The Importance of Critical Thinking in AI-Powered Healthcare Systems

When it comes to AI-powered healthcare systems, there is a growing concern about the potential for bias and misinformation to affect patient care. One approach that has been explored is providing descriptive advice to doctors, which can highlight the risks and challenges associated with certain diagnoses or treatments. However, research has shown that even descriptive systems can be problematic if they are not carefully designed.

For example, studies have found that when AI-powered advice is presented in a prescriptive manner, it can lead to biased decision-making by doctors. For instance, if an AI system advises doctors to call the police on minority patients consistently, without considering the specific context or circumstances of each case, this can perpetuate existing biases and inequalities in healthcare. In contrast, descriptive systems that highlight the risks and challenges associated with certain diagnoses or treatments may be more effective at promoting critical thinking and nuanced decision-making.

In health care specifically, we need to think carefully about not only what predictions are being made but also exactly how we deliver those predictions and what we find is with these descriptive systems you're giving them the rules and they're able to follow through on these judgments and have really good critical thinking skills but when we give them the end result just shortcut to the answer they don't actually go through the process and so they miss mistakes that the model might make.

For instance, if a doctor is given an AI-powered diagnosis of a patient's condition without being shown the underlying data or reasoning, they may not be able to critically evaluate the advice and make their own informed decisions. This can lead to missed opportunities for nuanced and personalized care, as well as perpetuation of existing biases and inequalities in healthcare.

To address these challenges, researchers are exploring the development of normative solutions that provide guidelines for best practices in patient care. These solutions are based on established care pathways and evidence-based medicine, and aim to reduce variation in how different patients are treated. While this approach may not always be successful, as there is huge variation one person's diabetes is not another person's diabetes one person's pregnancy is not a person's pregnancy, it can help to promote more equitable and effective healthcare outcomes.

In addition to descriptive, prescriptive, and normative solutions, researchers are also exploring the use of self-supervised learning techniques to improve the accuracy and fairness of AI-powered healthcare systems. For instance, by analyzing differences in how different models view data, researchers hope to identify biases and errors that may not be apparent through traditional supervised learning methods.

Furthermore, researchers are working to develop systems that can flag incorrect data from the get-go in large datasets, without relying on human oversight or annotation. This approach has significant implications for healthcare, as it could help to prevent the spread of misinformation and improve the overall quality of patient care.

Finally, researchers are exploring ways to deploy AI-powered healthcare systems in a way that promotes critical thinking and nuanced decision-making. By providing clinicians with access to transparent and auditable data, and by enabling them to evaluate and challenge AI-generated advice, researchers hope to create more equitable and effective healthcare outcomes.

The future of AI-powered healthcare systems holds much promise, but it also raises significant challenges and complexities. As technical contributors to these systems, we have a critical role to play in promoting fairness, transparency, and accountability. By working together to address these challenges, we can create systems that truly support the needs of patients and clinicians alike.

In our lab, we are excited about the potential for AI-powered healthcare systems to improve patient care and outcomes. One area of focus is on developing systems that can flag incorrect data from the get-go in large datasets, without relying on human oversight or annotation. This approach has significant implications for healthcare, as it could help to prevent the spread of misinformation and improve the overall quality of patient care.

We are also exploring ways to deploy AI-powered healthcare systems in a way that promotes critical thinking and nuanced decision-making. By providing clinicians with access to transparent and auditable data, and by enabling them to evaluate and challenge AI-generated advice, we hope to create more equitable and effective healthcare outcomes.

In addition, researchers are working to develop normative solutions that provide guidelines for best practices in patient care. These solutions are based on established care pathways and evidence-based medicine, and aim to reduce variation in how different patients are treated. While this approach may not always be successful, as there is huge variation one person's diabetes is not another person's diabetes one person's pregnancy is not a person's pregnancy, it can help to promote more equitable and effective healthcare outcomes.

The development of AI-powered healthcare systems also raises significant ethical considerations. As clinicians and technical contributors, we have a critical role to play in promoting fairness, transparency, and accountability. By working together to address these challenges, we can create systems that truly support the needs of patients and clinicians alike.

Overall, the future of AI-powered healthcare systems holds much promise, but it also raises significant challenges and complexities. As technical contributors to these systems, we have a critical role to play in promoting fairness, transparency, and accountability. By working together to address these challenges, we can create systems that truly support the needs of patients and clinicians alike.

By exploring descriptive, prescriptive, normative, and self-supervised learning techniques, researchers hope to develop AI-powered healthcare systems that promote critical thinking, nuanced decision-making, and more equitable care outcomes. These approaches have significant implications for healthcare, as they could help to prevent the spread of misinformation and improve the overall quality of patient care.

