Medicine today is, by and large, not personalized. If you and I go to the doctor with similar symptoms will be treated similarly, even though we have different gender, different backgrounds, different ages and so forth. But the dream that I have always had, and I think is becoming closer and closer with AI is that understanding the specifics of an individual or individual patient, including his the history of electronic records, the genomic sequences, the the different hereditary histories, should affect to perhaps even dictate the treatment. today We're talking to Dimitris Bertsimas. Dimitris is a faculty member at MIT And he's a professor of operations research.

He's also the Vice Provost of Open Learning. And we're going to get into topics that, involve his research career and getting into healthcare. So thank you for joining us today. My pleasure. All right. Let's start with your operations research. I think, you know, for most people, they may not know what that entails. Can you just give a basic definition of what operations research means? Yeah, it is indeed. The operations research is a well-kept secret. Even though, in my opinion, is a fantastic area. So what is operations research is there has a scientific method for for designing and improving systems broadly construed, and optimizing them.

Examples include transportation, health, finance, marketing, engineering, and so forth. I don't know many fields that are capable of, addressing issues, including your own field biology, that are capable of addressing that level of breadth. Yeah. And one example specifically that you did early on in your career is the Panama Canal. We won't get into the details of that type of work, but you can maybe explain what does that mean, operations research on the Panama Canal? So the Panama Canal was built in 1915 by the US government, and it was being operated by the US government, US government personnel until 2000.

So in 1998, I was contacted by the Department of Transportation in the United States to help them build a model to optimize the operations of the canal, because you have vessels coming one way from the Atlantic to the Pacific and from the Pacific to the Atlantic. And how do you optimize, given that the American, controllers would depart? So we develop the model, we optimize the model. And, it, actually launched, it was operating when I left it. You know, I haven't followed since, but, it has actually motivated me also to think extensions and other ideas. But, this was a case where optimization methods have been used to, to optimize the operations of the canal, increasing its throughput from, my recollection is from 42 vessels a day to 48.

And then with, taking into account uncertainty to 45 in the end. Yeah. So your research really has direct, real world applications, resulting from the action and the studies of you and your students. And I think, well, you know, your background actually, is that you're a PhD in mathematics, and you then went to operations and you found your way into health care in more recent years. And so you had some personal motivations to why you got into health care research. Can you share some stories of why you made that shift from just operations research to getting into health care and operations? I mean, the story starts when I was in my teens.

I was a good student, as many of the professors at MIT are, and my dear and my mother wanted me to become a physician and doctor. When I recognize what it is, it says it doesn't fit with my sort of more mathematical, orientation. But I have always been interested in health care. Now, so I started, actually a company in 2000 for, given my interest in, helping people, finding, what are the appropriate benefits, health care benefits for them? But it really intensified when, in 2007, my father was diagnosed with, gastric cancer. Unfortunately, a bit late, he was stage four. So he lived two years, and, unfortunately, he passed away in 2009.

But, it made me realize that, there are many opportunities for somebody with my background to help in health. The same year I have lost my mother to, my mother was diabetic, as most of my family from my mother's side. And, again, it made me realize that, personalizing, diabetes care might be relevant in this case. So I thought, you know, when you lose both your parents, you actually, in some for some people, that's the definition of adulthood. So this is the time I thought maybe I should pay significant more attention. I always paid attention to health. I wrote papers earlier, but but I have increased very substantially after that period and, and it's continuing to this day.

Yeah. And so let's, explain diabetes a little bit because, your family is greatly impacted by diabetes. I have a lot of cases of diabetes in my family. The United States is greatly impacted. But this is a worldwide problem where nearly 600 million people have diabetes or there are many undiagnosed cases as well. At a simple level, there's type 1 and type 2 diabetes. Type 2 diabetes is the more based on lifestyle and diet. Do you want to give an explanation of type 2 diabetes? Well, you know, this is a personal question. I'm diabetic. I inherited this from my family, from my mother and my family. So it's about 10% of the world population is diabetic, 9% is, type 2, and around 1% is type 1.

