WEBVTT - Algorithm, M.D. 0:00:04.600 --> 0:00:08.320 Sleepwalkers is a production of I Heart Radio and Unusual Productions. 0:00:12.760 --> 0:00:16.840 Jan had become paralyzed through a disease process such that 0:00:17.400 --> 0:00:21.520 she couldn't move anything below her neck, and she volunteered 0:00:21.520 --> 0:00:25.759 to have the surgical procedure in which we implanted electrodes 0:00:25.760 --> 0:00:29.240 in her head and we were able to decode the 0:00:29.280 --> 0:00:31.520 signals from her brain and she was able to move 0:00:31.600 --> 0:00:37.120 a high performance prosthetic arm in hand. That's Andy Schwartz, 0:00:37.200 --> 0:00:41.080 a professor of neurobiology at the University of Pittsburgh. He 0:00:41.080 --> 0:00:43.720 helped build the technology that allowed a patient called Jan 0:00:43.840 --> 0:00:48.199 Sherman to control a robotic arm with her mind. In 0:00:48.320 --> 0:00:50.479 order to do that, Andy's team needed to give the 0:00:50.640 --> 0:00:53.240 arm away to know where Jan wanted it to move. 0:00:53.760 --> 0:00:56.440 They had to read signals directly from her brain and 0:00:56.480 --> 0:00:59.120 teach a computer to understand them. So they built a 0:00:59.160 --> 0:01:03.120 brain computer interface a b c I. We implanted electrodes 0:01:03.120 --> 0:01:05.400 in her head. You have to put on these bulky 0:01:05.440 --> 0:01:08.880 connectors with thick cables going to a bank of amplifiers 0:01:08.880 --> 0:01:11.400 and computers. It sounds like something out of a science 0:01:11.400 --> 0:01:15.440 fiction film. But once Andy could see Jan's neurons firing, 0:01:15.920 --> 0:01:20.240 reading her intentions was simpler than you might think. What 0:01:20.360 --> 0:01:23.920 we found is that the rate that these neurons fire 0:01:24.160 --> 0:01:27.120 is related to the direction that the arm moves. When 0:01:27.160 --> 0:01:29.280 you add their signals together, you can get a very 0:01:29.360 --> 0:01:34.840 precise representation of that movement. It's a very very simple algorithm, 0:01:34.880 --> 0:01:37.920 and it's like listening to a Geiger counter um and 0:01:38.120 --> 0:01:41.360 each click of the Geiger counters the same, but as 0:01:41.400 --> 0:01:45.000 you get closer to a radioactive source, those clicks come 0:01:45.040 --> 0:01:50.040 closer together. Listening for those clicks was the breakthrough necessary 0:01:50.040 --> 0:01:53.240 to translate Jan's thoughts into signals that the robotic arm 0:01:53.240 --> 0:01:56.760 could understand, and this meant that Jan could move again. 0:01:57.320 --> 0:02:01.480 She could reach out and touch her husband, and Andy's 0:02:01.520 --> 0:02:04.120 team filmed Jan using the arm to complete a more 0:02:04.160 --> 0:02:08.040 playful goal that she'd set to feed herself chocolate for 0:02:08.080 --> 0:02:16.280 the first time in ten years for a woman, one 0:02:16.440 --> 0:02:22.560 giant fight for DHDR. Jan was able to do that 0:02:22.880 --> 0:02:25.440 with the robot, and not only that, but she made 0:02:25.480 --> 0:02:29.200 graceful and beautiful movements. That's what really blew me away. 0:02:29.760 --> 0:02:32.560 It looks much like a real arm in hand, and 0:02:33.880 --> 0:02:37.720 for someone who studies movement, it was really quite beautiful. 0:02:40.040 --> 0:02:43.120 Andy is a scientist and a researcher not given to 0:02:43.160 --> 0:02:46.680 being sentimental, but the way he talks about Jan moving 0:02:46.720 --> 0:02:50.040 her robotic arm, it's like describing a dance a ballet. 0:02:50.600 --> 0:02:53.160 It's one of the most inspiring examples of humans and 0:02:53.200 --> 0:02:56.640 machines working together in concert that we've come across in 0:02:56.680 --> 0:03:00.639 all of our reporting for Sleepwalkers. But it also raises 0:03:00.680 --> 0:03:04.239 profound questions about the future of our health, our bodies, 0:03:04.280 --> 0:03:08.480 and our society. What are the implications positive and negative? 