Don’t be data poor — Anuj Iravane, Anterior

Transcript

Hello everyone, welcome to Don't Be Data Poor. My name is Anuj. I lead AI at Interior. Just a bit about Interior: we are a clinician-led AI company built for health plans, backed by Sequoia and NEA. What we do is run AI transformations for health plans, as part of which we build agents for several high-stakes healthcare administrative workflows in production. Things like prioritization, payment integrity, HEDIS measures, etc. It's okay if you're not familiar with any of these workflows, because a lot of the work that we do can actually be summarized in the same way. It's policy-guided decision-making over highly unstructured data. The unstructured data looks something like this, right? You have these scanned fax bundles containing medical records full of patient information. A not-so-fun fact is that I think around 70% of medical communication still happens via fax. Fortunately or unfortunately, this is the data that we end up working with the most. It is very rich and information-dense data that we see here. The data distribution here comes from a very long tail of rare cases with nuanced scenarios. It models an entire clinical trajectory for a patient, and every single person's journey is very different. It also presents itself in varied formats. You have things like bad handwriting, tables, checkboxes, key-value pairs, images, a lot of tough data to deal with. But I personally think it's a very fascinating source of data that we see here. It's like an observation through a very fuzzy lens over an entire person's lifespan. It's really unique. I'm sure you must have heard this enough times today already, but in healthcare, the baselines for accuracy are exceptionally high. Ninety-five percent is not good enough. At Interior, this is why we invest very deeply in data sets and evals. These unstructured medical records are a stable source of data for these evals, and we work with this kind of data in almost every workflow that we try to automate. But the problem is we can't really keep this data. It's PHI. It's highly protected. We can't retain it, we can't reuse it, and we can't even derive information from it. Most of our contracts prohibit us from doing anything like that. Even things like redacting it, anonymizing it, and keeping derivative copies are a strict no-no, completely off the table. So nothing really survives in any sort of data set that we want to persist over a period of time. So what this talk is about is: what do you do when the data set you most need is also the data you're least allowed to keep? The answer that we put our bets on is that we can synthetically generate this data ourselves. There's been a lot of focus on synthetic data recently. You have Frontier Labs striving to generate synthetic data for continued pre-training, for RL, for computer use, for agents. So it's a hot topic, and it's a hot topic on our minds as well. The moment you say generate, the first thing that comes to mind is, okay, can we try to use an LLM to generate synthetic data? I think you can. I personally believe LLMs are a fantastic tool to generate synthetic data, and several teams have already demonstrated this. There have been some papers in the healthcare space and outside the healthcare space. People have successfully used LLMs to generate synthetic data for different purposes. There are some known challenges in trying to use these LLMs to create data, especially if you're trying to one-shot the whole process. It's really hard to generate diverse, realistic-looking synthetic records, and this is even more of a problem when you're trying to do this at scale. Oftentimes, these medical records are over 300 pages long. It's like imagining that you wouldn't ask an LLM to write a novel for you in one shot, right? It's the same reason why you wouldn't use an LLM to just one-shot a synthetic record for you. LLMs seem to suffer from this very strange mode collapse problem when it comes to generating diverse data, creative data. I think there are two main reasons for it. The first one is, as Aish mentioned in stock earlier, there's very little exposure to this data source in the pre-training data corpus. Today's objectives for pre-training and post-training are largely not incentivized for creativity or diversity, really. They're incentivized to be helpful systems. So with these challenges in mind, I'll walk you through one of our approaches in how we managed to build a pipeline to generate synthetic data. Earlier, I mentioned our forward tasks look something like this, right? You have workflows and tasks that start with some unstructured data and a policy, and you execute your policy against that data. You follow this reasoning trace through it, and you arrive at some sort of an outcome, which is your label. So this is our forward task. The idea we had was to try and reverse this process. Can we actually start by sampling a random label, figuring out a reasoning trace for that label, and then trying to generate data backwards from that? The idea here is that if you can actually sample these two things with enough diversity, we will be able to generate data that's conditioned on a diverse set of inputs, allowing us to circumvent the diversity problem a little bit. So just a quick aside on policies. We've talked about policies a bit, but let me just clarify what these really mean, right? This is an example policy we have for a CPAP device for patients. This particular one