In conclusion, the importance of critical thinking in AI-powered healthcare systems cannot be overstated. By promoting transparent and auditable data, enabling clinicians to evaluate and challenge AI-generated advice, and developing normative solutions that provide guidelines for best practices in patient care, researchers hope to create systems that truly support the needs of patients and clinicians alike.

"WEBVTTKind: captionsLanguage: enhello everyone and welcome back to Cale office hours I am Daniela ruse the director of Cale where we are inventing the future of computing and making the world better through Computing today our Focus will be on health and how Computing and AI can help with diagnosing monitoring and treating disease Dina you have made so many impactful contributions to communication and to understanding Wireless signals how did you think to move from studying networking to applying Wireless signals to the healthcare sector maybe about eight years ago we started looking into how we can use um radio signals to monitor physiological signals so that was the early set of results basically just understanding that radio signal bounce around and they bounce out the human body because we are made of water so signal comes and reflects and then if you analyze those Reflections particularly if you analyze them with the power of neural network in machine learning you can start learning things that you couldn't imagine that these radio signals around us have or carry about us such as of course our respiration the pulsing of our blood even when you go to sleep we can tell whether you are in the dreaming stage for example ra versus you are in light sleep or deep sleep and then that took us to the Next Generation not just monitoring the physiological signals of people using radio signals without touching them but also collecting that data and starting to develop neural network model for uh diagnosing and tracking like what's called biomarkers of diseases for Parkinson's for other newgeneration diseases and the final the most recent thing is Gen AI for uh Medical Data with your algorithms and your insights in Wireless networking you can see what we cannot see with a naked eye how does this work it takes many years to diagnose Parkinson like somebody may had Parkinson 10 years ago and they haven't been diagnosed because one of the issue is that you wait for the motor symptoms you know in Parkinson the Tremor the stiffness of the motion and these motor symptoms appear late in the Disease by the time the motor symptoms appear the uh the dopamine neurons which are the neurons involved in in Parkinson about 40 to 80% are already damaged when we learned this we were just like oh but can we diagnose it from other signals now that we have these devices in people's home and we just take the wireless signal maybe from warless we know that we can get people breathing let me just try to detect whether they have parking from their breathing because there are some hypotheses that breathing like is impacted by the uh basically that Parkinson comes from the GS along the vagus nerve and then it has the breathing area in the brain so breathing changes earlier but actually it turned out that you can diagnose or machine learning algorithm can diagnose Parkinson with high accuracy from someone's breathing what is it that the intersection of understanding Wireless signals and understanding machine learning what is it at this intersection that helps you solve this problem we are not trained to look at radio signals but machine actually when you train them with the right output they they first they can sense the signal and they can start seeing patterns that the human brain is not trained to see from sensory signals that we don't have like electromagnetic waves where does this data come from the data that you use to train your models we designed a device that looks very much like your Wi-Fi box at home so sit in the background of the home and it's like analyzing these radial reflection transmit very low power Wireless signal analyzing the reflection and analyze them neural network and from that it can be like you don't have to wear devices on yourself people can live their lives and you can just get all of that insight to clinical data but it also convey the information to the medical doctor to the practitioner who's the expert and can help them can this model monitor for multiple diseases or is it finally tuned for Parkinson's U monitors multiple diseases particularly interested also I find the the biggest opportunity is in neurology and Immunology so it's very hard to to study neurological diseases even things like depression we ask people how do you feel but give you like a drug or treatment he said there's no actual concrete measurements to do that that's where we are seeing like our results from being able to concretely and objectively measure the physiology of the person and how it changes we can see a major uh opportunity to have a um what's called biomarket in this context digital biomarker for these complex diseases can you tell us a little bit more about Immunology what are you looking at in the space one essential thing of the immune system is that you are talking about a dynamic system a that is changing like it's just your immune response it's a response flare up there are flares you change you are in remission if you have like an autoimmune disease you are in remission and then you have a flare but in the whole medical system we don't have a the like all of the tests that we do are snapshots like when we talk in computer science we talk about snapshots ver versus like continuous tracking we don't have the continuous tracking and clinical data for that continuously can track a dynamic system like the immune system there is a hugee opportunity here to be able to move from snapshots particularly when you are talking about inflammatory response immune response to continuous data and where we can see where we are using our devices and Ai and to track these changes so are you finding that the Wi-Fi signals reflect differently if there is an immune system flare up when they bounce off my body your body the environment they change based on everything your respiration the pulsing of the blood the twitching of your eyes like different muscles all of the behavior that we have and things that we don't even think as of as Behavior so the physics of how the reflections happen don't change but when you extract this information from those Reflections now those change your respiratory system response to inflammation is going to be different your sleep and sleep patterns will change if somebody has a particular disease or taking particular drug so those then you can think about Ai and two levels one level is like taking radio signals and extracting from the physiological signals and then another level of neural networks that takes those physiological signal continuously extract it from the radio signal and start talking does this person have Alzheimer does this person have Parkinson how did their Parkinson change are they taking anti-depressant