Code & Cure
Decoding health in the age of AI
Hosted by an AI researcher and a medical doctor, this podcast unpacks how artificial intelligence and emerging technologies are transforming how we understand, measure, and care for our bodies and minds.
Each episode unpacks a real-world topic to ask not just what’s new, but what’s true—and what’s at stake as healthcare becomes increasingly data-driven.
If you're curious about how health tech really works—and what it means for your body, your choices, and your future—this podcast is for you.
We’re here to explore ideas—not to diagnose or treat. This podcast doesn’t provide medical advice.
Code & Cure
#14 - Medicare’s WISER Pilot: AI, Prior Auth, and the Cost of Care
What happens when an algorithm—not a doctor or a claims reviewer—denies your surgery? A single decision like that can trigger a much bigger conversation about how AI is reshaping access to care.
In this episode, we dive into Medicare’s WISER pilot and the complex world of prior authorization. What’s the goal? Reduce waste and streamline approvals. But where does it go wrong—and how can we fix it? With insights from AI researcher Vasan Sarati and emergency physician Laura Hagopian, we unpack how claims data trains decision-making models, why black-box algorithms erode clinician trust, and what real safeguards look like in practice.
We spotlight three high-volume services—skin and tissue substitutes, electrical nerve stimulators, and knee arthroscopy for osteoarthritis—and explore why these procedures made the list. Knee arthroscopy, in particular, becomes our case study: widely performed, weak evidence in most OA cases, but not without its exceptions. That tension reveals deeper risks: overreliance on flawed data, quick “human reviews,” and denials that feel rubber-stamped.
Then we imagine a better way. What if AI could argue with itself before deciding? Enter multi-agent models—a system where different specialized AIs represent the patient, the provider, and the payer. They debate function, evidence, policy, and risk—and their decisions come with plain-language justifications, escalation triggers, and audit trails. The goal: approvals that are not just faster, but fairer.
If you care about timely access, fewer roadblocks, and smarter AI guardrails in healthcare, this episode is for you. Subscribe, share it with a colleague, and tell us what you’d change about AI-driven prior auth. We’ll highlight listener ideas in an upcoming show.
Reference:
Private health insurers use AI to approve or deny care. Soon Medicare will, too.
Lauren Sausser and Darius Tahir
NBC News (2025)
WISeR (Wasteful and Inappropriate Service Reduction) Model
Cemter for Medicare and Medicaid Services
Credits:
Theme music: Nowhere Land, Kevin MacLeod (incompetech.com)
Licensed under Creative Commons: By Attribution 4.0
https://creativecommons.org/licenses/by/4.0/
Knee surgery denied. Not by a doctor, not by a claim specialist, but by artificial intelligence. Today we dig into how Medicare's new AI-driven process is changing the rules of access to care.
SPEAKER_00:Hello and welcome to Code and Cure. My name is Vasan Sarati, and I'm an AI researcher. And I'm with Laura Hagopian.
SPEAKER_01:I'm an emergency medicine physician.
SPEAKER_00:Yeah, today. So we're going to talk a little bit about the um the Wiser program, right?
SPEAKER_01:Yeah, it's a new program that uh Medicare is initiating. It's like actually a really good acronym, I think.
SPEAKER_00:Yeah, it's wasteful and inappropriate science reduction. Service, service, reduction. Service reduction, sorry. Uh Wiser, yes.
SPEAKER_01:It's like you're gonna make Medicare smarter with AI. Right, right. And that's it's a good marketing move. It's a good acronym, like I'm telling like I'm saying.
SPEAKER_00:Yeah, they just need a logo of an owl and they'll be all set.
SPEAKER_01:Exactly. Um, but uh it's basically about figuring out whether or not Medicare can kind of automate the prior authorization process. And it's a pilot program. Yep. Um, but the whole concept behind prior auth is that you know, a provider must get approval from their insurer from an insurer before something gets covered, whether that's like a medication, um, whether that's uh a treatment, whether that's a surgery. It's a way for insurers, including Medicare now, to make sure that something is medically necessary and it helps manage costs.