0:03:08.919 --> 0:03:12.079 As AI makes us ever more able to decode complex 0:03:12.120 --> 0:03:16.240 systems like the brain or the human genome, how is 0:03:16.280 --> 0:03:20.000 AI poised to change the world of medicine. I'm as 0:03:20.040 --> 0:03:36.640 Veloshin and this is Sleepwalkers, So Kara I found Andy's 0:03:36.640 --> 0:03:41.280 story completely mind blowing, no pun intended. You know it 0:03:41.440 --> 0:03:44.720 shakes one of the last remaining mysteries. For the most part, 0:03:44.800 --> 0:03:47.880 neuroscience is still very much a black box problem. We 0:03:47.920 --> 0:03:51.440 know what's happening in the brain, but neuroscientists can't always 0:03:51.640 --> 0:03:54.240 know why and how, which is a lot like the 0:03:54.240 --> 0:03:57.560 problem of black box AI. We are moving quickly into 0:03:57.600 --> 0:03:59.840 a world where sensors can read us better and better 0:04:00.040 --> 0:04:03.080 know they can read pupil dilation, carbon dioxide in the 0:04:03.160 --> 0:04:06.000 breath to understand people's emotions. Yeah, this is the stuff 0:04:06.040 --> 0:04:08.960 we talked about an episode four around using AI to 0:04:09.080 --> 0:04:12.240 better read biometrics with Poppy Chrum. The difference being in 0:04:12.280 --> 0:04:14.880 this case is that Andy isn't monitoring the outside of 0:04:14.880 --> 0:04:18.760 our bodies, so the privacy concern is less. To read 0:04:18.839 --> 0:04:21.800 Jan's brain, he had to drill into her skull and 0:04:21.880 --> 0:04:25.200 place electros into the surface of her brain and then 0:04:25.200 --> 0:04:28.200 connect them to a computer, so that's unlikely to creep 0:04:28.279 --> 0:04:30.680 up on you. Google is not going to be doing 0:04:30.720 --> 0:04:34.919 that to get geolocation, not yet. Because that said, and 0:04:35.040 --> 0:04:37.159 he told me one of his big priorities is trying 0:04:37.200 --> 0:04:40.200 to figure out how to achieve the same effects without 0:04:40.279 --> 0:04:43.440 needing invasive surgery, so that people can use this technology 0:04:43.520 --> 0:04:46.839 at home. Talking about reading the brain, one cutting edge 0:04:46.839 --> 0:04:50.520 application for AI is to restore language. A few months ago, 0:04:50.560 --> 0:04:52.400 I was actually reading this paper in Nature, and I'm 0:04:52.400 --> 0:04:56.040 not sure all podcast listeners read Nature. Well, thanks, Carol, 0:04:56.080 --> 0:04:57.880 you do the hard work so that we don't have to. 0:04:58.000 --> 0:05:01.320 I guess why might they pay me the big bitcoin um? 0:05:01.320 --> 0:05:03.800 But you know, basically this article was about decoding the 0:05:03.880 --> 0:05:07.919 human brain to produce language based on how the brain 0:05:07.960 --> 0:05:11.440 tells the mouth to move well. Funny enough, this is 0:05:11.480 --> 0:05:15.080 something I spoke to Andy about a few months before 0:05:15.440 --> 0:05:18.159 your paper and Nature was published, Kara, and he was 0:05:18.160 --> 0:05:21.559 talking about exactly this. So if you record from motor 0:05:21.640 --> 0:05:26.280 areas associated with producing language, you can start to recognize 0:05:26.360 --> 0:05:30.280 certain words and phrases. Even I think that's realizable in 0:05:30.279 --> 0:05:33.240 the your term. What Andy is saying is that to 0:05:33.320 --> 0:05:36.520 create spoken words directly from the brain, we don't actually 0:05:36.600 --> 0:05:38.800 need to read thoughts. We can just look at the 0:05:38.880 --> 0:05:42.440 last step, the moment when thought becomes action as neurons 0:05:42.480 --> 0:05:45.599 fired to move our tongue, our lips are jaw to 0:05:45.680 --> 0:05:49.600 create sound. And just like with jan, an algorithm can 0:05:49.680 --> 0:05:53.039 listen to those neurons and allow thoughts to become actions. 