are they responding to is anti-depressant how do you use generative AI to get to better results being able to to take some simple signals or let's say like something from your Fitbit or some like your your pulse or something and transform that using generative AI to clinical test that was the use of generative AI that we are looking at and we actually do have initial data that shows that this is doable and the generated data is very much similar to the actual real data from the medical test for that person your generative AI system for medical data is it um what is the raw data used to train it we move from one medical modality that is cheap to a medical modality that is expensive a cheap medical modality it's it's kind of like a wellness modality like if I take the the PPG from uh like the puling from like AR device or uh the breathing from radio signal these are cheap available and easily and from that I generate something that is expensive like an EEG signal for example for the brain Dina it sounds like you have invented a new technology that is noninvasive and that gives doctors a kind of a superpower because it allows them to see things that are impossible to see with today's Technologies medicine I think there's a huge potential for competition in general I think even without a there are plenty of opportunity for competition but with machine learning also the the opportunities go even higher thank you for putting so much thought and effort in making healthc care better let's go meet MIT professor John gag John has been working with artificial intelligence and machine learning to develop the field of healthcare information and he has been looking at how machine learning can enhance the diagnosis of cardiac disease my first work in healthc care and AI was in in cardiovascular medicine and I had the good fortune that Colin staltz who's a member of cesil and a faculty member in eecs is also a practicing cardiologist Colin and I have been looking a lot at can we use non-invasive signals to get good approximations of data that is usually gathered invasively so the best example is probably an ECG measures the electrical activity of the heart every patient in the hospital is typically has an ECG connected to them what's been interesting is the things you might not expect to be able to get out of that signal but we think you can with the right kind of Technology machine learning machine learning to learn it but then of course when you're running it in practice you have to have forward passes that run fast the role of machine learning is to train a model that will give the right predictions is that right we use machine learning in healthc care in a lot of different ways sometimes it's used to train models that we'll actually deploy but other times we use it to do what I'll call science and and learn about things and once we've learned about them then the model maybe you don't need it anymore you you you've learned something you've done some experiments it's helped you learn things if we think of experience learning as you generating hypotheses that can then be tested in clinical settings or in the lab then you don't have to trust the model because all the model is doing is giving you things that you're going to test once the model generate finds hypotheses you can then maybe look for uh mechanistic explanations or look for rules maybe it's just discovering a rule and you can explain the rule to somebody and oh yeah that's a sensible Rule and so I think a lot of the use of machine learning today in healthcare is in the science part of it where we're discovering things that can then be used in the clinic we started with communicable diseases and we did that in conjunction with MGH with the infection control Department there so the infection control Department's responsibility is to reduce the incidence of healthc care Associated infections in the hospital we were approached to look at the question of could we use machine learning to find ways to reduce the prevalence of Health Care infections in in the hospital and the one we focus on to start with was uh CI the most common one it's an intestinal disease bacterial disease our models predicted which patients were most likely to contract the disease if you know that someone is likely to contract the disease you'll be more aggressive about knowing when to test them and that will let you discover they have it sooner and our treatment sooner and so what we did is we built models that were quite effective at determining which members of the population were most vulnerable so if we look at what makes people susceptible uh some of what we found was already known when the uh ant some antimicrobials uh increase the likelihood of Contracting sedi and some more than others we found zip code was important now he said how on Earth can where a patient lives matter some of what they think of Hospital acquired was actually the bacterium was already in the patient system and then something that happened in the hospital allowed it to flourish and get worse and so the bacterium came from the community and it was say the treatment in the hospital that made it flourish suddenly it made sense and we looked at these zip codes and for example there were often zip codes that had a lot of nursing homes in them so John for this uh work in communicable diseases can you tell us a bit about the techniques that you have used about the computational aspects of the solutions one of the things that's interesting about a communicable disease is is the fact that you need exposure and that means you have to look at Network effects and a lot of machine learning algorithms don't really deal with Networks and so a lot of what we did was was looking at at connectivity networks to try and and and build models based upon that and you use machine learning for that yes how did you train your models we were able to get access to all of the MGH patient records you can look at the patient record and see which nurses are visiting which patients which doctors are visiting which patients where they are we actually looked at the architecture of the hospital but the most important network was the caregiver Network what are you working on right now the biggest investment in my lab right now is in medical imaging and machine learning has is in the process of totally revolutionizing Medical Imaging so if you think of the amount of data produced by an MRI it's three-dimensional high resolution typically and it's more data than a human can possibly look at there are a small number of tasks that you have to do for any kind of image one of them is segmentation so that involves you take an image and you find Regions that are anatomically significant and separate them from the rest of the image so segmentation is very important but not just for things like tumors for a lot of say neurological diseases you want to look at the shape of say the hippocampus you know what is this exact shape is the shape changing and to do that you have to find it and segment it very precisely we have built something which we call univers EG