idea of because I don't think a NOTE CONF {"raw":[100,100,100,99,52,100,100,100,100,100]} 00:48:03.220 --> 00:48:04.900 neurologist is imputing data here. NOTE CONF {"raw":[100,100,68,100,100]} 00:48:04.900 --> 00:48:05.080 Right. NOTE CONF {"raw":[100]} 00:48:05.080 --> 00:48:09.010 So neurologist has this intelligence, the experience to make the NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:48:09.010 --> 00:48:11.410 assessment with whatever information is available. NOTE CONF {"raw":[100,100,100,100,100,100]} 00:48:11.470 --> 00:48:13.150 So I think that's the real eye. NOTE CONF {"raw":[99,98,100,100,100,100,60]} 00:48:13.150 --> 00:48:15.520 And hopefully, you know, we are, you know, going in NOTE CONF {"raw":[100,100,89,89,100,100,98,98,100,100]} 00:48:15.550 --> 00:48:16.240 that direction. NOTE CONF {"raw":[100,100]} 00:48:17.500 --> 00:48:18.050 Thank you. NOTE CONF {"raw":[100,100]} 00:48:18.070 --> 00:48:19.280 Thank you for your answer. NOTE CONF {"raw":[98,98,100,100,100]} 00:48:19.310 --> 00:48:19.780 What? NOTE CONF {"raw":[99]} 00:48:19.990 --> 00:48:22.760 It is a very, very good talk. NOTE CONF {"raw":[82,100,96,100,100,100,79]} 00:48:22.780 --> 00:48:24.910 I'm very impressed with it. NOTE CONF {"raw":[94,100,63,64,96]} 00:48:24.940 --> 00:48:26.230 Thank you very much. NOTE CONF {"raw":[100,100,100,100]} 00:48:26.260 --> 00:48:26.650 Thank you. NOTE CONF {"raw":[100,100]} 00:48:28.100 --> 00:48:28.550 Think. NOTE CONF {"raw":[100]} 00:48:30.650 --> 00:48:34.790 There was second question by the previous person. NOTE CONF {"raw":[100,100,100,100,100,100,100,100]} 00:48:34.790 --> 00:48:35.210 Yeah. NOTE CONF {"raw":[99]} 00:48:37.540 --> 00:48:38.110 Please go ahead. NOTE CONF {"raw":[100,100,100]} 00:48:46.480 --> 00:48:49.690 I think someone still has their hand raised. NOTE CONF {"raw":[82,100,100,100,100,100,100,100]} 00:48:51.820 --> 00:48:53.650 I think the follow up question. NOTE CONF {"raw":[42,68,99,100,100,100]} 00:48:53.980 --> 00:48:54.790 Yes, please. NOTE CONF {"raw":[100,100]} 00:48:54.790 --> 00:48:55.060 Yeah. NOTE CONF {"raw":[100]} 00:48:55.060 --> 00:48:55.330 Yeah. NOTE CONF {"raw":[100]} 00:48:55.810 --> 00:49:04.030 So my my second question is, was kind of already NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:49:04.030 --> 00:49:10.570 cued up by your comment about AI and the problem NOTE CONF {"raw":[67,100,100,100,100,100,90,100,100,100]} 00:49:10.870 --> 00:49:14.890 that we're seeing in the. NOTE CONF {"raw":[100,100,100,100,100]} 00:49:16.520 --> 00:49:21.770 In the environment with AI. NOTE CONF {"raw":[100,100,100,100,96]} 00:49:21.980 --> 00:49:26.690 The black boxes for AI are now considered in many NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:49:26.690 --> 00:49:31.400 settings to be proprietary, and so questions are not permitted NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:49:31.400 --> 00:49:37.310 to be asked about the AI because as these systems NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:49:37.310 --> 00:49:41.000 are commercialized the way that. NOTE CONF {"raw":[100,91,100,100,100]} 00:49:41.980 --> 00:49:46.960 Then this works in our particular economy is the black NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:49:46.960 --> 00:49:51.490 box becomes proprietary and there is no oversight in the NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:49:51.490 --> 00:49:53.170 absence of regulation. NOTE CONF {"raw":[100,100,100]} 00:49:54.130 --> 00:49:57.760 And so and right now, since a lot of the NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:49:57.760 --> 00:50:01.570 regulatory apparatus in the United States has been disassembled. NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100]} 