Type 2 diabetes is it is quite. It is the idea that, the body, has too much sugar. It doesn't metabolize, sugar appropriately. And, left unchecked, it could create significant difficulties for for life. You have significant complications if the if your, if it's, not appropriately controlled. So, it is correlated, of course, with, hereditary matters as it is in my case. Lifestyle, obesity does not help. In fact, it's detrimental. These are the two major reasons for type 2 diabetes. And, and it's typically being, addressed by lifestyle, dieting and, and there has been, insulin, the major discovery in the 20s, the 1920s, that continues today.

But recent recent drugs, GLP-1, it's called Ozempic and others are are capable of, helping manage the disease. You cannot cure it, but you can you can chronically manage it. Yeah. So it's really been in the media on the news a lot. And to give some of the biological background of diabetes, insulin that you mentioned, that's the peptide hormone that the body uses as a signaling molecule. That signal goes to cells. The cells receive the insulin and tells the body to uptake the glucose, the sugar in the blood. And then when we have, too much glucose, the cells start to become insensitive to the, the insulin.

And then that's when we have the issue of type 2 diabetes. That's really insulin resistance. and getting into those cells, not being able to receive that. And so in your early operations research, you did do some work with diabetes, and studying that. Can you explain what that looks like when someone at MIT, as a professor, we don't have a hospital here, how do you get involved with health care research and what does that look like in terms of the type of data you look at and then what the study is? So I work with hospitals. In this particular case, I worked with, the Boston Medical Center, which is the hospital for the underprivileged in Boston, where the prevalence of diabetes is typically higher. It correlates with, with, income, and social status.

So we had data about trajectory, electronic medical records from, from patients that were treated in the BMC in the Boston Medical Center for at least 12 years at the time we published the paper. And, so we know and because of the social aspect of the people involved, most of their care is there. So this is continuous medical records. That's different. For example, for a hospital like Mass General Hospital, where people go there for very serious, they don't go there for more routine situations. So we had electronic records, all the visits as well as their when I say electronic medical records, I mean their vitals, the glucose measurements and so forth every visit.

And then we tried to personalize the treatment. In other words, there are multiple ways of addressing diabetes. One way is, of course, lifestyle. What sort of, waiting but not actively taking drugs. Then there is, metformin, which is, generic drug. This was introduced in the United States around the 1970s that many people take, Sulfonylureas, oral medications, and more recently and then insulin. So the idea would be what is the appropriate given the history of a patient, what is the appropriate combination of treatments to maximize the effect, like minimize the glucose on the on the blood? And we've published using, machine learning AI methods and optimization.

We published in 2017 I don't recall 100% the year, but I think it's 2017 with some of my students. The first paper of using AI in Diabetes Care, which is a major journal of diabetes. So this was a case where we actually utilized real data, we utilized algorithms, and we published a paper that, that was the first of its kind to to show how personalized management. It was really motivated by my family. You know, if you look at my mother, my mother tried to treat it with oral medication. My aunt, the older sister. Yeah. Treated with, with insulin early on, actually, and my grandfather early earlier years did not treat it.

So you see very different approaches. And at that time there was no treatment. There wasn't so much treatment. Yes. So this was the first in diabetes that I know. But since then I have, I'm involved in, using genomic research to understand, using, genomic sequencing. Because diabetes, one of the ways that we get into diabetes is pancreatic decline. What is called beta cells. And, we are currently utilizing genomics to try to understand the, the decline. So as an early measurement of before the regular measurements of the tendency of the people to develop diabetes and therefore act when it's time. Yeah. So you can look at genomics to help with understanding the beta cells.

Beta cells are the cells that produce that insulin we've been talking about. And so if they are not producing insulin at the appropriate level, that's what leads to diabetes as well, especially type 2 diabetes. Yeah. There's type 1. And then there's type 2 is more of the resistance. So you mentioned machine learning. And the conversation today, around the globe. Really AI is on the news in the media, in academic research and education. And this has, started really because generative AI has really, led to AI being accessible to almost anyone. But in terms of your work, how would you you define artificial intelligence?