Universal segmentor it's important because the other Universal segmentors that are out there and there quite a few work only for natural images really they don't work very well for medical images and you need something that's very Broad because there are an enormous number of different modalities ultrasound to x-ray to to Mr to Optical to ooc all sorts of things he collected an enormous amount of data from many different Medical Imaging modalities many different tasks we then invented ways to generate synthetic data if you look at our synthetic data it doesn't look like any medical image anyone has ever seen or any task anyone has ever seen but it makes it prevents the model from overfitting to the actual tasks you show it during the training other thing we've done with Universe egg is is it's what's called context based learning at inference time when someone is trying to say segment something that they don't know how to segment what we do is you give it a small number of examples and that's used as a prompt to say oh that's the kind of that's what they're looking for and then based upon that small context and all of the other examples that's been trained on it then can generate quite good segmentations and we think eventually will be as good right now if you train a a one of model for one modality and one Anatomy it's slightly better than the universal segmentor as you think about computational tools and their potential in healthare where do you see us going in the next 3 to 5 years the large population studies will be the first thing impacted I think by a lot of the machine learning technology and I think it'll have enormous impact on things like tele medicine instead of having a person move to the clinic we acquire data and send the data to a computer and the data the computer looks at it and I think we can make very quick inroads in using machine learning to build models that will improve the health care system if Health Care gets less expensive we'll get better outcomes an awful lot of people don't get the treatment they need today because they can't afford it if we can reduce the cost it isn't just saving money it will give us better outcomes John thank you so much for sharing with us your work your results and your wisdom about the use of computational tools in healthcare now let's go to see what Professor marier gasmi has been up to I want to start by asking you about your work in rule violations and so you have shown that AI models fail to reproduce human judgment in identifying rule violations what is this all about well the issue is that when we label machine learning data sets in standard settings we're always asking people if a feature exists is a dog big or is a meal sugary but that's not actually why we train machine Learning Systems we usually train them to predict a violation of some rule is this meal too sugary for the school rules or is this dog too large for the building's Rule and so in that setting it's really important that we tell people you're going to be judging whether a dog is large for a rule violation because what we found is that people's Judgment of whether a dog is large actually flips when you tell them that it's being used to create a rule in a machine Learning System how does this translate to medicine if we have labels that don't reflect the ultimate judgment that's being made it's possible that some people will be denied services that a human judgment would have given them in the first place so what can we do about this we can look at giving more uh specific prompts to people who are judging whether a feature is present in an image or in text or in tabular data and let them know that this judgment will be used in a specific human rule setting and when they have that context we think that they'll make better judgments so we can have machine learning systems that mimic human judgments what we found is that when you try to improve the models that are using this data you have to make assumptions about how harsh the system should be and so when we looked at algorithmic improvements if you can make good gu guesses about where the threshold might be for the data so for example I know you've labeled this feature about a large dog factually but if I asked you subjectively whether it's a large dog in violation of a rule if you can guess what that would do to somebody's judgments then you can try to fix some of this on the algorithm side but couldn't you tweak it too much the other way so couldn't the system be abused then that's why we think there needs to be a holistic difference in How We Gather these labels in machine learning data sets you can't fix it all with the algorithm this is something we really need to change from the ground up and how would this look for medicine well in medicine what we've been looking at recently is how we can give doctors descriptive or prescriptive recommendations for different kinds of actions when you think about the ways that you could give advice to somebody usually it's based on an if then rule so if the meal is sugary then it's a rule violation or if the patient has a risk of mental illness then refer them to a psychiatrist and so we want to to look at whether these descriptions of whether a meal is sugary or a dog is large were the right thing to give a patient or their Advocate versus the Judgment this prescriptive action of saying that a dog is in violation or a meal is in violation or a patient should see a psychiatrist and so what we did is we looked at different settings such as Mental Health crisis lines where you could either say that a patient has a risk of violence or that you should call the police to help that patient and we wanted to look at what happened when we gave very biased advice from an AI system to doctors either descriptively saying there's always a risk of violence for minorities or prescriptively saying you should call the police on minorities consistently what we found is that if you give really biased AI advice to doctors in a descriptive way you just say that there's a high risk of violence for all the minority patients doctors don't listen to it they retain their original Fair judgment but when you tell them to call the police disproportionately then they do listen to it so this prescriptive recommendation that tells what to do somehow short circuits critical thinking skills and so in health specifically we need to think carefully about not only what predictions were making but also exactly how we deliver those predictions and what we find is with these descriptive systems you're giving them the rules and they're able to follow through on these judgments and have really good critical thinking skills but when we give them the end result just shortcut to the answer they don't actually go through the process and so they miss mistakes that the model might make and so it's crucial that when you have a system helping people it helps them through the thinking process not through this sort of efficient shortcut that they could have we find that when people are very good at the task then you can train a machine learning system