CBS Faculty Live: Leveraging AI to Improve Healthcare with Professor Carri Chan

Okay, welcome. Good afternoon. Thank you for joining us. My name is Elisa Douglas, director of Alumni Engagement and Programs at Columbia Business School. Welcome to our CBS Faculty Live Virtual Talk. This virtual series brings our CBS thought leaders to you no matter where you are located. And in just a few moments, we will hear from Professor Carrie Chan on leveraging AI to improve healthcare. This talk will inform on health, healthcare leaders can leverage predictive analysis to transition into a proactive future. And she will take the next hour to present and then we'll of course open it up for your questions. Please use the q and a function during that time to submit your questions. We'll do our best to take as many as we can. And now let me give you a little more background on Professor Chan. Carrie is the Cain Brothers and company professor of Healthcare Management Decision Risk and Operations Division at CBS. She is also the faculty director of the Healthcare and Pharmaceutical Management Program at CBS. Her research is in the area of healthcare operations management. Her primary focus is in data driven modeling of health care systems. Her research combines empirical and mathematical modeling to develop evidence-based approaches to improve patient flow. She has worked with clinicians and administrators in numerous hospital systems, including Northern California, Kaiser Permanente, New York Presbyterian, and Montefiore Medical Center. Now I would like to hand it over to Professor Care Chan. - Thank you Elisa, and thank you to all of you for, for being here. I see a number of familiar names in the list, so it's great to see all of you here. My hope today is to just shed a little bit of light on some of the research that I've been doing in this space, both the academic research and then the actual implementation of some of this AI in healthcare. And of course with all of the buzz that comes with AI today, I wanted to just give us a first, a little bit of a level setting and then I'll, I'll dig into some details. So just some quick hit facts I suppose since all of you are logging in here today. Some of these things you probably already know globally, healthcare expenditures comprise 10% of GDP in the us. This is much higher around 18%. It is a large part of every economy, and I've been saying for many years, and it's really accelerated recently that healthcare is on the cusp of a regime change. So the World Economic Forum actually estimates that hospitals generate about 50 petabytes of data per year. It is growing much more rapidly the amount of data within healthcare relative to other industries. But that said, there's still a lot of opportunity for impact and growth. About 6% of health systems have some form of AI strategy. And of course this data is changing daily, but remarkably, despite the fact that so much data is being generated, only a really small portion of it's actually being used. So estimates that have about 97% of health data is going unused. And this is in large part because much of this data is unstructured data. It's data in clinical notes or there's challenges with interoperability and importing that data into a usable format. But with a lot of the changes that have been happening recently, this is gonna have a big impact in shifting what is possible. In fact, the amount of savings that have been estimated with widespread AI adoption in healthcare is about $365 billion. And so the hope is with this promise of the technology, maybe we can actually start to cut in to this massive amount of cost, improve access, improve quality. And so a lot of my work focuses on that. So where are some of the biggest challenges within the US healthcare system? You've probably heard some of these statistics estimates are anywhere from about 700 to $900 billion in waste

and it's cut into different parts of the eco ecosystem. We have estimates from fraud and abuse, pricing issues, administrative complexity, but there's also a lot related to the actual delivery of care. Are patients getting the right type of care at the right time? Are we making sure that there isn't redundant care, unnecessary redundant care? Can we better align? And I think this is really where the opportunity for AI comes into place. So at a high level, the way that I think about AI in healthcare is there is a lot of excitement and discussion around the use of predictive analytics. And this has been happening for many years. This is not something that just came out in the last, you know, three to five years when chat GPT was introduced. This was happening decades ago, but there was always this intent to predict things that were happening. But less of a focus on what do you do once you have those predictions? How do you actually make decisions based on a prediction and what are the potential benefits of making those types of decisions? So that's really what I wanna highlight some examples of where we might do this. If we think about some of the biggest challenges within our healthcare system, the traditional model has been a very reactive model. Patients get care and seek care when they are very sick, when they need to go to the hospital, see the doctor. This is reactive approach ends up often resulting in higher costs and worse outcomes because you wait until the health state is fairly dire and sick before making corrective approaches. A lot of the dialogue around shifting the cost curve in healthcare has been around trying to be more proactive. And so this is really where the potential and the promise of predictive analytics can come into play. If we know what some of the future risks are going to be, can we intervene earlier using fewer resources, lower costs, and actually have better outcomes? And so this is kind of the excitement and promise that I wanna