0:05:53.839 --> 0:05:57.520 And this could transform lives. So if you could start 0:05:57.520 --> 0:06:02.560 to recognize words and language from brain activity, that would 0:06:02.600 --> 0:06:05.800 be very helpful for people who are locked in with 0:06:05.880 --> 0:06:08.880 a l S and can't communicate. We talked in the 0:06:08.960 --> 0:06:12.240 last episode about how many technological breakthroughs have come out 0:06:12.279 --> 0:06:15.360 of DARPER. That's the branch of the Defense Department charged 0:06:15.360 --> 0:06:20.920 with inventing technological surprises, well neural prosthetics or robotic limbs 0:06:20.960 --> 0:06:23.520 controlled by the mind have been an area of heavy 0:06:23.560 --> 0:06:26.880 investment for the agency. You may remember Gil Pratt from 0:06:26.880 --> 0:06:29.400 earlier in the series. He's now the CEO of the 0:06:29.440 --> 0:06:33.919 Toyota Research Institute, but previously he was at DARPA where 0:06:33.960 --> 0:06:37.360 he worked with none other than Andy Schwartz. So he 0:06:37.440 --> 0:06:39.680 was involved in a project at DARPA which was called 0:06:39.720 --> 0:06:44.680 Revolutionizing Prosthetics. And the project that I started was to 0:06:45.040 --> 0:06:48.479 see if we could actually help some of the experimental 0:06:48.520 --> 0:06:52.640 patients that he had to perform even better. So Kara 0:06:52.760 --> 0:06:55.479 I was interviewing Gil Pratt about his work at Toyota 0:06:55.600 --> 0:06:58.239 and on self driving cars, and we got talking about 0:06:58.279 --> 0:07:01.320 his interest in these human machine owner ships, and I said, 0:07:01.360 --> 0:07:04.920 let me tell you a story as a scientist called 0:07:04.960 --> 0:07:08.880 Andy Schwartz and his patient Jan and Gil was like, um, yeah, 0:07:08.960 --> 0:07:13.440 I worked on that. Yeah, no, no no, no, literally, it's funny. 0:07:13.480 --> 0:07:17.040 And it also shows the long arm of DARPA intended. 0:07:17.120 --> 0:07:21.840 Again that's far away from military applications. You know, darp 0:07:21.840 --> 0:07:25.280 AS interest in prosthetics is actually because of veterans, many 0:07:25.360 --> 0:07:29.080 of whom lose limbs on the battlefield, which highlights again 0:07:29.120 --> 0:07:32.280 the dual use nature of so much innovation. Right, um 0:07:32.280 --> 0:07:34.680 and The work Gil was doing at DARPA with Jan 0:07:35.160 --> 0:07:37.320 was all about doing about a job of interpreting her 0:07:37.440 --> 0:07:42.080 brain waves using the existing models. My group made a 0:07:42.360 --> 0:07:46.200 system called arm assist that watched what Jan was trying 0:07:46.240 --> 0:07:49.280 to do. It inferred Okay, now she's trying to pick 0:07:49.360 --> 0:07:51.600 up a block. Now she's trying to move it over 0:07:51.640 --> 0:07:53.840 to the left. Now she's trying to drop the block. 0:07:54.040 --> 0:07:55.800 You know, in a way very similar to if you 0:07:55.920 --> 0:07:58.880 use power Point and you have like snapped to grit 0:07:59.040 --> 0:08:02.200 or snapped the object turned on, it will help you 0:08:02.240 --> 0:08:04.480 move the mouse to where it thinks you want to go. 0:08:05.160 --> 0:08:08.080 This system helped her move the arm to where it 0:08:08.200 --> 0:08:11.080 inferred she wanted to go based on a very noisy 0:08:11.240 --> 0:08:13.800 signal that was coming out of her brain. That noise 0:08:13.840 --> 0:08:16.520 in the signal was partly because the connection between Jan's 0:08:16.560 --> 0:08:20.440 brain and the decoding computer was weak. So Gil's team 0:08:20.440 --> 0:08:25.400 Adoppa designed a program to boost Jan's intentions. We tested 0:08:25.440 --> 0:08:29.200 her by randomly turning the assists on and off. And 0:08:29.360 --> 0:08:32.280 you can think of the assist as a guardian, like 0:08:32.360 --> 0:08:34.800 what we're developing for the car to stop you from 0:08:34.840 --> 0:08:37.959 having a crash, and we had the guardian in this 0:08:38.000 --> 0:08:42.160 case help as little as possible, but still be effective 0:08:42.160 --> 0:08:45.040 at helping her to reach the goal. The amount of 0:08:45.080 --> 0:08:48.600 help that we gave her was so low that she 0:08:48.640 --> 0:08:51.199 couldn't tell whether it was on or off. But when 0:08:51.240 --> 0:08:54.960 Guardian was turned off, Jan's success rate fell. And in 0:08:55.000 --> 0:08:58.320 a human machine partnership is not just the computer that 0:08:58.400 --> 0:09:03.080 adapts to the human brain also adapts to the algorithm. 