based on past labels in this data and then that model will be so good that you could probably just use a prescriptive system and say go ahead and follow the model we think that most of the data we have while it's good it has errors it has flaws there's variation and so we want people to continue to understand how to make these decisions with advice but not always follow it now in addition to descriptive and prescriptive Solutions you're also considering normative Solutions normative Solutions are things that we believe that people should do in most cases that others have judged as being the right decisions to make this doesn't always happen in a health care setting but there are some really wellestablished care pathways you can look at in machine learning for health where we know the right things to do we want to reduce the variation in how different patients are treated and have a more normative solution because we know the right answer so how do you go from descriptive and prescriptive to normative for many chronic conditions there's huge variation one person's diabetes is not another person's diabetes one person's pregnancy is not a person's pregnancy and so because there's a length of time over which you have different complications there's going to be so much room for variation but a septic patient is often a septic patient they're so acutely ill that the way that you can treat them the number of levers you have to pull they're very small and the variation is also very small and so we'd like to have more normative actions this work is so important and it raises such profound ethics issues with deploying AI in medicine when you give advice to somebody who's already Bound by an Ethics code as all clinical staff and clinicians are are what is your additional responsibility as a technical person who's contributing to their interaction with a human system there's no easy answer but what we're trying to do right now is understand where all the biases in the process are from collecting the data to defining the labels to developing the algorithm to deploying specific advice and we think by codifying those and auditing them and letting the people actually using the systems know what's happening we can have a better end to-end system where doctors make those judgments what else are you excited about what's the future in your lab having systems that can flag incorrect data from the get-go in large systems in an unsupervised way this is definitely a problem in health it's also a problem in places like computer vision where we know that images of abuse and inappropriate treatment make it into large image data sets and then if you're not looking through these huge sets of images they can make it into Downstream models so if we have systems that can successfully audit data sets then we can have a better set of data to learn from from the beginning and then all the way on the other side I'm really excited about understanding how we can take deployments and deliver advice in a way that doesn't perpetuate biases but helps clinicians create more Equitable Healthcare Systems within their practice we don't want to create supervised systems for identifying bad quality or inappropriate data Within These large data corpora because that just doesn't scale instead what we want to do is look at the different views of data that exist within large scale embedding spaces and then use these different views as self-supervised learning where one view of an image is contrasted with another view and we think that differences between the ways that different models view data will highlight when data is inappropriate or incorrectly labeled marsia that sounds so interesting and important and I can't wait to hear about the new results thank you for having mehello everyone and welcome back to Cale office hours I am Daniela ruse the director of Cale where we are inventing the future of computing and making the world better through Computing today our Focus will be on health and how Computing and AI can help with diagnosing monitoring and treating disease Dina you have made so many impactful contributions to communication and to understanding Wireless signals how did you think to move from studying networking to applying Wireless signals to the healthcare sector maybe about eight years ago we started looking into how we can use um radio signals to monitor physiological signals so that was the early set of results basically just understanding that radio signal bounce around and they bounce out the human body because we are made of water so signal comes and reflects and then if you analyze those Reflections particularly if you analyze them with the power of neural network in machine learning you can start learning things that you couldn't imagine that these radio signals around us have or carry about us such as of course our respiration the pulsing of our blood even when you go to sleep we can tell whether you are in the dreaming stage for example ra versus you are in light sleep or deep sleep and then that took us to the Next Generation not just monitoring the physiological signals of people using radio signals without touching them but also collecting that data and starting to develop neural network model for uh diagnosing and tracking like what's called biomarkers of diseases for Parkinson's for other newgeneration diseases and the final the most recent thing is Gen AI for uh Medical Data with your algorithms and your insights in Wireless networking you can see what we cannot see with a naked eye how does this work it takes many years to diagnose Parkinson like somebody may had Parkinson 10 years ago and they haven't been diagnosed because one of the issue is that you wait for the motor symptoms you know in Parkinson the Tremor the stiffness of the motion and these motor symptoms appear late in the Disease by the time the motor symptoms appear the uh the dopamine neurons which are the neurons involved in in Parkinson about 40 to 80% are already damaged when we learned this we were just like oh but can we diagnose it from other signals now that we have these devices in people's home and we just take the wireless signal maybe from warless we know that we can get people breathing let me just try to detect whether they have parking from their breathing because there are some hypotheses that breathing like is impacted by the uh basically that Parkinson comes from the GS along the vagus nerve and then it has the breathing area in the brain so breathing changes earlier but actually it turned out that you can diagnose or machine learning algorithm can diagnose Parkinson with high accuracy from someone's breathing what is it that the intersection of understanding Wireless signals and understanding machine learning what is it at this intersection that helps you solve this problem we are not trained to look at radio signals but machine actually when you train them with the right output they they first they can sense the signal and they can start seeing patterns that the human brain is not trained to see from sensory signals that we don't have like electromagnetic