The future of computational health economics

Transcript

Russ Altman (00:04): This is Stanford Engineering's The Future of Everything, and I'm your host, Russ Altman. Today, Professor Sherri Rose will tell us how algorithms that are used to determine healthcare reimbursement can be unfair and can lead to worse outcomes for some patients, but she has ideas about how to fix it. Surprisingly, maybe using AI and machine learning. It's the future of computational health economics.

(00:28): How healthcare is paid for in the United States is a mystery to most of us, but we do know a few basic things. Healthcare systems get paid more money for patients that have a lot of diagnoses and they get less money for patients that are relatively healthy and don't have a lot of diagnoses. Unfortunately, this system of payments can be gamed. So if you happen to have some extra diagnoses coded, you might get a little bit more money. Or if you happen to not take care of a very sick patient, you can keep all the money that you got in payment. So it's worrisome.

(01:01): AI and machine learning might be part of the solution, according to Professor Sherri Rose, a Professor of Health Policy at Stanford University. She studies health economics and she has ideas about how we can look at the algorithms used for reimbursement and risk adjustment to make them more fair. That's at a macro level, that's how everything gets paid for. But she's also interested in looking at the algorithms that doctors might use at the bedside and what the quality of those algorithms is and whether they need to be more tightly evaluated before we let them loose on patients.

(01:36): You and your team work on something called computational health economics. That sounds awesome, but I need a little help. What does it mean? What are the questions we're trying to answer?

Professor Sherri Rose (01:46): Absolutely. I'm happy to elaborate on that. So computational health economics aims to bring advances from statistics and machine learning to solve major problems that we face for different types of health economic outcomes, which could be health spending and includes a variety of health outcomes as well. And specifically, my work in computational health economics aims to improve health equity.

Russ Altman (02:12): Great. So that is a beautiful definition and it sounds very meritorious. So I know that part of your work is the big picture economics of healthcare, like health insurance and who gets it and why they get it. And I know you've studied a lot of sources of inequity, so can you give us a little tick list of some of what you think are the biggest sources of how the money flow actually impacts whether people get better or not?

Professor Sherri Rose (02:42): Absolutely. And this is one of the things that I think is not widely understood when we are trying to improve health equity in the healthcare system, is that how we move money around the system has a huge impact on who has access to care, who's able to see a clinician and what types of treatments they might be able to receive for health conditions that they have. So for example, there's a formula called plan payment risk adjustment formula that aims to adjust for all of the health conditions that an individual has. And of course, it doesn't actually do that, but it's trying to make sure that health insurers get payments for the health conditions that you might have in any given year.

(03:26): And so how that formula moves money around the healthcare system creates lots of different incentives to provide care. It also creates incentives to insurers about which types of enrollees they'd like to have in their health insurance plans.

Russ Altman (03:43): Ah. So if I could just stop you there, just so I make sure I understand. So based on my diseases, what problems I have and my age probably and a bunch of things, that changes the level of payment that my insurance company might get from my employer, for example?

Professor Sherri Rose (03:59): Yes. From your health plan. Absolutely.

Russ Altman (04:02): I see. I see. And then what you were about to say before I interrupted you is that these formulae may be gamed. And so that sounds interesting. So could you please tell me a little bit more about that?

Professor Sherri Rose (04:13): Absolutely. So gaming is the correct word because we know that once something is a benchmark or a metric that somebody is going to game it, and health insurers are by and large profit maximizing. And so their goal is for any individual person to have as many health conditions coded as possible. So a lot of health conditions are zero, one binary flags that somebody is coded for having the condition or not coded. And I very specifically say coded for having the condition because there's error in both directions. You might be coded as having a condition you don't actually have, and you might be missing a code for one that you do have.

(04:52): And so the insurer's goal is to make sure that you're coded for as many conditions as you do have. And unfortunately, there's also fraudulent coding where sometimes insurers will code people for conditions they don't have in order to increase and maximize those profits.