show you and is actually some instances happening and being seen as we speak. So what I wanna highlight, I'm gonna start with some examples of proactive care where we actually use some of these risks models that patient risk to, to intervene earlier. I'm going to also talk about some examples in workforce management. There was recently a nursing strike in New York City. And so this is about trying to use these types of predictive analytics so that we can better manage workforce so that our staff are not constantly being burnt out, that we can make sure that there's the right amount of STA staff when they, the patients are showing up. And then some time spent on some of the things that I'm going on doing research on right now and, and some concluding thoughts. So let me start with a, a story about early warning scores. And this is actually a project that I've been working on with the Northern California Kaiser Permanente for many, many years. And this is what actually got me to working in the healthcare space in the first place. So it takes place in intensive care units. These are units that have the highest costs for patients in large part because the patients are so unstable that they need essentially constant monitoring by physicians. So they can cost anywhere from 3000 to $10,000 per day. The amount of spending in the US is around a hundred billion dollars on ICU care. And so you can imagine for patients that need ICU care

timely access to care is also of the utmost importance. And in fact, delays to ICU care can have some pretty substantial negative impacts. So this came as an idea with Kaiser where they wanted to build a AI model to predict whether or not a patient was likely to deteriorate and need to go to the be transferred to the ICU. And so we wanted to work with them to understand how should they actually make decisions based on these early warning scores. So the first thing to do if we wanna consider deploying such a model is, you know, the premise around proactive care is that earlier interventions should result in better outcomes. So we need to verify that this is actually the case. If we can get care to patients who are likely to deteriorate before they actually deteriorate, do they have better outcomes? And how much better? Presuming the answer is yes, then the question is, when do we do these proactive interventions? If the first one doesn't hold, there's absolutely no reason to do any preventive care. So this dev model that was developed is called the advanced alert monitor and it takes information from the last 24 hours of a patient stay in the hospital or if they've just arrived at the hospital up to 24 hours before. So this can include vital signs, lab tests and trends. So existing models before they that were developed actually looked at vital signs and has measures. If anybody's vital signs were off normal, that would increase someone's risk. But some people have naturally low resting heart rates. And so what's more concerning is not that the heart rate is low, but that their heart rate is very erratic, switching from high to low rapidly over time. And so also looking at these trends and then would predict would a patient, what's the likelihood a patient's gonna deteriorate in the next 12 hours? And so every hour this window of information would shift and there would be a new prediction for whether or not a patient is gonna deteriorate. And so the question is how do you actually use this information provided by the A M score? So we did a bunch of analysis and I, at a high level, the empirical analysis that we did found that indeed, if you're able to intervene on a patient who is at high risk of deterioration before they have that adverse event, we would see a substantial reduction in mortality risk, so reduced by 2. 56% and their length of stay in the hospital would go down by about a day and a half. So the, what we found is indeed strong statistical evidence that proactive care is beneficial. Great. The challenge here though is we can't give ICU care to every single patient. There's a capacity limited resource. Some patients may not ever need to go to the ICU and if we admit everybody we're going to needlessly, you clog up a very limited resource. So what we did was we built a number of models to try to estimate what the actual impact was going to be and estimate what would be good policies. So what we were actually able to find is that a simple threshold policy that said if we admit patients above this risk threshold, we should, if we have excess space, we should admit them into the ICU. If we don't have space or their risk is below that threshold, we should not admit them. And we're actually able to show that these types of policies have really nice properties that we can show

that they actually save the most number of lives. And so these two things together were evidence that we took to our providers at Kaiser Permanente and it was helpful for them to actually roll out implementation across their 21 Northern California hospitals. We, they ended up doing that. But just as a, for those of you who are thinking about operationalizing these types of models in practice, there were a number of very real world challenges related to change management, frankly. So this actually was a 10 year long pursuit from ideation of the initial early warning system to when this implementation happened across all 21 hospitals. We did the initial evidence collecting, then they rolled out the early warning score at two hospitals, but they only had the score available, no alerts, no interventions recommended whatsoever. The idea being that they, it was important for providers to trust that the model was actually predicting deterioration after those two years. They actually then at the two hospitals implemented a alert based on the threshold. And then after two years of monitoring that and showing the