waves where does this data come from the data that you use to train your models we designed a device that looks very much like your Wi-Fi box at home so sit in the background of the home and it's like analyzing these radial reflection transmit very low power Wireless signal analyzing the reflection and analyze them neural network and from that it can be like you don't have to wear devices on yourself people can live their lives and you can just get all of that insight to clinical data but it also convey the information to the medical doctor to the practitioner who's the expert and can help them can this model monitor for multiple diseases or is it finally tuned for Parkinson's U monitors multiple diseases particularly interested also I find the the biggest opportunity is in neurology and Immunology so it's very hard to to study neurological diseases even things like depression we ask people how do you feel but give you like a drug or treatment he said there's no actual concrete measurements to do that that's where we are seeing like our results from being able to concretely and objectively measure the physiology of the person and how it changes we can see a major uh opportunity to have a um what's called biomarket in this context digital biomarker for these complex diseases can you tell us a little bit more about Immunology what are you looking at in the space one essential thing of the immune system is that you are talking about a dynamic system a that is changing like it's just your immune response it's a response flare up there are flares you change you are in remission if you have like an autoimmune disease you are in remission and then you have a flare but in the whole medical system we don't have a the like all of the tests that we do are snapshots like when we talk in computer science we talk about snapshots ver versus like continuous tracking we don't have the continuous tracking and clinical data for that continuously can track a dynamic system like the immune system there is a hugee opportunity here to be able to move from snapshots particularly when you are talking about inflammatory response immune response to continuous data and where we can see where we are using our devices and Ai and to track these changes so are you finding that the Wi-Fi signals reflect differently if there is an immune system flare up when they bounce off my body your body the environment they change based on everything your respiration the pulsing of the blood the twitching of your eyes like different muscles all of the behavior that we have and things that we don't even think as of as Behavior so the physics of how the reflections happen don't change but when you extract this information from those Reflections now those change your respiratory system response to inflammation is going to be different your sleep and sleep patterns will change if somebody has a particular disease or taking particular drug so those then you can think about Ai and two levels one level is like taking radio signals and extracting from the physiological signals and then another level of neural networks that takes those physiological signal continuously extract it from the radio signal and start talking does this person have Alzheimer does this person have Parkinson how did their Parkinson change are they taking anti-depressant are they responding to is anti-depressant how do you use generative AI to get to better results being able to to take some simple signals or let's say like something from your Fitbit or some like your your pulse or something and transform that using generative AI to clinical test that was the use of generative AI that we are looking at and we actually do have initial data that shows that this is doable and the generated data is very much similar to the actual real data from the medical test for that person your generative AI system for medical data is it um what is the raw data used to train it we move from one medical modality that is cheap to a medical modality that is expensive a cheap medical modality it's it's kind of like a wellness modality like if I take the the PPG from uh like the puling from like AR device or uh the breathing from radio signal these are cheap available and easily and from that I generate something that is expensive like an EEG signal for example for the brain Dina it sounds like you have invented a new technology that is noninvasive and that gives doctors a kind of a superpower because it allows them to see things that are impossible to see with today's Technologies medicine I think there's a huge potential for competition in general I think even without a there are plenty of opportunity for competition but with machine learning also the the opportunities go even higher thank you for putting so much thought and effort in making healthc care better let's go meet MIT professor John gag John has been working with artificial intelligence and machine learning to develop the field of healthcare information and he has been looking at how machine learning can enhance the diagnosis of cardiac disease my first work in healthc care and AI was in in cardiovascular medicine and I had the good fortune that Colin staltz who's a member of cesil and a faculty member in eecs is also a practicing cardiologist Colin and I have been looking a lot at can we use non-invasive signals to get good approximations of data that is usually gathered invasively so the best example is probably an ECG measures the electrical activity of the heart every patient in the hospital is typically has an ECG connected to them what's been interesting is the things you might not expect to be able to get out of that signal but we think you can with the right kind of Technology machine learning machine learning to learn it but then of course when you're running it in practice you have to have forward passes that run fast the role of machine learning is to train a model that will give the right predictions is that right we use machine learning in healthc care in a lot of different ways sometimes it's used to train models that we'll actually deploy but other times we use it to do what I'll call science and and learn about things and once we've learned about them then the model maybe you don't need it anymore you you you've learned something you've done some experiments it's helped you learn things if we think of experience learning as you generating hypotheses that can then be tested in clinical settings or in the lab then you don't have to trust the model because all the model is doing is giving you things that you're going to test once the model generate finds hypotheses you can then maybe look for uh mechanistic explanations or look for rules maybe it's just discovering a rule and you can explain the rule to somebody and oh yeah that's a sensible Rule and so I think a lot of the use of machine learning today in healthcare is in the science part of it where we're discovering things that can then be used in the clinic we started with communicable diseases and we did that in conjunction with MGH with the infection control Department there so