Russ Altman (05:09): So who polices that? That sounds like a problem because we know that people... Especially when money's on the line, people can stretch the truth. So I'm sure that if you added to my medical record that I have lupus, that somebody's going to make more money from me, who's policing the accuracy of the description of each patient?

Professor Sherri Rose (05:30): So the formulas are regulated by the federal government as far as which variables are included. And then sometimes this relies on whistleblower reporting. So when there's fraud for Medicare, which is a system of health insurance that we have in this country for individuals who are 65 or older and certain disabilities as well as end-stage renal disease. So whistleblower reporting can be a huge component of finding out that there's actual fraud occurring. But as far as what might be considered aggressive coding where they're maybe sending a clinician or some kind of health provider to someone's home to make sure that they get their diabetes code that year, that type of behavior is not illegal and is not going to be supervised in any particular way.

Russ Altman (06:25): Okay. Okay. So many people who would hear this would then think, "Okay, fine, they're playing all these kinds of games, money is flowing. It may or may not be accurate," but here's the question. Is it affecting the quality of care that the individuals at the end of this pipeline actually get? Or is it just moving money around different ways and it really doesn't change the way me and my doctor interact or the decisions that my doctor may make about me?

Professor Sherri Rose (06:52): It absolutely has an impact on care. So if you think about somebody who has... Let's use the diabetes example again. That's an individual that on average is going to need more care than somebody who does not have diabetes. An individual living in a rural community, who lives very far away from their clinician maybe is not going to get the care that they need, but the health plan, the insurer is still going to be compensated as long as they have a code for diabetes, they're still going to be compensated at the average level of cost for an individual with diabetes. So now they're being overcompensated without providing socially equitable care for their patients who are actually not being seen as frequently as they need to be seen. And so they have no incentive to provide the care that would be required because that patient or that group of patients that is not being seen makes them money, is profitable to them.

Russ Altman (07:58): And I see. And I am aware of this phenomenon where a patient is far away, they live rural and they can't get to their healthcare provider and they just can't get the bus, they can't get a ride. And so it really is almost like a sin of omission because they just don't show up to their appointment and the insurance company can just shrug its shoulders and said, "Well, we would've paid for it if they showed up, but they didn't show up." And in a funny way, they're actually starting to blame the patient, whereas all of this was entirely predictable.

(08:27): So I wanted to get into who winds up getting the short end of the stick? I'm going to guess that I'm a professor near a major city, who works at a major medical center. I'm going to guess that I'm not going to be the one who gets shafted by these rules, but maybe I'm wrong. Who are the people who are going to wind up really suffering from this system?

Professor Sherri Rose (08:52): There are many different groups in the healthcare system that are marginalized in general, and then specifically by the risk adjustment formulas, it includes racial and ethnic minorities, individuals with disabilities, and broadly, individuals who have multiple chronic health conditions, individuals as we discussed, who live in rural communities and older adults as well.

Russ Altman (09:13): So what can we do about this? I know that part of your research is not just documenting all of these phenomenon, but thinking about how the system needs to change. Is that clear to you?

Professor Sherri Rose (09:25): I would say there's a lot of different avenues that we've pursued over the last decade, looking at how we can improve the efficiency and fairness of planned payment risk adjustment formulas so that they move closer to having justice for all these groups that we said were marginalized so that the risks, benefits, harms and resources are more equitably distributed across these groups. And so some of those things have included considering trying to make the formulas less gameable, and that involves potentially using machine learning and variable selection methods, including different types of fairness constraints to try and make sure that the formulas may perform better for some of these marginalized groups.

(10:09): And so there's a lot of different pathways towards trying to make these formulas better, but I would stress that it's going to be a continual process because as I said before, once you define a new formula or a new metric, there's going to be ways to game the system. So I'm really optimistic that there are many improvements that we can make and that there's this important role for researchers where we're not only doing primary research, but we're also working to communicate that research to policy makers in order for them to see what paths are possible. But it's not like we can make a few changes, it'll be perfect. It won't be perfect, but it would be improved. But then continuing to study this over time because there's going to continue to need to be changes down the line as well.