SPEAKER_00:Right. And this is something that private insurers have been doing anyways.
SPEAKER_01:Well, Medicare does this anyways, too, I'm sorry, but but they don't use AI to do it until now. Got it. Right. And so it in some ways it makes sense, right? You as an insurance company, you would not want to spend inappropriate money for something that's, you know, not evidence-based, right? You want to make sure that if someone is having something done, whether that's an expensive medication or a surgery, that it's actually necessary for them to have it done.
SPEAKER_00:So just to be clear, uh the the process would be that a doctor and a patient would work together and they would decide that a medicine or a procedure needed to happen. And then the insurance company would have to, or whoever is providing the insurance should have a separate say in deciding whether that med decision is going to be covered by the insurance. And they would like to do that before the the procedure or the medication. Or the medication, right?
SPEAKER_01:Yeah. It's like, you know, when you say like that, and some sometimes people think about this, like it's it's meddling a little bit, right? The doctor might have decided, hey, we should do this, we should do this procedure. And the insurance comes back and says, like, no, we're not covering that.
SPEAKER_00:Yeah.
SPEAKER_01:Um, on the one hand, you know, could a provider recommend something that's not actually medically necessary and is expensive, and then, you know, the insurance comes back and says, no, don't do that. We're not covering that. Then it sort of makes sense. But there's all sorts of stories out there that are not like that, right? Um, but from the insurance standpoint, you want to make sure it's necessary and you want to control your costs, right? You don't want to cover things that are expensive and um some things that are deemed inappropriate. Right. And like a common example might be, you know, if you walk in with two days of low back pain um and you don't have any sort of whatever red flag symptoms, you don't need an MRI right away. Usually we tell people to like wait it out. And if the symptoms aren't gone after about six weeks or so, then we start to think about imaging. Um, so the whole, the whole concept is like, let's let's figure out what's necessary, let's figure out what lower cost alternatives there are, and let's try to control the costs.
SPEAKER_00:Um That being said, that being said, so it's a bit of a dance between the insurers and the and the and the providers in that sense.
SPEAKER_01:Right. And and who's stuck in the middle? And this is the problem with American healthcare so many times. The person's, you know, it's the human, the patient is stuck in the middle, and they're like not the one with power in this situation. And they're just like kind of at the whim of what's being decided almost without them. Right. Right. And so, you know, uh a provider might make a recommendation, then the insurance might say, no, we're not gonna cover that. And maybe it was appropriate to say that, but maybe not. Like, right? There could be patient harm that comes from this. There could be poor outcomes that come from this, there could be delayed care that comes from this. Um and not not only that, but it's a huge administrative burden as well, right? These prior auth forms take forever to fill out, right? And AI can be used to try to automate filling out the prior auth forms, but providers have to have this extra administrative burden of, oh, you need to go on this expensive medication like Humerima. Now we have to fill out special forms. We have to hire someone in the office to fill out these special forms so that you can get on the med that you need to be on. So it's just like this this added burden. And that kind of a burden increases burnout, not surprisingly. And we've talked about burnout before, how it's multifactorial and administrative burden is is one piece of that. Um, and you know, in some studies, we're seeing the number of prior authorization denials increasing. So now you're spending your time not with patients, but like sifting through all this paperwork, diverting your time and resources um in the office to work on these tasks.
SPEAKER_00:Yeah, because now you have again, you have to do that dance, right? You have to figure out what how to fill out the forms or what what you know um procedure or you know, prescription that you can give that will be not denied, will not be denied. And you know, that's that makes it more challenging for sure.
SPEAKER_01:Right. And so you could imagine that someone who a patient who's stuck in the middle, um, and is in this sort of waiting cycle, yeah, a month or two months to see if they can get this test or get, you know, that medication, they might still be symptomatic and bounce back into the healthcare system.
SPEAKER_00:Right.