00:50:03.920 --> 00:50:04.280 It. NOTE CONF {"raw":[100]} 00:50:05.030 --> 00:50:08.090 Is it part of. NOTE CONF {"raw":[100,91,100,100]} 00:50:08.900 --> 00:50:16.040 Is it part of your design to also essentially provide NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:50:16.280 --> 00:50:20.480 best practices, if you will, for the governance elements that NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:50:20.480 --> 00:50:21.470 you're talking about? NOTE CONF {"raw":[100,100,100]} 00:50:21.470 --> 00:50:25.460 In other words, if you're going to use this system NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:50:25.460 --> 00:50:30.380 or these systems, then you must build your own improvement NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:50:30.380 --> 00:50:35.690 models for constantly testing and validating the engine. NOTE CONF {"raw":[100,100,100,100,100,100,100,100]} 00:50:35.690 --> 00:50:41.720 And that that should be nonproprietary information not only for NOTE CONF {"raw":[100,100,100,100,100,98,100,100,100,100]} 00:50:41.720 --> 00:50:46.100 the benefit the community, but also for the continuous improvement NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:50:46.100 --> 00:50:50.390 of the technology for patient care, plus the confidence of NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:50:50.390 --> 00:50:52.580 clinicians who are depending on it. NOTE CONF {"raw":[100,100,100,100,100,100]} 00:50:52.580 --> 00:50:56.960 They won't be blocked from asking questions about, well, how NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:50:56.960 --> 00:51:01.850 do I know this is really giving the optimal answers NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:51:01.850 --> 00:51:04.580 and how will you improve? NOTE CONF {"raw":[100,100,100,100,100]} 00:51:04.580 --> 00:51:09.560 My confidence in that is where will that governance belong NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:51:09.560 --> 00:51:11.880 in the absence of regulation? NOTE CONF {"raw":[100,100,100,100,100]} 00:51:11.900 --> 00:51:14.360 Would you would you recommend. NOTE CONF {"raw":[100,100,100,100,100]} 00:51:14.840 --> 00:51:17.210 Oh, that's a very. NOTE CONF {"raw":[100,100,100,100]} 00:51:18.050 --> 00:51:18.350 Video. NOTE CONF {"raw":[96]} 00:51:18.350 --> 00:51:19.490 Interesting question. NOTE CONF {"raw":[100,100]} 00:51:20.360 --> 00:51:22.760 We're going to elect you to Congress if you come NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:51:22.760 --> 00:51:23.870 up with an answer. NOTE CONF {"raw":[100,100,100,100]} 00:51:26.060 --> 00:51:26.300 Uh. NOTE CONF {"raw":[71]} 00:51:27.020 --> 00:51:31.310 I am part of some committees, and I think I NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,83]} 00:51:31.310 --> 00:51:34.220 was actually at the NIH a couple of months ago NOTE CONF {"raw":[100,100,100,100,100,100,100,85,100,100]} 00:51:34.220 --> 00:51:38.810 to discuss something similar, not exactly about governance, but sort NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:51:38.810 --> 00:51:44.540 of really thinking about the overwhelming amount of submissions that NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:51:44.540 --> 00:51:47.930 are going to the FDA these days on on different NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:51:47.930 --> 00:51:51.320 kinds of products, not in terms of just diagnosis, but NOTE CONF {"raw":[100,100,100,100,100,100,100,100,97,100]} 00:51:51.320 --> 00:51:56.680 assistive tools and triaging tools and management, administrative tools, etcetera, NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,70]} 00:51:56.680 --> 00:51:57.890 