Russ Altman (11:08): So I was very interested that you brought up the topic of... You said either AI or machine learning, I can't remember. Very interesting because as you know in the popular discussion these days, AI is not the hero of the story. Very often people are worried about exactly the issues that you're trying to combat in terms of inequity, unfairness. So take me through why these AI machine learning algorithms might be the source of more fairness, where what we hear about in the news is how they can be the source of less fairness?

Professor Sherri Rose (11:40): Absolutely. I have so many thoughts on this, so I'll try to walk through a coherent stream of consciousness here. So firstly, when people talk about machine learning and AI, it is absolutely overhyped, but so many of the algorithms that are called machine learning and AI are often just a very standard type of regression, is something that we've known about in statistics for decades and decades and decades. What is included in the machine learning and AI bucket has grown to basically include any time we're analyzing numbers.

Russ Altman (12:15): Yes.

Professor Sherri Rose (12:15): And so this is not a new problem that any of these types of tools that we might be using, these quantitative tools can lead to systematic unfairness. There are more tools and some are worse than others. And as I said, broadly, many of these are massively overhyped. But the issue of worrying about and considering really carefully the impact of an algorithm before it would be deployed is another area that I work in, and it's something that we don't do functionally at all.

(12:49): It's often putting something into practice without having really studied it very well in the small samples that it might have been studied in, or even if it's a large sample, it's a single database. And in healthcare and medicine, it might be at a single medical center.

Russ Altman (13:06): Right. Right.

Professor Sherri Rose (13:06): And so that doesn't address things like generalizability, how it would apply to a larger population, but those algorithms are usually not evaluated for their impact on marginalized groups, even prospectively. But then it's thrown out into use and not continued to be studied. It's not evaluated at all. And then, oh, we look and it's been in practice for two years, for five years or 10 years, and it's causing a lot of disproportionate to harm to groups that were already marginalized by the healthcare system.

(13:40): And so I absolutely agree that machine learning and AI can have huge impacts, negative impacts, downstream impacts on historically and currently marginalized groups. But I'm saying that the problem's actually bigger and that this is not a new phenomenon. This is something that we've known about for a long time, and that the whole motto in Silicon Valley of move fast and break things absolutely does not apply when these algorithms impact real people. And in healthcare, that's almost always the case. And so we need to be really thoughtful and methodical before we would deploy a new algorithm. And I've said algorithm a number of times recently. I mean that very broadly. Any kind of tool that takes data as an input and return summary outputs is an algorithm.

Russ Altman (14:33): Yeah. Thank you. That really makes sense. And I do want to talk about those algorithms because I presume because you do computational health economics that one of the main classes of algorithms would be these ones that make these risk adjustments based on the diseases that somebody has. Are those closely held by the companies in the sense that it's hard for you to examine them and understand their performance? Or are these publicly available and inspectable? And as you try to think about using your ideas and methods to create new algorithms, what's the path to get them to be used by either the government or the insurance companies?

Professor Sherri Rose (15:14): So some of the algorithms that I study, like planned payment risk adjustment algorithms are... I say available. You have to dive into really long documentation from the federal government in order to figure out all the nuances of exactly how they built it.

Russ Altman (15:32): Yes.

Professor Sherri Rose (15:32): But there is that documentation available. In some of the other areas that I study, especially algorithms that are within the healthcare system and might actually be used at the bedside by clinicians, many of those are not transparent, they're completely opaque. Even if you dive into the documentation from these for-profit companies that sell their algorithms to healthcare systems, you can't really completely figure out what that algorithm is. And some of these impact millions and millions of people in the US. There's been a small amount of attention on these opioid overdose risk algorithms from this narcs care algorithm that really is having a huge impact on individuals getting pain prescriptions.