impact, they eventually were able to do a rollout across 21 hospitals. And so what they found, we were able to find is that having this early warning score and the alerts resulted in a about seven hour reduction in ICU length of stay and a 31% reduction in 30 day mortality rate. So the idea here, again, the, and I Imran, you had a question, the reduction was very much in part because getting access, the idea being getting access to care before things are really bad, you can see there's a benefit in reduction of total time and mortality risk. Now some of the patients who were admitted may never have needed it in the first place because our, these prediction scores are all probabilistic. But overall it showed that there could be a pretty substantial impact. And so here's this idea where these prediction models, if we have the right capacity available to intervene early, it can have a really substantial impact on improving overall survivability. And you know, the added benefit, if you're able to reduce length of stay, that also means other patients who might not have been able to get access to an ICU, there's now kind of additional capacity available for them. So we've been kind of expanding and I've been working in many different areas related to these early warning scores. So another area that's certainly relevant in the news today, yesterday, in fact I think the New York Times had an article about how 50% of new colorectal cancer diagnoses are for those who are 50 and younger. And so nowadays, I think actually in 2023, it was the second leading cause of cancer deaths. It might be moving up the list currently, but many people don't get the gold standard for of screening. Where nowadays the recommendation is anybody 45 plus who has average risk should get a colorectal, a cancer screening through a colonoscopy. Colonoscopies are not the most pleasant of things to procedures to do really the prep, it's a 24 hour prep. But they are remarkable in being able to detect cancer, colorectal cancer early and the prognosis when diagnosis is early is, is much, much better than later. So you can see some of these statistics here today. So this is a project we did with Geisinger Health Systems in Pennsylvania where they had a nursing team

that would look at all patients who are overdue for their colonoscopy. And if an individual is flagged as being high risk, the nurse would call up the individual and say, Hey, I've noticed that you're overdue for your colonoscopy. You've been flagged as being at higher risk. Can I help you schedule a colonoscopy today? And so then they looked, we looked at what the impact of this would be. And you know, not surprisingly that individualized outreach increased colonoscopy uptake by about 7% here. But what we found the impact on actual mortality was astounding. So a 6. 2% reduction in two year mortality, and we actually saw fewer types of oncology visits within two years as well. And I think the idea here is that, you know, colon rectal cancer, certainly, you know, screening earlier and having colonoscopy can have benefits. Sometimes there are pre-cancerous polyps that if you can catch during the colonoscopy this could have that impact. But also it's getting people in the door. And so for those who are at high risk for colorectal cancer, they may also be high risk for other types of cancer. And so getting them to engage with the health system can have actually a really big impact in improving overall survivability and quality. So I, I'm really excited about these opportunities. Lots of health systems are utilizing and deploying many, many pilots around these types of predictive analytics. But there is a real question about how do you do the right outreach? What's the right type of intervention? And if you do it, what is the potential benefit going to be? Because none of these interventions are, are lists, right? This nursing outreach requires having a team of dedicated clinicians to reach out. It means having enough colonoscopy capacity to intake those who have additional likelihood of getting screening. I'm gonna switch gears a little bit here and shift towards using predictive analytics on the staffing side of things. And again, quite timely, especially given the, the nursing strikes that we just had. And I think the reality is our health systems, the financial situation is quite challenging, especially with the shift of the aging demographics of having more government medicare pay, but also from a policy standpoint where the spending of Medicare and Medicaid has, you know, been not growing as quickly and there's actual pressures to continue to, to slow its growth. And so that puts our health systems in a challenging financial situation. At the same time, there is a lot of sick people coming into the hospitals. There are certainly many situations where there's where staff are feeling overburdened and overwhelmed and so burnout is a real problem. And so you seem to have these two forces that are opposed to each other. And what can we do? And I think this is where there's real potential for predictive analytics to help us better match the, the supply and demand so that we can lower costs and improve access and quality simultaneously. I don't think that you have to sacrifice one for the other. So let's think about emergency department staffing. There's a number of challenges. There's a lot of uncertainty and demand by nature of the fact that these are emergencies. We don't know exactly when emergencies are going to arrive. That said, there are a lot of seasonal patterns. So within a day, the middle of the day tends to have more demand than the middle of the night. Flu season tends to have more demand then in the summer. And then a very real challenge as opposed to other types