0:09:03.120 --> 0:09:05.960 And the's algorithms are simple and they rely on human 0:09:06.000 --> 0:09:10.160 programmed rules. So when they see cause in this case, 0:09:10.320 --> 0:09:13.760 enough neurons firing at the same time, they create an effect, 0:09:14.040 --> 0:09:18.840 a movement. But what's tantalizing is that effect that output 0:09:19.200 --> 0:09:25.360 hints at something far deeper and infinitely complex personality. I 0:09:25.440 --> 0:09:28.960 like Moby Dick where he talks about Captain Ahab walking 0:09:29.000 --> 0:09:31.800 on the deck of the ship, and by observing his movements, 0:09:31.840 --> 0:09:34.679 you can really understand what he's thinking about. If you 0:09:34.760 --> 0:09:39.400 think about movement, it's really a communication between your innermost 0:09:39.440 --> 0:09:42.600 thoughts in the outside world. Andy and his team saw 0:09:42.640 --> 0:09:45.560 this come to life before their very eyes when they 0:09:45.600 --> 0:09:50.559 connected another subject, Nathan, to a neural prosthesis. The robotic 0:09:50.679 --> 0:09:53.880 arm moved in accordance with the personality of the user 0:09:54.480 --> 0:09:57.440 Jan when she moved. She was very careful and gentle, 0:09:58.160 --> 0:10:01.240 and Nathan is more are of a video gamer. He's 0:10:01.280 --> 0:10:03.840 a younger guy. He's more of a competitor, so he 0:10:03.880 --> 0:10:07.880 would move much faster, and so he would pick up 0:10:07.880 --> 0:10:10.600 an object and then instead of placing it carefully in 0:10:10.600 --> 0:10:13.240 a receptacle, he would basically toss it into the receptacle 0:10:13.840 --> 0:10:16.920 to be faster. Today we can see the difference in 0:10:17.040 --> 0:10:20.800 Nathan and Gen's personalities from how the brain controls a 0:10:20.920 --> 0:10:25.720 prosthetic arm. We can infer personality from output, but we 0:10:25.760 --> 0:10:29.680 don't even have sophisticated enough tools to ask why. As 0:10:29.720 --> 0:10:32.880 we get better and better at these computational approaches, will 0:10:32.960 --> 0:10:36.880 gain a much better understanding of the way the brain works. 0:10:37.000 --> 0:10:43.360 Rather than having some major event causing some simple consequence. Instead, 0:10:43.679 --> 0:10:46.880 it's more like a perfect storm where many factors come 0:10:46.920 --> 0:10:52.079 together to generate a consequence. And if we understand which 0:10:52.120 --> 0:10:56.000 of those events are important and how they're combined, then 0:10:56.080 --> 0:11:00.960 we can start to understand brain function better. If you 0:11:01.000 --> 0:11:04.400 see a cup and you're thirsty, if we could realize 0:11:04.440 --> 0:11:06.719 that what you want to do is to drink from it, 0:11:07.400 --> 0:11:10.760 then we could understand that you want to grasp the 0:11:10.800 --> 0:11:14.200 cup from the side, and if we could distinguish that 0:11:14.360 --> 0:11:17.800 from you grasping a cup with the intention of passing 0:11:17.800 --> 0:11:20.079 it to me, you might hold it now from the top, 0:11:20.440 --> 0:11:23.079 then we could do a better job generating the correct 0:11:23.360 --> 0:11:27.360 movement for Andy. The next frontier is more deeply understanding 0:11:27.400 --> 0:11:31.240 the human brain carrot by developing new models and better algorithms. 0:11:31.559 --> 0:11:34.120 And that's really a computer science problem as much as 0:11:34.120 --> 0:11:36.880 a medical problem. And then really the next frontier in 0:11:36.920 --> 0:11:40.800 medicine is all about using AI to decode complex interactions. 0:11:41.200 --> 0:11:45.319 And in fact, you reported a piece on exactly this phenomenon. Yeah, 0:11:45.400 --> 0:11:48.319 I spoke to a woman named Regina bars Ali. She's 0:11:48.360 --> 0:11:51.600 a computer scientist m I T and her own experience 0:11:51.600 --> 0:11:54.400 of how she was diagnosed with breast cancer actually inspired 0:11:54.480 --> 0:11:58.000 her to work on bringing AI into the realm of medicine. 