SPEAKER_01:And then they cost money, but in different ways. Right? They're not getting the treatment that was recommended because it was denied. Um, and so they might end up in the ER. They might end up investing more money in the hospitalization, they might have an additional office visit, they might have some sort of poor health outcome. So it's it is this dance because of course we want to make sure that the treatment someone is getting is appropriate, is medically necessary, and is cost effective. But at the same time, you can see how this could potentially go wrong.
SPEAKER_00:Well, yeah, because determining whether something is cost-effective is more than just that one point uh where they determine whether that specific prescription or procedure is necessary. There's a whole host of other factors that come into play. And that makes it even more complicated to determine what cost-effective even means.
SPEAKER_01:Yeah, and you know, there could be algorithms, um, and there are certain things that are, you know, clearly yes, this should be covered. Clearly, no, this shouldn't be covered. But there's probably some gray areas out there, right? Where there's some nuance. Um, and each individual needs to be looked at to some extent to figure out oh, what's happening in this person's life and this person's medical history to see if it's necessary for this exact person? Yeah. Um, and so it's an it's an interesting application of AI because you know, are there ways that it could improve efficiency here? Um, potentially, right? Like I was saying, sometimes it can take a month or two to get these like prior auth denials to come back, even right. And then you're like, well, it got denied. Now what do I do? Right. So, you know, could it speed things up? Could it uh kind of weed out the low value services more easily? The potential is there. Yeah. Um, and the prior auth processes already already exists, but on the flip side, it's like, well, how is AI seeing the nuance that might be present with one person? And this is why I'm glad they're launching this as a pilot program, right? They're only doing this in a few states, and they're only doing it for a few different services with this wiser model. They're doing it for skin and tissue substitutes, they're doing it for electrical nerve stimulators, and they're doing it for knee arthroscopy for knee osteoarthritis.
SPEAKER_00:I wonder why they picked those specific ones as opposed to others. Um, it's always interesting to ask that question. I'm sure there is no clear answer, right?
SPEAKER_01:Yeah, I mean, I I don't know exactly. I could sort of conjecture, but I think um they may be sort of more straightforward in terms of what does the evidence say? Like, can we create an algorithm for this? Got it. The easier it is to create an algorithm or say, oh, this is inappropriate or this is low-value care, um, the easier it is to probably like give AI a rule, right?
SPEAKER_00:Yeah, yeah. I mean, it depends on the nature of the AI system. I think that, you know, I I do think that AI has been used quite a bit, or I should say more generally, computers have been used quite a bit to assist with this whole process of prior authorization. I know that uh, you know, over a decade ago, it was primarily used to help fill out forms or other things relating to the process of doing the prior authorization. But more recently with machine learning, um, AI has sort of come into the come into play in different ways. And we don't know exactly what models uh they're gonna be using here. We also don't know what exactly insurers, what models insurers use. That's a lot of it is proprietary and behind closed, you know, sort of closed um closed doors. But um they typically use machine learning models that, you know, swallow in a whole bunch of clinical data. And um they have data sets where they have prior insurance claims data sets that they use to train these models up and to help them make the decisions whether something should be denied or accepted. And um again, it's very unclear what the specific models are and also which part of the um prior authorization process the models are actually being used, right? Are it it seems like that again they could be used in various different stages. Um and you know, but the the notion of having it be mode for predictive analytics is might be where was might where it might be that that's where they're using it the most um to to to especially for this program.
SPEAKER_01:Yeah. Um that's interesting to think about, okay, it's not it's my mind goes to okay, they're gonna just deny things with AI, but you're saying that's it could be used in other parts of the process, like synthesis of the clinical background or whatever.
SPEAKER_00:Yeah, because I mean, denying everything is not going to be beneficial to anybody.
SPEAKER_01:No, I mean, we talked about how you can actually increase healthcare costs by doing that, right? Because you get people bouncing back, going to the ER, seeking a second opinion, uh, getting hospitalized, whatever.