of industries is these are highly skilled individuals where you can't have, if we compare to the gig economy, someone come work for one hour if you, or half an hour if you have an increase in demand, people are coming for shifts whether they're eight hours or 12 hour shifts. And that can introduce some, some challenges. So what we've done is we've actually done some work with a couple of health systems in the New York area to come up with some algorithms to help better manage staffing. So let me start with an example on surge staffing, which happens when demand exceeds supply. So there are more patients that arrive in a shift, then there are staff and this can be quite challenging to deal with. This certainly leads to burnout. There are quality concerns because patients aren't getting timely access. So we wanted to understand if we could use real time predictions to do a better job of this surge staffing. So how does this work? Typically hospitals will set weeks or months in advance their base staffing. So think of this as the total number of full-time employees you're gonna have per shift and they're being paid at their standard base rate, their regular salary rate. But once you get to the beginning of a shift and you see that there's more patients and their staff, a number of things can happen. You can ask existing staff to stay overtime, so stay longer than their initial shift. You can call in additional staff to do an extra shift or you can call in agency nurses. All of these help increase the amount of staff to what you need, but you're paying at a premium to get these people to come in at a last minute. And so we have this tension where the cost of base staffing is lower but you have a lot of uncertainty of what's happening weeks or months in advance at the surge staffing level, you know, if you need more staff but you have to pay a premium to be able to get them in. And so we wanted to see if we could use algorithms to help guide these types of decisions. And why is this so challenging? Well let me give you some context. In the case of the flu season, which we're still a bit in the middle of, we know every winter we are going to see an increase in hospitalizations due to the flu. But what we don't know is how bad each flu season is going to be. So this is data from the CDC of the number of hospitalizations in the US due to the flu each season. And so you can see there's some seasons where it's quite low and high. So 2017 to 18 was a very high year and then last year year. This year the data CDC has not, has not posted the number of hospitalizations. So this is the best that I can do. But there's a big difference between having 140,000 hospitalizations across the entire US versus close to a million, right? We don't know six months in advance if it's gonna be a 2011 like flu season or a 2024 like season. But once you're in it, once the flu patients are showing up at the hospital, you know if it's a bad flu season or not. And so that type of real-time information can be very beneficial to improving staffing decisions. So what we wanted to understand is how do we actually make these decisions at these two timescales? So at the first stage we're trying to understand what is the number of FTEs, what's our base staffing? And the challenge here is we don't know if it's gonna be a bad flu season with a lot of demand or a good flu season with less demand. Then we think about when we're in the flu season we need to optimize how much do we surge up to? And the challenge here is that even if we know it's a bad flu season, we don't know exactly when the patients are gonna show up on

the ho at the hospital. Are they gonna show up at Monday at 8:00 AM or are they gonna show up on Friday at 3:00 PM? And so there's still some uncertainty but not as much as you know, weeks or months in advance. And so we wanted to build an algorithm that led us leverage the fact that we would get more information and optimize the staffing. So what did we do is we looked at the information that's available at these two different timescales at the base stage, very limited information, day of week, time of day, is it a holiday? You know, fun fact the Monday after the Super Bowl there tends to be an increase in demand and emergency departments. Maybe it has something to do with how excited people are getting or disappointed people are getting with the game or the fact that maybe the diet that day is probably not as, it's quite different than on other days. Then at the search stage we actually use real time information. How many patients are currently current in the emergency department? How sick are the patients that are in the emergency department? Is it snowing out? Is it very dry? We also use Google search trends. So we find, you know, lots of people will go into Google and nowadays maybe chat GPT and say here are my symptoms, do I have the flu? And so that's actually quite predictive of how many patients are likely to show up in the next shift. We built a algorithm that tells hospitals how many people they should staff at the base stage and how many at the surge stage based on these predictions. And we're able to show that it performs actually quite well in a theoretical sense. So we took this information to our partners at Hackensack Meridian over in New Jersey and we ran a pilot study where every day, this is actually the a screenshot of the email that we sent to the nursing manager. And so we would make the prediction of the number of arrivals of patients over the course of the day and our recommendation of how many nurses they should staff. And so if there weren't enough nurses, they would have to call in additional staff. We rolled this out and what we found was by having our recommendation available, we estimate that they would be able to save $1. 