And is this something new, or is this a continuation of work over years? Yeah. The way I defined, I'll take a step before I define AI. So, in the 2010 or so, the, the area of utilizing data, to build models for predictions and prescriptions, decision-making. I call this analytics. The data science is from data to models to decision to predictions. And the decisions is the prescriptive part. This is mostly operational research. The unit I call it analytics. The definition of data was structured data with the emergence of the ability of analyzing non-structured images and language. We are now entering the world of artificial intelligence, where the use of unstructured data for building models, typically what's called deep learning models for prediction and prescriptions.

This area I call AI, Notice that this is not necessarily the same as the public perceives it, because the public perceives primarily the language, processing. But really it's the totality of the of all of this unstructured data that's the key element, structured and unstructured, that this is the ability that came recently for languages in the 2022 range. And images were a little bit earlier 2018 or so. So it's in the last decade, we have seen very significant advances in this field to the point that everyday life is affected, to the point that individual, you know, members, people that do not have, not have formal education, utilize these technologies.

Yeah, I think everyone has probably some experience with AI at this point and wouldn't have expected that five years ago. And you mentioned personalized medicine, personalized healthcare. What does it mean to you in terms of your research work or just in general? In general, in life. 265 So medicine today is, by and large, not personalized. If you and I go to the doctor with similar symptoms, we'll be treated similarly, even though we have different genders, different backgrounds, different ages and so forth. But the dream that I have always had, and I think is becoming closer and closer with AI is that understanding the specifics of an individual or individual patient, including the history of electronic records, the genomic sequences, the the different hereditary histories, should affect to perhaps even dictate the treatment.

So personalized medicine or precision medicine, as people say, is the idea of utilizing treatments specific to a person. And and it's coming closer and closer. I gave you the example for diabetes. Let me give you an example of oncology. So my father was, had a gastric adenocarcinoma for cancer. And, he developed (it) in 2007. He was given chemotherapy, which worked for a little bit of time, maybe two years. But then it did not. And he passed away in 2009. At the same time, and, actually a friend, a lady, had similar treatment, I don't know, adenocarcinoma, stage four. She's still alive today. She was given similar treatments, actually.

It worked for her, but it didn't work for my father. So this is a case where understanding, and we are. We are not there. I don't claim that we are, but as a direction for, scientific discoveries and, our role as, as researchers, I think there is room to utilize specific data for the patient. Genomics is one, but not only electronic records, images and so forth. And in effect, treatment that is distinct from person to person. That's the promise of precision medicine. Yeah, I think everyone probably has their personal experiences of someone they know that a medicine or a treatment worked for somebody, but somebody else with the same disease it did not work.

So I think there's a lot of examples And cases where people can understand this concept and why we need to look deeper into how we treat people. For your own case, you mentioned you're a diabetic. How does data collection play a role in your own management? In a significant way, actually, I'm using, a continuous monitoring device, that needs to be changed every ten days or so. So I have the entire history of my glucose. I don't have to pinch myself, as has been the case in the past. So I see actually when I eat the food, a particular type of food, the glucose increases, the speed at which increases. The speed at which goes down and so forth.

So people talk about glycemic index, but glycemic index is personalized. You know, if I eat rice, another person who is diabetic or not, eats rice. The curves are different. So it helps me intuitively, regulate the food, the speed at which things happen. It makes you more conscious, of course. Which is already positive and to, to manage the disease. I envision that in the future when you all this like for type 1 diabetics, where you have monitoring device and a pump that, that regulates, the, the pump would have the insulin. The pump would have the insulin and so forth. That's not needed for a type 2 diabetics.

I don't have the pump, but the monitoring device in a way affects my thinking of what to eat, what not to eat, how to, over time and so forth. It's not, you know, humans are imperfect. You know, it's not that I, I perfectly control, but it definitely it affects my thinking. It influenced my life to some degree that, overall has been beneficial to me. Yeah. And so you, day to day manage your, own diet with that type of data and monitoring. I try. I do not always succeed. Well that is the typical person. Right. If we could all be perfect, then we would solve a lot of problems. That's right. So then we have, you have this data, you have access to genomic data if you would go do that type of testing. So would you feel comfortable sharing your data with researchers like your own team and being part of that database? I would.