(16:29): And this is something that's used in all 50 states. And if you dive into the documentation trying to figure out what the algorithm is, you will not be able to understand exactly what they did. And it's a lot of vague language about even the form of the algorithm, what type of algorithm it is and what variables they used. And so we're actually have been working on a project trying to reverse engineer this in order to try and evaluate it more comprehensively because it has led to a lot of downstream negative impacts, including people with chronic pain conditions not being able to get their medication and this opioid overdose risk algorithm not really serving the purpose of trying to prevent overdoses, but really just penalizing people from having previous prescriptions.

(17:21): And so I would say that in the medical space broadly, a lot of the algorithms are not transparent. We don't know how they were created, and that is part of one of many issues in trying to improve them or even saying, "This algorithm shouldn't exist," which should always be an answer.

Russ Altman (17:42): This is The Future of Everything with Russ Altman, more with Sherri Rose next.

(17:47): Welcome back to The Future of Everything. I'm Russ Altman, and I'm speaking with Professor Sherri Rose of Stanford University. In the last segment, Sherri told us that the healthcare reimbursement system can be unfair. It can lead to gaming, which leaves certain populations, even though there's money set aside for their care, not receiving that care for a variety of reasons. In this segment, she will tell us that even the programs used at the bedside for doctors might not be fair and might not really even be ready for use in any situation. It's kind of a Wild West of AI systems in medical care, and we need to pay attention to that.

(18:27): She ends up optimistic, however, because a new generation of researchers is considering the full picture of the social situation of healthcare delivery as well as the algorithms, and that'll lead to better outcomes. So tell me about the source of your concern and what kind of work you're doing?

Professor Sherri Rose (18:44): Absolutely. The problem with a lot of algorithms in the medical space and broadly the public health space as well, is as I noted briefly earlier, these algorithms are not well validated. So we have a lot of individuals whose incentive is to produce an algorithm that might be potentially technically interesting or they want to quickly, "Solve," a problem by bringing machine learning to an area that it hasn't been brought to before. But unfortunately, we don't have minimum standards for this type of work, for machine learning in health. There's no minimum standards set by journals that they require or funders that are required. And so it leads to a lot of low quality publications in the medical literature and the health literature broadly.

(19:34): And I think that it becomes a launching pad for bad algorithms then coming into practice, because then we also have other groups doing this. We have startups, we have for-profit companies creating algorithms that have no basic minimum standards. And so to be very blunt, the vast majority of machine learning for health algorithms are garbage. And they're not going to generally be useful. They may be incredibly harmful, and we should absolutely not be deploying them because they haven't been well-designed, they haven't been well studied, and they aren't going to lead to the impact that we would like them to have and could actually create more problems.

Russ Altman (20:17): And I'm going to guess they don't come labeled as garbage. So it's not an easy task of just saying, "Oh, I just want to buy the ones that are not garbage." Which gets me to my next question, which is, okay, as a patient or as somebody who cares about healthcare, how can I tell, A, if my physician is using an algorithm, and B, whether it's one of the garbagey ones or not? This sounds like a hard problem.

Professor Sherri Rose (20:39): That's absolutely one of the issues is you don't even know that an algorithm has been used as part of your care and whether it may be guiding a clinician to make a different decision. And sometimes this can due... They might be afraid of being sued. So that's an example. Going back to the opioid overdose risk algorithm, if somebody says this individual is, "High risk," to give an opioid prescription, and a clinician might be afraid to do that, even if otherwise, they would have done that.

Russ Altman (21:13): Right. Okay. So let me step back because I'm sure there are people in the industry who would be thinking that we're not being fair to them by saying that their algorithms might be garbage. So how do you even tell if an algorithm is garbage? You made a strong statement and it makes sense to me, but would a student or you be able to get one of those algorithms and then put it through its paces and say, "I'm sorry, but this is not performing like it should," or is even the definition of what is reasonable and what is not reasonable up for debate?