0:11:58.720 --> 0:12:01.400 When we come back, we'll here from Regina and take 0:12:01.440 --> 0:12:04.760 a look at other ways AI is changing diagnostics and 0:12:04.840 --> 0:12:15.040 the future of our health. I remember still, I went 0:12:15.080 --> 0:12:18.480 to mamogram and they told me a high density but 0:12:18.559 --> 0:12:22.199 you shouldn't worry. Half of the population have high density breasts, 0:12:22.240 --> 0:12:26.079 so don't worry about it. That's Regina bars Lie. Regina 0:12:26.160 --> 0:12:28.960 is a professor of Electrical Engineering and Computer Science at 0:12:29.040 --> 0:12:31.520 m I T. I couldn't believe in it. I didn't 0:12:31.559 --> 0:12:33.960 feel anything. There was nothing wrong, you know. I was 0:12:34.160 --> 0:12:37.720 continuing my morning runs and being fine. It was later 0:12:37.760 --> 0:12:40.720 on after her mammogram that Regina found out she had 0:12:40.760 --> 0:12:44.280 breast cancer, and as it often does, the news came 0:12:44.280 --> 0:12:47.080 as a surprise. And if I'm looking at myself, I 0:12:47.120 --> 0:12:51.760 really cannot explain what was wrong that I got this 0:12:51.920 --> 0:12:55.560 disease that clearly didn't have any family history. I'm exercising, 0:12:55.640 --> 0:13:00.000 I'm meeting healthy and for many many patients that I'm 0:13:00.000 --> 0:13:03.560 me during my own journey, UH, their diagnosis came to 0:13:03.600 --> 0:13:07.120 them as the biggest surprise. The thing is, according to 0:13:07.160 --> 0:13:09.960 current medical standards, breast cancer risk is based on a 0:13:10.000 --> 0:13:13.000 few factors. Are you a woman, and are you old? 0:13:13.200 --> 0:13:14.880 Do you have the Brocko gene? And do you have 0:13:14.960 --> 0:13:18.240 breast cancer in your family? But those are relatively simple 0:13:18.280 --> 0:13:21.320 inputs and they don't account for what complex systems we 0:13:21.360 --> 0:13:25.080 are above eighty percent of women who diagnosed with breast 0:13:25.080 --> 0:13:29.680 cancer there first in their families, so it's not clear 0:13:30.040 --> 0:13:33.240 you know what causes it. According to the Susan Becoman 0:13:33.280 --> 0:13:37.160 Breast Cancer Foundation, breast cancer is the most common cancer 0:13:37.280 --> 0:13:40.560 amongst women around the world. Every two minutes a case 0:13:40.559 --> 0:13:42.559 of breast cancer is diagnosed and a woman in the 0:13:42.640 --> 0:13:46.200 United States, and every minute somewhere on earth a woman 0:13:46.240 --> 0:13:50.079 dies of breast cancer. That's more than four hundred women 0:13:50.120 --> 0:13:52.880 per day. So if you're listening to this podcast, you 0:13:53.000 --> 0:13:56.079 probably know someone who has been or will be diagnosed 0:13:56.080 --> 0:13:59.200 with breast cancer in their lifetime. And with that many 0:13:59.240 --> 0:14:02.040 people affected that they are bound to be some oversights, 0:14:02.080 --> 0:14:06.200 like in Regina's case. I discovered that my own diagnosis 0:14:06.320 --> 0:14:10.640 was delayed because the malignancy was missed in the previous mammograms. 0:14:10.960 --> 0:14:14.720 And I also discovered that this is not a unique experience. 0:14:15.080 --> 0:14:18.319 So the question I asked is is it possible for 0:14:18.440 --> 0:14:22.480 us to know ahead of time what's to come? In 0:14:22.520 --> 0:14:26.200 other words, can we look into the future prior to 0:14:26.200 --> 0:14:29.080 our diagnosis. Regina had already been a long time computer 0:14:29.120 --> 0:14:31.920 scientists at M I T. She thought often about machine 0:14:32.000 --> 0:14:35.560 learning in her work, but this new personal hardship redirected 0:14:35.560 --> 0:14:39.600 her thinking. When you go to the hospital, you see 0:14:39.680 --> 0:14:42.520 like real pain of other people. You see people who 0:14:42.560 --> 0:14:46.560 go through key most radiation, even though the hospital just 0:14:46.640 --> 0:14:49.760 want stop away from M I T. I just was 0:14:49.800 --> 0:14:52.640 not aware there is so much suffering. And at that 0:14:52.720 --> 0:14:55.760 point when I came back, I was thinking, we create 0:14:56.040 --> 0:14:59.120 so much you know, exciting new technology it and I 0:14:59.240 --> 0:15:03.200 see why I we are not trying to solve it. 0:15:03.200 --> 0:15:07.440 It's it's just a travesty. And so Regine and her 0:15:07.480 --> 0:15:09.960 colleagues that M I T began training a deep learning 0:15:09.960 --> 0:15:13.600 model on over ninety thod mamograms from mass General Hospital, 0:15:14.240 --> 0:15:16.400 and with such a large data set, they were able 0:15:16.440 --> 0:15:19.520 to predict a patient's risk of breast cancer by comparing 0:15:19.560 --> 0:15:23.640 one mammogram to tens of