And why should people feel comfortable with sharing personalized health care data? And how would that help? Science? Yeah. Look, let me take oncology as an example. Oncology is the second largest killer. Heart disease is the first. Since 1970 that, the then president of the United States, Nixon declared war on cancer. We have made little progress. We have made progress on blood cancers, but on solid cancers, we have perhaps elongated life by a month or two. We have not materially affected that. Part of this, in my opinion, is the lack of data. If you look at where the data is, if you take, let's say, our own Broad Institute, who has an excellent genomic data, not so great phenomenological data, electronic records.

If you go to Dana-Farber, for example, they have excellent electronic records, not so great genomic information. Drug companies, oncology drug companies have have, different types of data. So this data is not widely shared. That is and I think this is a great opportunity for the world. I recognize that there is the, the aspect of privacy that, which might be, of course, a concern, but I think the broader good, treating and creating personalized medicine for these major killers of the world is, makes me comfortable because of the bigger good to. And not only potentially, I have already released them. It's not like in theory. Yeah. So, anonymized though, people not need to know that it is Dimitris Bertsimas, but other patient 157 has this data.

I mean, that's not relevant Who the person is. I mean, but the ability to have this data is a great opportunity for humanity. And I have been advocating and to quite a bit, but also reasonably successful with collaborations with hospitals where we have electronic medical records, not always genomics. It's not so widely available for all the patients in major U.S. hospitals. That has led to what I told you about diabetes, what we do about many other diseases, using anonymized data. Yeah, maybe explain anonymized data for people that are non researchers. Yeah. So typically when you have an electronic medical record, It is known that it is my name.

Where I live, my age my social security number, my ID and so forth, that's in the record. But if we do want to do research on the topic, what you care only is about what blood pressure I have. What diabetes measurement. You don't care about the particular name, the address, the social security number. You need. 403 You only need to have an idea of a person so that if this, this ID also has, a CAT scan, then you can relate that this comes from the same, but who the person is, and so forth. That's not relevant. So anonymization is the process of taking out of the record the identifying characteristics of the person, like address like name, like Social Security number, which allows, the ability to not expose the details, the health details of a person, but still allows the ability of research to, to understand diseases.

Yeah. As a researcher, then you never know the person who that data connects to. That's right. I don't know. And you don't need to not know. And you can still, have an impact and collect that information as you, do your studies. And now that we're in this, day of generative AI, and you've had the ability to apply it to your own health care research, is there a recent story that you, feel has been really impactful that you can see the difference already in the health care industry that you can share? Yeah. So as you, as you perhaps know, Mary Ellen, the development and the improvement of generative AI is only in the last three, three and a half years.

It's not a very long history, but the progress has been extraordinarily strong. So in my group we utilize generative AI tools, together with the other records to do that differential diagnosis, both at the hospital level. So somebody goes to the ED, to the emergency department. And, the typical situation today is that the patient is being interviewed by a nurse as part of triage. Then waits more. At some point, they write a document. This goes to the internal ED, they start further exams and so forth. It takes quite a bit of time to from the time the person arrives until something happens. So the average time in the ED is of the order of 6 to 8 hours, depending on the hospital, before either discharge or entering the hospital.

So what we have tried and currently in pilots is to utilize the then notes from the from the triage together with electronic records that the person might have, including the vitals, make a differential diagnosis using generative AI as well as the other data. And we call it multimodal because you have tabular data as well as language notes. Then this, then we after the differential diagnosis, we also order test. These are advisory to physicians. So by the time all this record goes to the physicians in the ED, we already have a differential diagnosis. Differential diagnosis means that you have three possible diseases, five, with corresponding likelihoods.