SPEAKER_00:Yeah. And so I can't imagine that these systems use any more complex machine learning models than other use cases. Uh and machine learning, you know, if you go back to the basics of it, is essentially giving it a bunch of data and asking it to find patterns in the data, and then using those patterns to predict a future item in data and tell you whether or not something is, you know, should be accepted or denied. And so they uh it depends on the nature of the data that they're using, but the more data they have, the more points of contact they have uh for that specific person, the more likely it is that they're going to be doing a better job predicting. Uh-huh. You know, the more information they use um for that patient, the more accurate the prediction is likely going to be. And when I say more information, I mean more relevant information to make that determination, right? And so the assumption is that the insurance companies, or even in this case, Medicare, has all that information. The information that was previously used by humans to make that decision is available to the aid to train the AI system to be able to find those patterns of human decision making and apply them again in a new case.
SPEAKER_01:It's interesting because you know, uh the AI could synthesize more information than maybe a human would be able to when they're looking at so many of these over and over again. So there's like potentially an advantage there.
SPEAKER_00:Yes. And there's in fact, so that's kind of what's being touted here is that the benefit is that it reduces human error and human bias. And human error can come from just tired people, right, looking at things. Yeah, all day long. Right. And that doesn't happen with a machine. Um, and then there's issues of bias where humans might be biased in a specific example, in a specific unique case, and that doesn't happen here because the AI system doesn't care about one specific human versus another. It's just another data point, and so it's making a prediction based on that. But on the flip side of that is that you end up um not caring about the unique needs of the particular patient, which might actually be important and which might actually be the most important thing, in fact, overriding all the other data points, right?
SPEAKER_01:Right. And that's why you need, and we talk about this all the time. That's why you like need a human in the loop to review what the what the output from that AI is and say, oh, geez, like this makes sense, or perhaps not, perhaps this doesn't make sense.
SPEAKER_00:Yeah. And they do have talk about human in the loop and having sort of a meaningful um human review. Uh, but of course, the challenge there is how much time does a human have to review this? And I think there was some data that I saw a while back about this um from Cygna, and they were talking about how a human maybe spends one to two seconds looking at each case. And that's crazy. Yeah. Right. So that's clearly not enough time.
SPEAKER_01:And also, that feels really crappy if you're the patient on the other end of it. Like, oh, they spent two seconds to decide, like, and I found out two months later to decide, like, no, I'm not getting this medication that I that my doctor recommended for me.
SPEAKER_00:Right. So I mean, the one to two seconds is a very short amount of time, obviously. And so what is it? I don't think that even means that's even meaningful in terms of a human review. And if that human is tired, then they're just gonna accept whatever the AI system is saying and not think about it, right? And so I think that that's another piece of it which um is important. And, you know, again, we've talked about this before. Um but these machine learning systems are often black boxes, especially with deep neural networks, which is probably the case here. Um, they're gonna be big black boxes. So we don't know how they make the decision. We just know that they've identified some pattern in the data that they were trained on about how humans make decisions on this c on on these prior art situations. Um, but we don't know beyond that um what the actual logic is. Um and so it might make a decision, it might find some pattern that we didn't think about, or some pattern that really humans don't use, but is relevant in that specific set of data and use it, in which case there's no way to know what how it made its its decision.
SPEAKER_01:Yeah, it's weird because like if if a clinician were to be reviewing it, wouldn't you want that explainability there? Like this was denied for this reason, or this was approved for that reason. Yeah. Right because otherwise it's like they're not reviewing if it's just a simple yes or no and they don't they have two seconds to like look at it, they're not reviewing the whole chart or the synthesis or anything like that. You'd want that sort of explainability there.
SPEAKER_00:Yeah, but I think that so, so um the you know, in AI research and people developing these algorithms and these architectures, they are working on these problems, right? There are people, there are people working on explainability and interpretability and providing and all of these details. So that means that we will have the technology, or we maybe even have bits and pieces of this technology already available for this purpose. And so we should be able to do this. However, um, what are the incentives for these companies to actually do these kinds of things? Um, is there is there an incentive for them to actually incorporate AI models that are more interpretable?