4 million in staffing without having any impact on patient quality or access. So no change in the emergency department waiting time, total time spent in the emergency department, no change on the likelihood a patient would get so frustrated with waiting that they would leave without being seen. But just by leveraging the predictions of when the patients were gonna be showing up to the emergency department, we could be smarter about when do we surge patients, when do we not, excuse me, when do we surge staff? When do we not surge staff? How many staff should we staff at the base level weeks in advance? And so by just having a better match of supply and demand that's facilitated by these predictions, we can actually cut costs substantially. So what I wanna do maybe in the remaining time is talk about some ongoing work that is still very new and I'm, I'm very excited about. I'm working on it with my colleague Hannah Lee and a PhD student in the statistics department. And so the idea here is really trying to think about how to integrate AI into these service settings. And so here I'm gonna talk about it in the healthcare space, but you can also imagine other settings in education where you might have an early warning score for advisors to reach out to a student at risk of failing or in job search where you are targeting candidates who are most likely to want to join a an organization. And the traditional approach

especially from the computer scientists and the AI frankly community has been trying to make sure that all of these predictive models have the best possible predictive accuracy. If we can predict things the best, then it should be able to do the best. The problem with this is when you're operationalizing these types of models, you ignore the fact that there may be capacity constraints, which means that this actually you could either have too much intervention and not a enough capacity to management or you could actually have more interventions and you have idle capacity. So an approach that many in the operations field have taken is to try to make sure that we deploy these types of models to use up all of the capacity. But the challenge here is that it actually doesn't try. Think about why are we deploying these models and what is our objective in doing so. So let me give you an example of how this might work. Suppose we had a prediction, and I'm gonna use this in the context of an infection, let's say of sepsis. So this is a project we're doing with the Columbia Medical Center where we have predictions of a patient likely having sepsis and so there's an alert that nudges providers to check on patients who are at high risk of having sepsis. And so there is then some people who will actually get requests to have treatment. So with this nudge it might be the case that not everybody is able to be seen because there's too many people to be seen. And so you have to select amongst the ones that have the alert to actually go intervene on, even those who are not alerted may be raised to be checked on because a provider saw them and felt like there might have been some risk. And so they can, what we refer to as like self request, once we have all these requests, there's a limited amount of capacity for those who can receive the service. And so the question is, is we wanna maximize actually the number of people who actually intervene on who are at high risk. And so that's what we're trying to understand, who should be targeted and how do you choose what algorithm? And I, the second one is one that many health systems have asked, they say we're constantly being pitched by different companies that say they can intervene or they have these different models that are really high quality, but how do we know which one to use? So that's what we're trying to understand and the biggest challenge is these capacity constraints. You can't treat everybody. But also we don't quite know when we intervene who is going to respond. Some will and some will not. So I just wanna give you a little intuition about why this could be challenging. So let's just do a quick thought experiment. Suppose you had a hundred individuals and you can, they individually, there's some probability that they will need request care even if you don't outreach to them. But if you do outreach, their in likelihood of getting re of requesting care is going to increase quite a bit. So if we set some threshold, say let's flag the top 10%. And so some of those who are flagged we'll see seek care and some who are unflagged will also seek care. But we actually, the number who do seek care is less than our overall capacity. And so we actually have under utilization and that means that if we actually flagged more people, we could have intervened and helped more. So you could say, well let's actually increase the number of people we flag and indeed then you can actually make sure you use up all of the capacity. But we find so that actually if you keep increasing at some point

there are some people that will request care that are in the unflagged portion that could actually push out those in the flagged portion. And we refer to this as cannibalization. Some of these independent requests will actually crowd out the higher risk individuals. And so what this tells us is that actually there's only some portion of thresholds that you're likely to utilize if you have capacity constraints. And what this means if for those of you who recently graduated and took business analytics, the traditional approach of trying to manage or pick algorithms is looking at an area under the curve. It says we care about the true positive rate and the false positive rate over all possible thresholds at which we intervene and we want that to be highest but aren't in the analysis says that actually there's certain thresholds you would never use. And so you could have a algorithm that has a higher A UC but a worse performance when and if you operationalize it. So I wanna just show you an example of this. In, in sepsis we compared a, the epic prediction model for whether or not a patient has sepsis to a model. We trained ourself using XG boost and we found actually the epic model had higher predictive power, it had a higher area under a curve. And so by traditional measures it would say that the epic model is better. But what turns out to be the case is that if you are in a very capacity constrained and environment, what you care about is the true positive rate at lower false positive rates. 