And then you have a test to try to clarify to make it, far more precise. This is advisory. For example, do we need x ray? Do you need the ultrasound? Do you need the, what types of blood tests to clarify. So the physician might modify this, this list and then the, the orders are made. We have found that this can decrease the length of stay within the ED from 6 to 8 hours to half that. And in addition, it makes about 85%, you know, hospitals are being paid by, the tests they do by the diagnosis and so forth, and especially with the tests, 85% of the tests are accepted by insurance, 15% are not. We have found, because we do it using, systematically in a consistent way.

Our rate of, rejection by insurance is more like 5%. So we have, so better outcomes for the patients and also better income for the hospitals. Overall, that's one example from the hospital side. We also utilize it for, giving advice to humans. So let's say, so people use it already used Chat GPT or Claude. Yeah. So for a patient you're saying in this case. To basically give you advice, we have built a system based on GenAI that is capable of doing better differential diagnosis and ask questions that inform the information gathering better and so forth, We are currently pilots in both cases, both for patients as well as hospitals.

Yeah. So for the doctors that are receiving this, these AI guidelines, are they happy to receive these? Are they welcoming this input? And then this is maybe a guardrail to some of the fears to having AI make decisions that can be false. AI does not make decisions right. It gives advice. Clearly this is by example because I work with people that are excited about this. The AI revolution. So the people I talk to are of course, are positive. We are not always, sometimes they say this is not correct. We iterate. This is an effort to improve. But, regardless, I, I do not believe that AI alone is the answer. What I believe is that given the progress we have made, the combination of humans and I, it gives better outcomes than humans alone or I alone.

So that's what I'm advocating at the moment given the current state of the art. Of course, the state of the art has increased drastically. So. But that's my current belief. Yeah. And if you could save some patients time for waiting for in half from 8 hours to 4 hours, I mean, that's something that everyone would welcome. We all have an experience waiting in hospitals or doctor's offices. So this, hopefully, the pilots are already in place. So we'll have to see where this, how this turns out as time goes on. Right. And it's clear that there is sensitivity on the matter because, you know, everything, you know, humans are not perfect.

AI is imperfect as well. So, we would like to make certain that we don't make serious mistakes in this effort. So this is not something I take lightly. But I do believe, that that AI will make a material difference in health in the future. And I have definitely positioned my group, research group to, to impact positively. Is there a future that you see in this in the healthcare field, it could be related to the generative AI impact or just your operations research in general? What do you think is possible in five years? It's difficult to predict the future. As you know, it's the only certainty that some of it would be wrong.

But my view is this that, I expect in the in the next five years that, personalized medicine will be with us. So I imagine a world in which a patient visits a physician, it knows the system, knows the history of the patient, its electronic record, whatever is available, the AI listens to the discussion. Potentially the physician can get advice about further questions. And a personalization plan, a personalized medicine based on AI is proposed to the physician. Of course, the physician is the ultimate decision maker who has the final responsibility, but it opens the horizon, in ways that are not available today.

I mean, the current situation is that, you know, a physician, let's say PCP, primary care physician sees 40 patients a day for 15 minutes. I mean, there's not much, time to. You can't collect the data in that time. It really can can make a material difference. And I think better outcomes are here. I honestly believe that, that AI will, appropriately, combined will make, material difference in health. As I also believe in education, but that's another story about that. But, but in health, I, I definitely believe that. And we can check the recording in five years to see how accurate your prediction is. But, you know, as everything in life, if I'm directionally correct, I will be happy. Yes.

Yeah. You have to aim somewhere and set goals, right? That's right. So I think your career is, a primary example of lifelong learning. You trained as a mathematician, and you mentioned that you were inspired early on because your mother wanted you to be a doctor. But at some point, you still continued on with your education in mathematics. Then you went to operations, which is similar to mathematics, but maybe you had to learn there. And now you are also learning healthcare in those systems. So how do you approach continuing to grow this way as an adult? Well, I have you know, there are very I, I try to live my life somewhat on some principles.