SPEAKER_01:Right. It's like that's helpful for the patient and the provider, uh, but not uh always the top thing for the uh insurer.
SPEAKER_00:Yes. So it's all about the incentives, right? What is the incentives? How are the incentives lined up? And uh and I think that uh if explainability is important, then I would like to see more of that in this program. And I would like to see when this program gets into play in this pilot program, that they people actually care about that and want to know that. And you know, and and and that's required. That's it's almost like in Europe, they have all of these rules about how AI systems have been deployed, must have certain characteristics about them and explainability and all those things are pieces of that. Um and those are regulations that are in place to ensure that companies do that. Uh, maybe we need something like that here as well in the United States. So I I don't know exactly what's gonna happen with the program specifically, but I would urge them to think about you know how they're gonna incentivize explainable AI models.
SPEAKER_01:Yeah, I completely agree with that. Uh let's humor me. Let's take one of these examples because I think that might let us dig in a little more. And I think, you know, knee arthroscopy for knee osteoarthritis is a is a great example because um a lot of the clinical guidelines show that having um someone do this surgery on your knee for arthr for a specific type of osteoarthritis, arthritis in the knee, has basically like almost no clinic clinically meaningful improvement. It leads to, it doesn't lead to better outcomes, essentially. Like people's um quality of life, their ability to function and do activities of daily living, their pain levels, um, they're not any different. So it's like, why do a procedure if it doesn't help, right? Right. Why not just do physical uh therapy? Why not just do these non-surgical interventions like an injection that that are less less costly and they work just the same? Um, and also knowing that surgery is a big thing, right? It could lead to infections, it could lead to um uh it can lead to other complications as well. And so this is an example where it feels like, okay, this is maybe an easier place to implement um AI because we the evidence base for doing this procedure for this indication just like isn't there yes.
SPEAKER_00:I mean, so that just suggests a you know sort of blanket denial, right? For this kind of thing. Like, why do you even need AI for this?
SPEAKER_01:Yeah. I mean, that's the thing is like, yes, why do you need AI? This is the so possibly some people could get surgery for it if they have like um severe mechanical symptoms or if they're unresponsive to conservative management at those points. Maybe like if they have locking of their knee, um if they've got like a loose body and it doesn't respond to physical therapy, maybe those are the ones who do need it. And it's a rare case, but it's not like never. Yeah. So will the AI be able to sort of identify well, I think that's those case example.
SPEAKER_00:This is a great example because I I worry that this is the kind of example where the AI will fail. Because especially in the case where if it's evidence-based that most of the times it's not worth doing the surgery, then um what most of think about all the data that the AI is being trained on. Deny, deny, deny, deny, deny, and maybe a couple of accept, accept, but most of it is deny, which means that the ta the data is very unbalanced towards denial, which means that the likelihood of the next data point being deny is very high, regardless of something specific. Now, that's assuming that you have taken into account the extra little features that make something from go from deny to accept, like some maybe some specific aspects of that patient, right? Even if that was taken into account, the AI system has to recognize that that feature has more weight than everything else. And that's what causes it to accept it. And but that's assuming that there that's being fed into the AI system, right? Like you you just described a situation where a patient may actually need it. And but there's all these other factors that are needed to be seen in order to make that determination. Now, is that being fed into the AI too? Was that part of the training data? Right. Right. And so otherwise you're going to get the issue, which is I think the biggest issue I I personally think is that you have um an averaging that it's happening that doesn't take into account unique situations for that patient. And if if it's not done correctly, if it's not done well, then you're going to have that problem even here. Um and especially here where there isn't balanced uh data available. Yeah.
SPEAKER_01:This is fascinating because prior auth comes up in general as like a problem for providers, right? You're like, oh, another prior auth form we have to fill out, or another denial we have to deal with. And this just adds another layer to it. Um, of okay, how is AI making this decision? And um, and is it really taking into account the unique needs of my patient?