and so on and so forth. NOTE CONF {"raw":[83,100,100,100,100,100]} 00:51:58.550 --> 00:52:02.750 Um, it's still an ongoing debate as to who is NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:52:02.750 --> 00:52:05.570 going to take the responsibility or what kind of responsibility NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,97]} 00:52:05.570 --> 00:52:08.090 should be even taken by these people who are using NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:52:08.090 --> 00:52:09.020 these tools. NOTE CONF {"raw":[100,100]} 00:52:09.650 --> 00:52:17.540 Um, but I think the, the, the definition of responsibility NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:52:17.630 --> 00:52:21.380 I think would change depending on the use case, in NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:52:21.380 --> 00:52:22.370 my opinion. NOTE CONF {"raw":[100,100]} 00:52:22.760 --> 00:52:25.100 Ultimately, the clinician is making the decision. NOTE CONF {"raw":[100,100,100,100,100,100,100]} 00:52:25.100 --> 00:52:27.440 Ultimately, clinician is managing the patient. NOTE CONF {"raw":[100,100,100,100,100,100]} 00:52:28.070 --> 00:52:31.220 Any tool that you are providing to the clinician hopefully NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:52:31.220 --> 00:52:35.090 has to either improve their, for lack of better words, NOTE CONF {"raw":[100,100,100,100,65,100,100,100,100,100]} 00:52:35.090 --> 00:52:39.440 improve their building capacity or improve their clinical clinical efficiency, NOTE CONF {"raw":[100,100,54,100,100,100,100,84,100,100]} 00:52:39.440 --> 00:52:39.850 if you will. NOTE CONF {"raw":[100,100,100]} 00:52:39.970 --> 00:52:40.100 Right. NOTE CONF {"raw":[64]} 00:52:40.100 --> 00:52:42.950 So ultimately, we need to think about how these tools NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:52:42.950 --> 00:52:45.800 are going to improve efficiency in these practices. NOTE CONF {"raw":[100,100,100,100,100,100,100,100]} 00:52:45.800 --> 00:52:48.770 And for that specific part is what I'm interested in, NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:52:48.770 --> 00:52:53.540 which is ultimately the clinician is interacting with the patient, NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:52:53.540 --> 00:52:55.520 they are the decision maker, they are the ones who NOTE CONF {"raw":[100,100,100,100,92,100,100,100,100,100]} 00:52:55.520 --> 00:52:56.570 are managing it. NOTE CONF {"raw":[100,100,100]} 00:52:56.570 --> 00:53:00.290 But can I somehow reduce their their burden? NOTE CONF {"raw":[100,100,100,100,100,100,100,100]} 00:53:00.440 --> 00:53:02.480 Can I somehow make sure that they are able to NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:53:02.480 --> 00:53:03.800 do these things in the right way? NOTE CONF {"raw":[100,100,100,100,100,100,100]} 00:53:04.010 --> 00:53:05.660 But on the other hand, if you're asking a question NOTE CONF {"raw":[100,100,100,100,100,100,100,100,63,96]} 00:53:05.660 --> 00:53:09.580 about whether the clinician has to trust something here, um, NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,86]} 00:53:10.850 --> 00:53:13.280 there are already lots of tools, by the way, that NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:53:13.280 --> 00:53:15.230 you guys are using, which you don't know there is NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:53:15.230 --> 00:53:16.510 AI behind it, right? NOTE CONF {"raw":[89,100,100,84]} 00:53:16.550 --> 00:53:17.890 So you're using it already. NOTE CONF {"raw":[100,100,100,100,100]} 00:53:17.900 --> 00:53:22.360 Clearly those tools are probably not not have not reached NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:53:22.360 --> 00:53:25.970 to the stage where, you know, at that critical juncture NOTE CONF {"raw":[85,100,100,100,100,100,100,100,100,100]} 00:53:25.970 --> 00:53:28.130 where you have to decide on the patient, for instance. NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:53:28.130 --> 00:53:30.170 But in the back end, they're already happening. NOTE CONF {"raw":[100,100,100,100,100,99,100,100]} 00:53:30.590 --> 00:53:30.890 Right. NOTE CONF {"raw":[100]} 00:53:30.890 --> 00:53:36.410 So so I think these tools have to be in NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:53:36.410 --> 00:53:37.610 this assistive mode. NOTE CONF {"raw":[100,100,100]} 00:53:37.610 --> 00:53:40.190 I don't think they should sort of do anything beyond NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:53:40.190 --> 00:53:40.670 that. NOTE CONF {"raw":[100]} 00:53:40.670 --> 00:53:43.640 And for them to be in the assistive mode, I NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:53:43.640 --> 00:53:49.310 think the regulation can be less stringent given that these NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:53:49.310 --> 00:53:52.730 tools are simply improving their efficiency. NOTE CONF {"raw":[100,100,100,100,100,100]} 00:53:54.050 --> 00:53:56.360 And that's where I can that's what I can talk NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:53:56.360 --> 00:53:56.630 about. NOTE CONF {"raw":[100]} 00:53:56.630 --> 00:53:58.520 I don't think I can talk about beyond that at NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:53:58.520 --> 00:53:59.210 this point. NOTE CONF {"raw":[100,100]} 00:54:00.700 --> 00:54:01.270 Great. NOTE CONF {"raw":[100]} 00:54:01.450 --> 00:54:02.260 Well, thank you. NOTE CONF {"raw":[100,100,100]} 00:54:02.260 --> 00:54:04.240 And thank you for your work and thank you for NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:54:04.240 --> 00:54:06.010 your conscientiousness about it. NOTE CONF {"raw":[100,100,100,100]} 00:54:06.010 --> 00:54:06.490 Thank you. NOTE CONF {"raw":[100,100]} 00:54:07.390 --> 00:54:09.670 Well, if you all are interested, I can talk to NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:54:09.670 --> 00:54:12.850 Dr. Merman and then I can, you know, give the NOTE CONF {"raw":[99,23,100,100,100,100,100,100,100,100]} 00:54:12.850 --> 00:54:16.410 tool to see if there is any way to, um. NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,88]} 00:54:17.900 --> 00:54:19.730 This tool can be useful because I still don't know NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:54:19.730 --> 00:54:21.020 the answer, to be honest with you. NOTE CONF {"raw":[100,100,100,100,100,100,100]} 00:54:21.020 --> 00:54:24.350 I still don't know the actual answer as to whether NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:54:24.350 --> 00:54:28.940 this is going to change or assist clinical practice in NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:54:28.940 --> 00:54:29.510 any way. NOTE CONF {"raw":[100,100]} 00:54:29.600 --> 00:54:33.890 I mean, I think it's a great, exciting research question. NOTE CONF {"raw":[100,100,100,100,99,100,100,100,100,100]} 00:54:33.890 --> 00:54:36.860 So if you actually do PubMed search on machine learning NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:54:36.860 --> 00:54:41.240 and you'll probably see 10,000 papers, but what does that NOTE CONF {"raw":[100,92,100,100,100,100,100,100,100,100]} 00:54:41.240 --> 00:54:41.630 mean? NOTE CONF {"raw":[100]} 00:54:41.870 --> 00:54:45.560 So why should we do these kinds of research thing NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:54:45.560 --> 00:54:47.060 if they're not going to be useful at the end