the infection control Department's responsibility is to reduce the incidence of healthc care Associated infections in the hospital we were approached to look at the question of could we use machine learning to find ways to reduce the prevalence of Health Care infections in in the hospital and the one we focus on to start with was uh CI the most common one it's an intestinal disease bacterial disease our models predicted which patients were most likely to contract the disease if you know that someone is likely to contract the disease you'll be more aggressive about knowing when to test them and that will let you discover they have it sooner and our treatment sooner and so what we did is we built models that were quite effective at determining which members of the population were most vulnerable so if we look at what makes people susceptible uh some of what we found was already known when the uh ant some antimicrobials uh increase the likelihood of Contracting sedi and some more than others we found zip code was important now he said how on Earth can where a patient lives matter some of what they think of Hospital acquired was actually the bacterium was already in the patient system and then something that happened in the hospital allowed it to flourish and get worse and so the bacterium came from the community and it was say the treatment in the hospital that made it flourish suddenly it made sense and we looked at these zip codes and for example there were often zip codes that had a lot of nursing homes in them so John for this uh work in communicable diseases can you tell us a bit about the techniques that you have used about the computational aspects of the solutions one of the things that's interesting about a communicable disease is is the fact that you need exposure and that means you have to look at Network effects and a lot of machine learning algorithms don't really deal with Networks and so a lot of what we did was was looking at at connectivity networks to try and and and build models based upon that and you use machine learning for that yes how did you train your models we were able to get access to all of the MGH patient records you can look at the patient record and see which nurses are visiting which patients which doctors are visiting which patients where they are we actually looked at the architecture of the hospital but the most important network was the caregiver Network what are you working on right now the biggest investment in my lab right now is in medical imaging and machine learning has is in the process of totally revolutionizing Medical Imaging so if you think of the amount of data produced by an MRI it's three-dimensional high resolution typically and it's more data than a human can possibly look at there are a small number of tasks that you have to do for any kind of image one of them is segmentation so that involves you take an image and you find Regions that are anatomically significant and separate them from the rest of the image so segmentation is very important but not just for things like tumors for a lot of say neurological diseases you want to look at the shape of say the hippocampus you know what is this exact shape is the shape changing and to do that you have to find it and segment it very precisely we have built something which we call univers EG Universal segmentor it's important because the other Universal segmentors that are out there and there quite a few work only for natural images really they don't work very well for medical images and you need something that's very Broad because there are an enormous number of different modalities ultrasound to x-ray to to Mr to Optical to ooc all sorts of things he collected an enormous amount of data from many different Medical Imaging modalities many different tasks we then invented ways to generate synthetic data if you look at our synthetic data it doesn't look like any medical image anyone has ever seen or any task anyone has ever seen but it makes it prevents the model from overfitting to the actual tasks you show it during the training other thing we've done with Universe egg is is it's what's called context based learning at inference time when someone is trying to say segment something that they don't know how to segment what we do is you give it a small number of examples and that's used as a prompt to say oh that's the kind of that's what they're looking for and then based upon that small context and all of the other examples that's been trained on it then can generate quite good segmentations and we think eventually will be as good right now if you train a a one of model for one modality and one Anatomy it's slightly better than the universal segmentor as you think about computational tools and their potential in healthare where do you see us going in the next 3 to 5 years the large population studies will be the first thing impacted I think by a lot of the machine learning technology and I think it'll have enormous impact on things like tele medicine instead of having a person move to the clinic we acquire data and send the data to a computer and the data the computer looks at it and I think we can make very quick inroads in using machine learning to build models that will improve the health care system if Health Care gets less expensive we'll get better outcomes an awful lot of people don't get the treatment they need today because they can't afford it if we can reduce the cost it isn't just saving money it will give us better outcomes John thank you so much for sharing with us your work your results and your wisdom about the use of computational tools in healthcare now let's go to see what Professor marier gasmi has been up to I want to start by asking you about your work in rule violations and so you have shown that AI models fail to reproduce human judgment in identifying rule violations what is this all about well the issue is that when we label machine learning data sets in standard settings we're always asking people if a feature exists is a dog big or is a meal sugary but that's not actually why we train machine Learning Systems we usually train them to predict a violation of some rule is this meal too sugary for the school rules or is this dog too large for the building's Rule and so in that setting it's really important that we tell people you're going to be judging whether a dog is large for a rule violation because what we found is that people's Judgment of whether a dog is large actually flips when you tell them that it's being used to create a rule in a machine Learning System how does this translate to medicine if we have labels that don't reflect the ultimate judgment that's being made it's possible that some people will be denied services that a human judgment would have given them in the first place so what can we do about this we can look at giving more uh specific prompts to people who are judging whether a feature is present in an image or in text or in tabular data and let them know that this judgment will be used in a specific human rule setting and when they have that context we think that