thousands of others instantly. My 0:15:23.800 --> 0:15:27.800 firm belief was that despite the standard risk factors, there 0:15:27.840 --> 0:15:31.800 is a lot of information in women's breast. Human eye 0:15:31.880 --> 0:15:35.200 which even seen you know, thousands or tens of thousands 0:15:35.280 --> 0:15:39.120 images over their lifetime, may not be really able to 0:15:39.200 --> 0:15:43.080 detect it with a great clarity. However, if machine which 0:15:43.160 --> 0:15:46.560 is uh you know, trained on sixty images where it 0:15:46.720 --> 0:15:51.600 knows the outcomes, it can identify the difference in pixel 0:15:51.680 --> 0:15:56.760 distribution that I likely correlate to a future things that 0:15:57.000 --> 0:16:00.400 may come. The amount of data which you can train 0:16:00.400 --> 0:16:04.280 a computer on versus a doctor is massive and the 0:16:04.400 --> 0:16:07.200 AI was able to detect smaller details than the human 0:16:07.280 --> 0:16:09.600 I could pick up, so this does feel like a 0:16:09.680 --> 0:16:15.080 perfect application for the strengths of machine learning. Fifteen months 0:16:15.080 --> 0:16:18.960 ago we put first our density model, which is every 0:16:19.000 --> 0:16:22.520 mammograms that goes through months general it shows a prediction 0:16:22.600 --> 0:16:27.760 to the radiologists and around of the times radiologists degree 0:16:27.920 --> 0:16:30.960 with the machine, and when they disagree, the more experienced 0:16:31.000 --> 0:16:35.440 radiologists typically sides up with the machine. So Regina's early 0:16:35.480 --> 0:16:38.040 efforts are doing really well, and the hope is that 0:16:38.080 --> 0:16:41.000 more hospitals around the country will begin using these models 0:16:41.000 --> 0:16:43.920 for early detection and risk assessment. And that's not only 0:16:43.960 --> 0:16:48.880 because understanding risk is important, but also because misdiagnoses have 0:16:49.040 --> 0:16:53.120 led to unnecessary surgeries. I read this book The Emperor 0:16:53.240 --> 0:16:55.920 Whole Melodies. It's a book about comes, so I really 0:16:56.000 --> 0:16:59.920 like let me just say one particular moment it really 0:17:00.080 --> 0:17:04.760 choices me out. It's about how men, male surgeons who 0:17:04.760 --> 0:17:09.359 are treating women with the surgeries this firmly believe that 0:17:09.480 --> 0:17:12.280 the more you cut out of you know, women's body, 0:17:12.920 --> 0:17:16.720 the beta is w likelihood. In the early twentieth century, 0:17:16.760 --> 0:17:20.480 there was an insurgence of doctors performing radical mastectomies, removing 0:17:20.520 --> 0:17:25.080 the entire breast, thus permanently disfiguring a patient, and radical 0:17:25.160 --> 0:17:28.680 mastectomies became the norm for much of the twentieth century. 0:17:28.840 --> 0:17:31.680 The reasoning was that removing a lump leaves the risk 0:17:31.720 --> 0:17:35.000 of tumors growing elsewhere, so why take the risk. By 0:17:35.040 --> 0:17:38.280 the nineteen eighties, it was clear that radical mastectomies weren't 0:17:38.280 --> 0:17:42.320 actually in effective treatment for many patients, but according to Regina, 0:17:42.680 --> 0:17:45.120 surgeons still remove too much tissue out of an abundance 0:17:45.160 --> 0:17:48.560 of caution. The reason it happens, it's not because there 0:17:48.640 --> 0:17:51.040 is some evil doctors that wants to look another surgery. 0:17:51.080 --> 0:17:53.880 It's just because people are uncertain and it's high risk, 0:17:54.000 --> 0:17:56.639 and many times conceptician would say, I am ready to 0:17:56.680 --> 0:18:01.160 go for the harshest treatment to minimize that chances. So 0:18:01.320 --> 0:18:05.040 what we demonstrated that with machine learning you can actually 0:18:05.520 --> 0:18:09.000 identify a city percent of the populations it doesn't need 0:18:09.280 --> 0:18:13.560 this type of surgery. Regina told me that had these 0:18:13.560 --> 0:18:15.480 deep learning models been in place at the time of 0:18:15.520 --> 0:18:18.760 her early mamograms, she might have detected her risk two 0:18:18.840 --> 0:18:22.199 years sooner. And in many cases, early detection makes a 0:18:22.240 --> 0:18:25.000 big difference in how a patient chooses to treat their cancer. 