SPEAKER_00:Yeah. And you know, I've talked about it the way I've talked about AI through this uh podcast so far has been sort of this old school machine learning format. You feed in a bunch of data, you do a prediction, predict, you train a predictive model, and then it does yes or no for everything. That might not be how we do it going forward. Uh going forward, we might leverage some of the large language models that people have developed, chat GPTs and such, uh, to build an architecture that has uh what they call an agent-based architecture, which is a system in which you have a bunch of different agents or different LLMs kind of all working together in concert, trying to figure out if this patient should be accepted or denied. This is a little bit different, if you noticed, from um the other case, because here you might be able to give it all the information, including the stuff that makes it special and unique. And you might be able to uh design a team of agents, each with different you know, sort of preferences, uh debating and deciding as a group whether or not something should be accepted or denied. And these sorts of architectures are you you're seeing them more and more in various settings, um, where you have one machine learning model uh questioning the judgment of another one, and they work together in order to figure out if this is in fact the case. And they do this very fast, right? It's not like we are sitting around and you know, chatting over a coffee or something about what this this is gonna what's gonna happen. It's more, it's instantaneous, very quick. But the point is that um you they are looking into AI more broadly than what I just described as the old school machine learning, unbiased data and predictive analytics. It's more than that now. And I think there is maybe potentially some um ideas that people are gonna come up with to help improve this overall approach here.
SPEAKER_01:This is so cool. I feel like we're gonna have to do an episode about this. It's like having a bunch of different people at the table with different motivations or different points of view that are coming together to figure out what's the best path forward. And it's not, I mean, at the end of the day, for prior auth, it ends up being a simple yesno, but like, how do you get there? And um, gosh, it would be nice to have, you know, a little patient agent advocating for themselves, in addition to, you know, like um the Medicare agent that's like, oh, we have to make sure it's um, you know, net needed and and uh as low cost as possible. And then you could have a provider agent that's like, my patient really needs this, and here's why. That's right. And the patient could be like, Oh, I can't walk because my knee is bothering me. Am I doing this right? Am I creating the correct? Am I creating agents?
SPEAKER_00:I think you are. I think that and and you know, you could, you could, you could these models could be given the information, given all the data, right? And they could process all the data. All the all the agents could be given all the data and talk about, talk to each other about it and come to a consensus, um, or some other means of uh a collective judgment uh that is you know acceptable. And then you have some sort of explainability, some degree of explainability, right? I mean, these agents are still pretending to be humans and and saying things, uh, but at least at that at some level you have this dialogue conversation uh recorded, and you have the ability to go and say, okay, this is how that judgment was made. If it was made in error, then here's here are the reasons why it could be made in error. So that's that's potentially a path forward, um, especially given the modern we have modern day LLMs that can do that kind of chatting.
SPEAKER_01:That feels like it's actually wiser to use these multiple agents. Do you like how I did that? I just I just brought it back. I just brought it back to the owl, to the wasteful and inappropriate service reduction model.
SPEAKER_00:Well done.
SPEAKER_01:Thank you. Thank you. Well, I think we can um we can end here. You know, prior authorization, it's um it's it's a tough thing. It's tough for providers, it's tough for patients, um, but it definitely has some utility, especially from the insurer's point of view, because we want to make sure people are getting things that are necessary and appropriate and um, you know, cost effective. Right. And so, you know, can AI help here? Uh, you know, we'll we shall we shall see. Um, but it's interesting to see what CMS and Medicare will be piloting um in a small subset of of the population for very select services and and and in select geography. So it'll be interesting to follow up on this one. And I don't know, I feel like I gotta learn about this multi-agent stuff a little bit more.
SPEAKER_00:Yeah, I think that is there is a huge future there. I'll put some uh links on the on the on in our show notes as well so people can explore this as well. Um, but yeah, it's very exciting.
SPEAKER_01:Yeah, and we'll see you next time on Coding Cure. Yeah.
unknown:Bye-bye.