One of the principles I have is to do what I love. And one of the things I love is to impact the world positively. Problems of the world don't have labels. They are not mathematical problems. They are not biology problems. They are problems. And the knowledge that we need to put to affect this is different. It's from varying fields, for many, many, many fields. Just operations research by its nature because the way it was structured, the way, it tries to address multiple types of problems in the world. So the motivation was there. So, so I found the opportunity to, to adapt. When I want to address a problem, I have to learn what is the problem for the problem, if you want to address problems, you better understand somewhat health care.

If you want to understand, personalization, you have to understand genomics. So among other things, I'm not saying. So I have found that continuous learning and adaptive, adaptiveness to to the question in front of you is, a key to success. It's also quite desirable for me because it increases my happiness because I have, found something I love to do, and I care about the topics I address. And it also gives me an opportunity to work with, young people who, you know, younger people generally are much more open to ideas, but don't they don't have the weight of experience. So the combination of all this, the field, OR (Operations Research) is a broader field.

The, the, my philosophy about having impact positively in the world, has led me, in being adaptive. I have found many times that if you don't change your perspective because we are not perfect, you know, half of the time of what I initially think is not correct. But if you adapt. Hopefully for the better, I have found the combination of this mix makes it also life very interesting. To this day, I'm not exactly a youngster, I feel I keep learning. And I consider this the the blessing of being a professor at an institution like this with so many talented people. And will that attitude apply to other careers and non academics as well?

That's for sure. And I have been a serial entrepreneur. I, you know, the major reason for being an entrepreneur is because I would like, if you leave your ideas for somebody else to try, good luck. Most likely nothing will happen. So whenever I believe I have something that can help the world, I try to apply it myself. Of course, I cannot apply it in research with students. It's not appropriate. They are here for, for learning and research. This is more, repetitive in that way. So I have found that the ideas that I start with a company sometimes don't work, and then I'll adjust. And they don't work again, I adjust again and so forth.

And similarly, I have followed a similar approach with, Open Learning. You know, not all of the ideas I have, are correct, but hopefully, particularly when you hear other ideas, you move. And, I have found this to be adaptiveness. This adaptivity is, a key aspect of for human success. I have found. And say you're a high school student, somebody working in industry or someone abroad who is just curious and maybe retired. How would they be able to use resources from MIT to address this lifelong learning, a goal? Yeah, I mean, one of the, advantages of, Open Learning, this is, of course, a biased perspective because I am responsible for it, is that it has the capability, you know, over the last 25 years from OpenCourseWare, MITx, and what we are currently doing for universal learning, we are developing materials that, aim to help people.

And, we have put it together in a, in a platform we call Learn.mit.edu. And that aims for helping people from a young age, high schools and so forth to university years to to professionals, to retirees. We have developed, and that's something the community has to be very proud about, significant learnings, courses, podcasts, you name it, and so forth, videos and so forth. And, so the MIT, if you look at the Central MIT website, this has lifelong learning, lifelong learning. If you click there, you go to Open Learning (Learn.mit.edu). So I feel that, it is, a significant contribution of the world that what MIT does.

Quite proud and happy to be responsible for this effort. And, my hope is that, under my watch, this would increase, and exponentiate, based on the fantastic basis, my colleagues, my predecessors have already built. Yeah. So, as Dimitri says, everyone can access MIT Learn. At learn.mit.edu to learn more, starting today, or, any time you have something you're interested in. Uou can check it out and try to explore more. And I can vouch that before, Dimitri was in charge of Open Learning and responsible for MIT Learn that he was also a participant. He participated in biology courses, taking the introduction to biology course, The Secret of Life.

And, as a learner of Biology as well. It is my favorite course, actually, more than mine. That's quite an endorsement. It is. It's a fantastic course. Inspiring course. With Eric Lander that I helped put together. So if you are curious in health care and biology topics, you could definitely start there as a point of, learning more. And I think for today, we can end there with this conversation. Thank you, Dimitris, for being part of this and for visiting today and sharing your input and insights into health care research. Thank you, Mary Ellen. like we're having a meeting and you just talk and answer questions.

But I'm the one guiding the conversation. Can I ask questions? You could, you could, and it's fine. It's just a conversation.