NOTE CONF {"raw":[100,73,100,100,100,100,100,100,100,100]} 00:54:47.060 --> 00:54:47.480 of the day? NOTE CONF {"raw":[100,100,100]} 00:54:47.660 --> 00:54:50.150 So I'm still trying to figure that out. NOTE CONF {"raw":[100,91,100,100,100,100,100,100]} 00:54:50.150 --> 00:54:56.030 And the value proposition, interestingly, is different for different people. NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:54:56.630 --> 00:55:00.140 If I'm talking with a practice manager, actually one of NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:55:00.140 --> 00:55:03.530 my became a good friend now because I bombarded them NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:55:03.530 --> 00:55:06.650 with thousands of questions every every week and they say, NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:55:06.650 --> 00:55:09.320 oh, you know, are you going to tell me about NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:55:09.320 --> 00:55:11.750 how is this going to change my billing capacity? NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100]} 00:55:11.840 --> 00:55:12.920 That's their problem. NOTE CONF {"raw":[100,100,100]} 00:55:13.130 --> 00:55:15.710 It's ultimately they want to know whether you can charge NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:55:15.710 --> 00:55:18.800 a complexity code of 5 or 4, and rather than NOTE CONF {"raw":[100,100,92,95,99,99,99,100,100,100]} 00:55:18.800 --> 00:55:21.320 spending 50 minutes or 45 minutes on a case, can NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:55:21.320 --> 00:55:22.340 you spend 30 minutes? NOTE CONF {"raw":[100,100,100,100]} 00:55:22.340 --> 00:55:22.490 Right. NOTE CONF {"raw":[52]} 00:55:22.550 --> 00:55:23.990 So can this help me do that? NOTE CONF {"raw":[100,100,100,100,100,100,100]} 00:55:24.230 --> 00:55:25.460 That's the question they asked me. NOTE CONF {"raw":[100,100,100,100,96,100]} 00:55:25.470 --> 00:55:25.890 Right. NOTE CONF {"raw":[86]} 00:55:25.970 --> 00:55:28.310 On the other hand, if I'm talking with Dr. Merman, NOTE CONF {"raw":[100,100,100,100,100,100,100,100,95,43]} 00:55:28.310 --> 00:55:32.900 he's like, oh, this is interesting because maybe this can NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:55:32.900 --> 00:55:36.050 be useful, especially in those complex cases where I want NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:55:36.050 --> 00:55:38.510 to understand the differential diagnosis of dementia and I need NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:55:38.510 --> 00:55:40.000 more supporting information, right? NOTE CONF {"raw":[100,100,100,96]} 00:55:40.010 --> 00:55:42.290 So that's a different question or a different utility or NOTE CONF {"raw":[100,100,100,100,100,100,76,100,100,100]} 00:55:42.290 --> 00:55:43.070 value, right? NOTE CONF {"raw":[100,100]} 00:55:43.190 --> 00:55:46.490 So but how can I manage these two different things? NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:55:46.490 --> 00:55:49.790 Because if it's not improving billing capacity, if it's only NOTE CONF {"raw":[100,100,100,100,100,96,100,100,100,100]} 00:55:49.790 --> 00:55:52.190 answering this question, how is this tool going to be NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:55:52.190 --> 00:55:54.860 practically be integrated in the clinical system? NOTE CONF {"raw":[100,85,100,100,100,100,100]} 00:55:54.860 --> 00:55:55.070 Right. NOTE CONF {"raw":[100]} 00:55:55.070 --> 00:55:58.410 So these are unknown for me at this point, and NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,82]} 00:55:58.430 --> 00:56:01.400 that's why I want to make sure I give these NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:56:01.400 --> 00:56:04.250 tools and understand these things in a more meaningful way. NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:56:06.070 --> 00:56:06.520 Very good. NOTE CONF {"raw":[100,100]} 00:56:06.520 --> 00:56:07.120 Thank you. NOTE CONF {"raw":[100,100]} 00:56:11.900 --> 00:56:14.590 I think I see a couple questions on the chat, NOTE CONF {"raw":[76,79,100,100,100,100,100,100,100,100]} 00:56:14.680 --> 00:56:17.680 which is if anyone's interested, the system is going to NOTE CONF {"raw":[100,100,100,93,100,98,100,94,96,60]} 00:56:17.680 --> 00:56:19.510 post is a real time project in the electronic health NOTE CONF {"raw":[73,99,100,100,100,100,100,100,100,100]} 00:56:19.510 --> 00:56:20.100 record system. NOTE CONF {"raw":[99,80]} 00:56:20.110 --> 00:56:20.590 I see. NOTE CONF {"raw":[100,100]} 00:56:20.950 --> 00:56:21.300 I see. NOTE CONF {"raw":[100,100]} 00:56:21.310 --> 00:56:23.620 I would love to learn more if there is any NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:56:23.620 --> 00:56:25.600 link or some information on that. NOTE CONF {"raw":[100,100,100,100,100,100]} 00:56:26.800 --> 00:56:29.740 Um, and then could you provide us with a few NOTE CONF {"raw":[96,100,100,100,100,100,100,100,100,100]} 00:56:29.740 --> 00:56:32.410 examples of how the model tool can be utilized in NOTE CONF {"raw":[100,100,100,100,100,66,100,100,95,100]} 00:56:32.410 --> 00:56:33.670 real world situations? NOTE CONF {"raw":[100,100,100]} 00:56:33.790 --> 00:56:35.670 That's the question that I'm trying to figure out. NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100]} 00:56:35.680 --> 00:56:37.210 I mean, when I say real world, I mean I NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:56:37.210 --> 00:56:38.950 can literally give this as a research tool. NOTE CONF {"raw":[100,100,100,100,100,100,100,100]} 00:56:38.950 --> 00:56:41.950 I, in fact, some of my colleagues use it, but NOTE CONF {"raw":[58,100,100,100,100,100,100,100,100,100]} 00:56:42.490 --> 00:56:43.480 that's research tool. NOTE CONF {"raw":[100,100,100]} 00:56:43.870 --> 00:56:46.240 So that's not really a real world situation, if you NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:56:46.240 --> 00:56:46.510 will. NOTE CONF {"raw":[100]} 00:56:47.080 --> 00:56:50.680 So so I'm still figuring that out and I'm happy NOTE CONF {"raw":[100,100,100,100,100,100,100,100,100,100]} 00:56:50.680 --> 00:56:52.180 to reach out if I have an answer. NOTE CONF {"raw":[100,100,100,100,98,100,100,100]} 00:56:54.630 --> 00:56:57.120 I mean, maybe people here can teach me if there NOTE CONF {"raw":[66,66,100,100,100,100,100,100,100,100]} 00:56:57.120 --> 00:56:58.290 is something that I'm missing. NOTE CONF {"raw":[100,100,100,100,100]} 00:57:01.580 --> 00:57:02.050 Thank you. NOTE CONF {"raw":[100,100]} 00:57:02.060 --> 00:57:03.590 Excellent presentation. NOTE CONF {"raw":[100,100]} 00:57:04.400 --> 00:57:04.940 Thank you. NOTE CONF {"raw":[100,100]} 00:57:09.760 --> 00:57:10.960 I think that'll close this out. NOTE CONF {"raw":[100,100,85,100,100,100]} 00:57:10.990 --> 00:57:13.420 Dr. Llama, again, fascinating. NOTE CONF {"raw":[97,45,100,100]} 00:57:13.420 --> 00:57:15.970 And we really appreciate your your presentation. NOTE CONF {"raw":[100,100,100,100,100,100,100]} 00:57:16.240 --> 00:57:16.730 Thank you. NOTE CONF {"raw":[100,100]} 00:57:16.750 --> 00:57:17.830 Thank you for your time. NOTE CONF {"raw":[100,100,100,100,100]} 00:57:18.400 --> 00:57:19.570 Good evening, everyone. NOTE CONF {"raw":[100,100,100]} 00:57:19.690 --> 00:57:20.050 Bye. NOTE CONF {"raw":[91]} 00:57:20.590 --> 00:57:21.010 Thank you. NOTE CONF {"raw":[100,100]} 00:57:21.010 --> 00:57:21.430 Bye. NOTE CONF {"raw":[86]}