they'll make better judgments so we can have machine learning systems that mimic human judgments what we found is that when you try to improve the models that are using this data you have to make assumptions about how harsh the system should be and so when we looked at algorithmic improvements if you can make good gu guesses about where the threshold might be for the data so for example I know you've labeled this feature about a large dog factually but if I asked you subjectively whether it's a large dog in violation of a rule if you can guess what that would do to somebody's judgments then you can try to fix some of this on the algorithm side but couldn't you tweak it too much the other way so couldn't the system be abused then that's why we think there needs to be a holistic difference in How We Gather these labels in machine learning data sets you can't fix it all with the algorithm this is something we really need to change from the ground up and how would this look for medicine well in medicine what we've been looking at recently is how we can give doctors descriptive or prescriptive recommendations for different kinds of actions when you think about the ways that you could give advice to somebody usually it's based on an if then rule so if the meal is sugary then it's a rule violation or if the patient has a risk of mental illness then refer them to a psychiatrist and so we want to to look at whether these descriptions of whether a meal is sugary or a dog is large were the right thing to give a patient or their Advocate versus the Judgment this prescriptive action of saying that a dog is in violation or a meal is in violation or a patient should see a psychiatrist and so what we did is we looked at different settings such as Mental Health crisis lines where you could either say that a patient has a risk of violence or that you should call the police to help that patient and we wanted to look at what happened when we gave very biased advice from an AI system to doctors either descriptively saying there's always a risk of violence for minorities or prescriptively saying you should call the police on minorities consistently what we found is that if you give really biased AI advice to doctors in a descriptive way you just say that there's a high risk of violence for all the minority patients doctors don't listen to it they retain their original Fair judgment but when you tell them to call the police disproportionately then they do listen to it so this prescriptive recommendation that tells what to do somehow short circuits critical thinking skills and so in health specifically we need to think carefully about not only what predictions were making but also exactly how we deliver those predictions and what we find is with these descriptive systems you're giving them the rules and they're able to follow through on these judgments and have really good critical thinking skills but when we give them the end result just shortcut to the answer they don't actually go through the process and so they miss mistakes that the model might make and so it's crucial that when you have a system helping people it helps them through the thinking process not through this sort of efficient shortcut that they could have we find that when people are very good at the task then you can train a machine learning system based on past labels in this data and then that model will be so good that you could probably just use a prescriptive system and say go ahead and follow the model we think that most of the data we have while it's good it has errors it has flaws there's variation and so we want people to continue to understand how to make these decisions with advice but not always follow it now in addition to descriptive and prescriptive Solutions you're also considering normative Solutions normative Solutions are things that we believe that people should do in most cases that others have judged as being the right decisions to make this doesn't always happen in a health care setting but there are some really wellestablished care pathways you can look at in machine learning for health where we know the right things to do we want to reduce the variation in how different patients are treated and have a more normative solution because we know the right answer so how do you go from descriptive and prescriptive to normative for many chronic conditions there's huge variation one person's diabetes is not another person's diabetes one person's pregnancy is not a person's pregnancy and so because there's a length of time over which you have different complications there's going to be so much room for variation but a septic patient is often a septic patient they're so acutely ill that the way that you can treat them the number of levers you have to pull they're very small and the variation is also very small and so we'd like to have more normative actions this work is so important and it raises such profound ethics issues with deploying AI in medicine when you give advice to somebody who's already Bound by an Ethics code as all clinical staff and clinicians are are what is your additional responsibility as a technical person who's contributing to their interaction with a human system there's no easy answer but what we're trying to do right now is understand where all the biases in the process are from collecting the data to defining the labels to developing the algorithm to deploying specific advice and we think by codifying those and auditing them and letting the people actually using the systems know what's happening we can have a better end to-end system where doctors make those judgments what else are you excited about what's the future in your lab having systems that can flag incorrect data from the get-go in large systems in an unsupervised way this is definitely a problem in health it's also a problem in places like computer vision where we know that images of abuse and inappropriate treatment make it into large image data sets and then if you're not looking through these huge sets of images they can make it into Downstream models so if we have systems that can successfully audit data sets then we can have a better set of data to learn from from the beginning and then all the way on the other side I'm really excited about understanding how we can take deployments and deliver advice in a way that doesn't perpetuate biases but helps clinicians create more Equitable Healthcare Systems within their practice we don't want to create supervised systems for identifying bad quality or inappropriate data Within These large data corpora because that just doesn't scale instead what we want to do is look at the different views of data that exist within large scale embedding spaces and then use these different views as self-supervised learning where one view of an image is contrasted with another view and we think that differences between the ways that different models view data will highlight when data is inappropriate or incorrectly labeled marsia that sounds so interesting and important and I can't wait to hear about the new results thank you for having me\n"