0:18:25.760 --> 0:18:28.640 Since developing the breast cancer detection model, Regina is now 0:18:28.680 --> 0:18:31.320 co leading m I T S J Clinic, which is 0:18:31.320 --> 0:18:34.359 a new initiative focusing on machine learning and health. What 0:18:34.600 --> 0:18:39.240 I hope that we as a society advance since then 0:18:39.680 --> 0:18:42.000 and we are ready to bring you know, the recent 0:18:42.520 --> 0:18:46.399 science and help women, and even if it means that 0:18:46.480 --> 0:18:51.560 we need to change our beliefs about how risk assessment works. 0:18:52.080 --> 0:18:54.280 And Regina hopes that as a society we can move 0:18:54.320 --> 0:18:57.400 toward greater acceptance of using machine learning to enhance medicine. 0:18:58.160 --> 0:19:01.520 Whichever one does a bit, Joel, that one should prevail. 0:19:05.520 --> 0:19:08.800 So big question, how do you feel, Kara about putting 0:19:08.800 --> 0:19:11.280 your health into the hands of an algorithm? You know, 0:19:11.359 --> 0:19:15.840 after speaking to Regina, who told me that she probably 0:19:15.840 --> 0:19:18.680 would have been diagnosed two years sooner using her models, 0:19:20.040 --> 0:19:23.880 it seems as though machine learning provides this unparalleled form 0:19:23.880 --> 0:19:28.520 of detection right because the most seasoned doctors simply can't 0:19:28.640 --> 0:19:31.160 compare thousands of data points at once. I don't think 0:19:31.200 --> 0:19:34.600 the issue is algorithms replacing doctors. It's more a matter 0:19:34.640 --> 0:19:37.000 of equipping doctors the sharper tools that they can do 0:19:37.040 --> 0:19:40.159 their jobs. They've got to provide a patient with information 0:19:40.320 --> 0:19:44.040 and allow that patient to make informed decisions. Algorithms don't 0:19:44.080 --> 0:19:46.200 have a bedside manner. You know. What you say about 0:19:46.200 --> 0:19:48.560 the shop of tools reminds me of our conversation about 0:19:48.560 --> 0:19:53.320 creativity and episode two, using AI to give artists, musicians, 0:19:53.680 --> 0:19:57.720 screenwriters new tools to do better work. At the same time, 0:19:58.240 --> 0:20:00.879 just like the art world, the medical profession sits on 0:20:00.920 --> 0:20:03.720 this enormous pedestal where we have to trust what they 0:20:03.800 --> 0:20:05.879 say because most of us don't have the tools to 0:20:06.000 --> 0:20:09.840 question them, right unless you're on WebMD. Yeah, the bane 0:20:09.840 --> 0:20:11.960 of every doctor's life. Yeah, you know, doctors have a 0:20:11.960 --> 0:20:14.960 hard enough time explaining medicine to patients. Imagine them having 0:20:14.960 --> 0:20:17.920 to explain artificial intelligence. Well, from experience, we can say 0:20:17.920 --> 0:20:21.520 good luck to him. What's crazy to me actually is 0:20:21.560 --> 0:20:24.760 that Regina and her co author, this woman, doctor Connie Lehman, 0:20:24.760 --> 0:20:28.359 who's a radiologist at Mass General, were rejected from every 0:20:28.400 --> 0:20:30.879 single federal grant they submitted at first. Why would they 0:20:30.880 --> 0:20:33.880 be rejected from all those federal grants? Because as much 0:20:33.920 --> 0:20:36.760 as there is a ton of buzz surrounding AI, I 0:20:36.800 --> 0:20:40.320 think people have to appreciate how new this frontier is right, 0:20:40.560 --> 0:20:44.360 and using machine learning to make predictions about people's cancer 0:20:44.960 --> 0:20:47.240 is very, very new, and it's going to take doctors 0:20:47.280 --> 0:20:49.199 a really long time to learn how to convey this 0:20:49.240 --> 0:20:52.800 information to patients. So that's where Regina is right now, 0:20:53.200 --> 0:20:57.639 figuring out how doctors can explain to patients, Hey, AI 0:20:57.760 --> 0:21:00.439 helped us determine your cancer risk. Well, part of problem 0:21:00.480 --> 0:21:03.400 is that we don't yet have explainable AI. So it's 0:21:03.440 --> 0:21:05.280 not just that it's hard to explain it to patients, 0:21:05.320 --> 0:21:09.160 it's actually a black box. You may remember Sebastian Thrun, 0:21:09.320 --> 0:21:11.960 the founder of Google X from earlier in the series. 