'cause you're not gonna be able to intervene on that many people. And so in those capacity constrained strain settings, it's actually better to use a model, the XG boost model that has a lower A UC, but better predictive power in the areas that really matter. So I know that was a lot of information, a lot of different ways that AI is being used, but I, what I wanted to you to take away from this is there's a lot of developments and advancements in the use of these predictive models. If you don't think about the types of decisions you're going to base on these and how you're gonna actually operationalize it, you're never going to be able to realize these benefits of being more proactive versus reactive. You're just gonna have these prediction algorithms that tell you somebody's risk. But in fact if you intervene at the wrong time or on the wrong patient, it could actually be worse than not having the predictive algorithms at all. So it's really important to think about what is the benefits of the predictive algorithms? What are some of the real world constraints that allow you to them and how can you partner closely with those who would actually use it to have benefits. So with that I will pause, stop and happy to take any questions that there may be from the audience. - Okay. Please take a moment to enter any questions you may have into the q and a function. And we'll start off with our first question here. Professor Chan, you describe healthcare as traditionally reactive. What are the biggest structural or cultural barriers preventing hospitals from becoming truly proactive? - That's a challenging question. I think that some of the, there's definitely financial implications. The way that our SY payment system has been set up as a primarily fee for service model there is relatively limited financial benefit

and factually it can hurt our provider systems to move to a more preventative scenario. I think conceptually they, our providers want the best for the patients, but when we live in a world where 30 to 40% of hospitals have negative operating margins, it is a real financial challenge if we continue to stay in that fee for service model to try to move and shift away from doing more and to be more quality focused. I think some of the others are that, you know, cul culturally, the, especially in the US there has been more of a focus on the acute care. There's a contrast in the number of primary care providers that there are in the US relative to other peer countries. So the US has about 30% that go into primary care, 70% into specialties. Most European countries are flipped and that's because the system has been set up to facilitate more preventative care, but that hasn't happened yet here. So I think there's like an access issue is that even, you know, people wanting to find primary care physicians to facilitate that. It's, it's been more of a, a challenge as a consequence. - Okay. We have a question here from Nini and pardon if I'm mispronouncing that. Could you go deeper in the ER staffing optimization math, she love to know a little bit more about the base versus surge models and any scenarios where you reduce from the base if the volume drops. - Yeah, so yes, the short answer is the base will definitely be lower if the predictive volume is going to be lower. So that means if the staff at night is gonna tend to be lower than the amount of staff during the day, because we know volume drops a lot in the middle of the night, people are much more likely to go to the emergency department at two in the afternoon rather than 2:00 AM. The challenge is that we don't, the b the base staffing can't be shifted real time. So once you make that decision weeks in advance, you're stuck with it. And so actually knowing that you have this opportunity to surge up, that gives you the flexibility to think about what's the right amount of staffing for base, knowing if I have to, I can increase my staff a little bit, but at a higher cost. So we do see that variation. Once you're in the shift, you've already staffed the people. So if you find out volume's gonna be much lower, you're not necessarily gonna send the base staff home. You just won't surge in those situations. - Okay. We have a question from Christie. How are these types of tools and models currently being implemented in hospitals? Are hospital administrator, administrator teams hiring consultancies to implement these types of tools? Is that actively happening now? - So we have the full spectrum of what hospitals are doing. So there are some that have their own internal teams and that are deploying it with their clinicians and their faculty and their analytics teams. They will develop them in-house or they'll partner with companies or you know, other academics to, to deploy them. And some are aren't hiring consultant companies to help with that. I think the challenge is that for many of these implementations, they tend to be pilots and it's actually not, you know, if you have a strong administrative and clinical team that believes in it, it certainly takes time. But getting these pilots going, you know, there's hundreds of pilots going on at health systems. I, I couldn't even try to tell you what all of them are. I think the biggest barrier is when do we translate from a pilot to actually widespread implementation. And I think that example that started with at Kaiser with their early warning system, it was a 10 year long pursuit for them to deploy that early warning score across their entire enterprise. In northern California, there's a massive amount