Professor Sherri Rose (21:46): I think both are true. So we haven't decided what's good enough to be used in healthcare, and that is also often going to be conditions specific. And we haven't agreed on the metrics. We haven't agreed on what we should be looking for and what the bar should be. And I can say that if you're reading the literature, you can tell what's garbage or what at least hasn't been well proven based on what's included in the paper. Sometimes really basic evaluation statistics are hidden in the appendix or not included at all. A lot of times it's, "We did machine learning, we did machine learning," and you have to dig and dig and dig to find that the actual algorithm that they deployed is very close to standard techniques from statistics and maybe some tiny little difference between a standard approach. But that's often one of the tells is that, okay, there's this tiny, "Improvement," compared to standard practice and it's being overhyped.

Russ Altman (22:49): Yes.

Professor Sherri Rose (22:50): This is going to change everything, solve everything that should be immediately deployed. The conclusions in the paper are often woefully overstated.

Russ Altman (22:58): Yes.

Professor Sherri Rose (22:58): But I think a discerning reader of the literature can often have trouble because a lot of this research, they do not make clear.

Russ Altman (23:12): That really makes sense because I know in the setting of clinical trials, which we've been doing for literally 100 years, when you submit a paper about a clinical trial, there are certain mandatory elements that you're supposed to include and the reviewers will even send it back without even reading it if you haven't included certain topics. But as far as I know, for AI and medicine, such a list of mandatory elements has not been created or agreed upon. And so it is a little bit of the Wild West right now, isn't it?

Professor Sherri Rose (23:41): Yes. There's been a lot of different proposals, and I've been part of them as far as, okay, what should it be? But when you have 10 proposals and not one of them has been agreed upon by the community and there hasn't been uptake, absolutely, you can send your paper to these clinical journals. And because machine learning and AI is such a hot topic, some of these journals are really incentivized to take papers that have really fundamental issues. And I've been a reviewer many times where I highlight the same three or four rejection worthy issues. And oftentimes the editors are really appreciative of the thoughtful comments but it's still, "We still want to take this." And then they take it-

Russ Altman (24:22): Because it's going to be on the cover of the newspaper the next day.

Professor Sherri Rose (24:25): Yeah. And they know that readers are going to look at it. And there's a clinical journal where... And there's probably more than one depending on the day where you look, and you see all their top 10 most read papers are the machine learning ones. And so now the incentive is, "Oh, let's publish a bunch of popular garbage."

Russ Altman (24:44): So it sounds like one of the solutions is editors who have some courage to say, "We want to see some minimal things reported, and we are not going to review your paper on AI in medicine unless it includes and addresses these." And that's something that we've seen in many areas of science. It would not be revolutionary. It would be just as expected. And so that makes good sense.

(25:07): Well, in the last couple of minutes, I wanted to ask you about the future and about how things are looking in terms of solving these problems? You have a vibrant lab with lots... I think, a very diverse lab. So tell me about what do the teams look like who are starting to work on these projects?

Professor Sherri Rose (25:24): I'm happy to elaborate more on that. That's one of the things that gets me so excited about the future are the students, because there's more and more students, especially from technical fields and computer science and biomedical informatics, statistics who are interested in grounding their work in algorithmic development in the social systems where these algorithms will be used and understanding things like social determinants of health and really being aware of the fact that the development of tools that are better cannot be divorced from how it would be used. So it's not just about technical solutions of optimizing loss functions. It has to be that, plus grounding it in these systems.

(26:05): And I really enjoy the fact that I work with students from a variety of different prior training, their current PhD programs. I also work with students in the health policy who focus on economics, for example, and having all of this disciplinary exchange makes the tools better.

Russ Altman (26:27): Yes.

Professor Sherri Rose (26:27): And it's really required for us to be doing that, to have thoughtful, not just disciplinary training, but just people's life experiences. We need more people who have experienced the gaps in healthcare to be part of the solutions in creating healthcare systems that actually function better for more people. And so if we systematically exclude people who have been failed by our health system, we're not going to get closer to health equity.

Russ Altman (26:56): Thanks to Sherri Rose, that was the future of computational health economics. You have been listening to The Future of Everything with Russ Altman. You can follow me on Twitter @Rbaltman, and you can follow Stanford Engineering @StanfordEng.