0:21:12.520 --> 0:21:15.520 As well as self driving cars and flying cars. He 0:21:15.600 --> 0:21:20.679 works on medical diagnostics and he recognizes a problem. One 0:21:20.720 --> 0:21:23.399 of the conundrums of machine learning is that when you 0:21:23.440 --> 0:21:25.520 open up in doing that brook you look at like 0:21:25.680 --> 0:21:28.680 hundreds of millions of numbers, but he can't quite understand 0:21:28.800 --> 0:21:31.359 what's happening. So people are are concerned. People look at 0:21:31.359 --> 0:21:33.359 that Brooks say, add the thing of diagnals and cancer, 0:21:33.760 --> 0:21:36.560 what does it do. We've talked about the black box 0:21:36.640 --> 0:21:39.200 problem in AI and how hard it is to trust 0:21:39.240 --> 0:21:42.520 decisions that can't be explained, but Sebastian is quick to 0:21:42.520 --> 0:21:46.000 point out that human beings can also be difficult to decipher. 0:21:47.000 --> 0:21:50.520 Let's remind ourselves our doctors are also black boxes. You 0:21:50.560 --> 0:21:52.680 can't open up the brain of your doctor and ask 0:21:53.119 --> 0:21:55.959 what was this here she using for diagonals and cancer. 0:21:56.520 --> 0:21:58.719 It's a fair point, and it's one has also been 0:21:58.800 --> 0:22:02.080 noted by Siddathan Kache. He is one of the world's 0:22:02.080 --> 0:22:05.879 foremost cancer doctors and the Pulitzerprise winning author of The 0:22:05.920 --> 0:22:09.199 Emperor of All Maladies. That's the book that Regina referred 0:22:09.200 --> 0:22:12.960 to earlier, and Siddartha has also written extensively about how 0:22:13.000 --> 0:22:16.359 AI is changing medicine. I've been a huge fan of 0:22:16.359 --> 0:22:18.959 his work for a long time and actually ambushed him 0:22:18.960 --> 0:22:21.440 after a talk he gave in order to persuade him 0:22:21.440 --> 0:22:24.879 to an interview for this podcast. So thinking about AI 0:22:24.960 --> 0:22:28.720 helping diagnose patients, it's worth asking is the human doctor 0:22:28.880 --> 0:22:31.960 so very different? One problem that I think is fascinating 0:22:32.080 --> 0:22:35.000 is when a patient comes into the hospital. If you 0:22:35.080 --> 0:22:39.760 ask a particularly astute physician, that physician can actually describe 0:22:39.800 --> 0:22:41.960 to you what the most likely journey of that patient 0:22:42.040 --> 0:22:44.800 will be in the hospital, whether they're likely to stay 0:22:44.800 --> 0:22:47.800 for twenty five days suffered through bacterial stepsis, you know 0:22:47.840 --> 0:22:50.160 all from peeking in through the door of in emriency room, 0:22:50.400 --> 0:22:54.440 Siddartha started to wonder how doctors make those lightning fast calls, 0:22:54.920 --> 0:22:57.840 and it got him interested in understanding what the brain 0:22:57.920 --> 0:23:01.080 is doing when a doctor makes a die ignosis. We 0:23:01.119 --> 0:23:05.120 actually understand very little about how human beings make diagnoses. 0:23:05.160 --> 0:23:07.480 I mean, the studies that have been done so far 0:23:07.720 --> 0:23:10.840 suggest that most people make diagnosis in a kind of 0:23:11.480 --> 0:23:15.560 recognition sense rather than an algorithmic sense. The classical description 0:23:15.600 --> 0:23:18.760 of how we make diagnosis was extraordinarily algorithmic, sort of 0:23:18.800 --> 0:23:22.360 goes down a series of elimination. It's not this, it's 0:23:22.400 --> 0:23:25.440 not that. Now. Whence the alto says algorithmic, he doesn't 0:23:25.480 --> 0:23:29.480 mean using a computer algorithm. He means using rule based logic. 0:23:29.960 --> 0:23:33.600 If this, then that, et cetera. But what he learned 0:23:33.680 --> 0:23:36.840 was that despite what the textbooks say, that's not actually 0:23:36.840 --> 0:23:40.480 how doctors make a diagnosis. When you put doctors inside 0:23:40.600 --> 0:23:44.040 MRI machines and ask the question how do they make diagnosis? 0:23:44.040 --> 0:23:47.280 In fact what lights up is parts of the brain 0:23:47.359 --> 0:23:50.359 that are much much more to do with pattern recognition. 0:23:50.680 --> 0:23:53.640 Here's a rhinoceros, Here's not a rhinoceros. Here's an elephant, 0:23:53.680 --> 0:23:58.439