of investment and patients that, and persistence that is required for that. And I think that right now some of the challenges, even when the pilots show great promise, the widespread implementation can be limited. That example that I gave at Hackensack Meridian, we saw the pilot, we saw there was an improvement in reduction in costs, but to directly integrate the algorithm into the workflow of the providers that would require, you know, a massive amount of investment in infrastructure and software development that, you know, would require, you know, it was challenging and, and meant that it's, it's actually not continuing to be used daily because of these real world challenges. - We have a question from Benjamin. How do you consider capacity constraints when they may not be fixed over the long term? For example, if current capacity is relatively high for historical reasons, an algorithm may tend to use that capacity for marginal cases since it's considered available. However, there may also be consideration to reduce capacity over the long term if the benefits of that capacity aren't high. I, - That's a great question and there have been a number of studies that refer to healthcare as, you know, supply driven demand in the sense that if you have the capacity it will be used. And so that can make it more challenging to understand what is the right amount of capacity to have. I think using data and these types of analytics, you're always going to have some variations. So maybe there are times when you know you could have done with less capacity and there are gonna be times where you could do with more capacity. And so leveraging the variation in the data that naturally arises because we don't know when patients are gonna show up to be sick. Sometimes providers, they themselves aren't feeling well and don't, aren't able to come to work. And so if you can leverage that variation to understand what are the needs, what are the implications by having lower staffing or more staffing that can create the evidence to make a better match of supply and demand. But you don't, I, you know, I feel very strongly, and for those of you who are integrated in the, the AI space, there's often this tension between what we refer to as exploration and exploitation. So exploration is the models are trying out things to learn about the environment, an exploitation that sounds bad, but it's actually saying we're doing, given the information we have right now, we're doing the best possible. And so in ai, this exploration means you may actually choose to do something that may not be the best it could be, but we don't know if it is. And so we're taking a bet so we can learn about it if there's something more, more information. I think the challenge in healthcare is that exploration means that I think there's serious ethical concerns. If you're doing something that you think might do harm you can't explore in the same way that you can in other industries. And so I think that there is a balance between trying to learn, but creating safeguards to make sure it's done in a way that does no harm. - Wonderful. And we're gonna close out with one last question here. As AI adoption accelerates, what governance, ethical or equity considerations should healthcare organizations prioritize to ensure these tools improve access to care without widening disparities? - That's a great question and I think there's kind of two things that I would note. So the first is what I was referring to with this exploration viewpoint where I think that the types of safeguards that are necessary are quite different than other industries. So I actually think that means that the development of algorithms, it might be, it certainly is worth it to sacrifice some predictive power. And so accuracy with the knowing that it's gonna create some of these safeguards

to limit the amount of harm. The, the other big issue is, you know, we all know now, if you didn't know before, you definitely know it now. AI models are only as good as the data that they are trained on. And there are many communities that have historically had challenging interactions with our health system. They've had limited access or less access, which means as a consequence, they don't show up in our data sets as much as the majority of people do. And so if we're training our models on data sets that have limited representation, the implications could mean that we have algorithms that work for many people but actually could harm other people and we just don't even know it. I'm not saying that it's done adversarially, but what it means is that a lot of effort needs to be put into making sure that data is representative. And you know, there's examples of this in therapeutics where because of different biological reasons, certain demographic of people either can't break down, they, their body doesn't create an enzyme that to break down the drug appropriately or they create a resistance to the medication faster. And it wasn't until there was full deployment of these therapeutics that these examples were realized. And so it's really important to make sure that the data has great representation. - Thank you so much Professor Chan. This was really insightful. We really, really value you taking the time and walking us through this. And thank you to our attendees for taking the time to be with us today. Our next CBS faculty live virtual talk will be held on March 17th at 6:00 PM Eastern with Professor Connor Walsh. He'll be speaking on the global rise of clean energy and in this talk he will discuss how the global marketplace is and will be affected by the exponential growth of clean energy. All righty. So that's all for our time. Thank you again for joining and we hope to see you on our next talk. Take care. Thanks everyone.