is for a medical necessity review workflow, and it outlines all these diverse sets of conditions that a patient might have in which a CPAP device should be approved or rejected. In this policy, as well as many other policies, you can think of these as essentially decision trees that outline all these sorts of conditions that dictate how some outcomes are met. At NTIR, we actually spend a lot of time and energy trying to model these policies explicitly as decision trees. We work with symbolic representations similar to decision trees, and it helps us achieve a better accuracy and consistency score when executing them in an LLM-based workflow. The reason why I'm bringing this up is that by having this sort of symbolic representation of a policy, you actually have a way to deterministically sample different reasoning traces for a given outcome. So back to the idea of reversing the process, right? This sampling of reasoning traces from the policies is what helps us get that diverse conditioning input to then generate medical records from. The key idea here is that the distribution we sample from is a much more uniform and effective prior distribution than what you'd normally get from an LLM. One added benefit of sampling this way is that, in theory, you're able to test for far more scenarios than you would likely get from production data sources. So what I mean by that is, say you get a sample of 200 cases from your customer, and you try to have an eval that measures performance against that, and you get a 95% score. It doesn't really tell you what your performance would be in those rare edge cases that are not in that data set. There will always be rare edge cases that are outside that distribution just because of the fact that our data is so highly variant. So for those familiar with Cynthia, they follow a similar pattern of sampling scenarios from a symbolic causal state representation. There are a few folks in the space who are working with these symbolic representations to generate diversity in synthetic data generation. So let me walk you through the rest of the pipeline. Once we have this diverse set of samples as a conditioning input, what we did was build an LLM-based pipeline that follows a coarse-to-fine pattern to progressively build up a medical record layer by layer. Here we first start with creating some patient invariants like biological sex, birth date, blood group. We use that, along with the reasoning trace, with an LLM again to produce an ordered list of events and provider encounters that a patient might have had, and we call this the patient journey. This is a high-level overview of what a patient might have gone through in their lifespan, captured by a list of events in natural language. In the real world, it is actually only during these provider encounters that documentation is really generated, at least for the data that we get. Most of our source data is generated during these provider encounters. So we model exactly that in our pipeline. We first generate a document plan for each encounter, and then based on that and the preceding history of the patient, we fan out into generating the actual documents to hydrate them with actual synthetic information. This coarse-to-fine layering is what allows us to keep the different prompt payloads in the pipeline very token-efficient from both input and output perspectives. This also helps enable scaling across longer patient journeys. You can scale this pipeline, have a much longer patient journey, and just fan out and generate documents that way without overloading the context windows of your LLMs. Finally, we have this sort of refinement loop at the end that uses a set of evals to provide feedback to improve specific parts of the generated documents. For example, one of the evals we have is an LLM-based check for consistency between all documents. This makes sure that there are no contradictions, inaccuracies, or conflicting information between two generated documents. This is important because we have a parallel fan-out process that is used to generate these documents independently. Because we started with the labels for this particular pipeline run, what we also have is an ability to use those labels and compare those against the generated medical records to see if the tasks that we originally used actually match and the data is in concordance with the task inputs and outputs. So we can do this sort of round-trip check to ensure that our data is actually in sync and, by default, get correct labels by construction. In theory, this is a really nice property to have. You can basically skip the expensive ground-truthing process you need for your data. One thing to clarify here is that so far, all the generation has been happening just in plain text and markdown text. It is possible to go from that to a rendered PDF, but we don't really see much value in doing that because we have state-of-the-art PDF parsers today. They are available to everyone, and they allow you to convert any sort of complex PDF into a nice markdown representation. So all of this synthetic generation and evaluation happens in the text domain. This is just an example of a pipeline that we created from scratch, and it's very easy to build. It's largely fully LLM-based. But who came up with this, right? Who am I to know anything about what a good medical record looks like? So how do we know if this is any good? I think this has been mentioned a few times today already, but you really don't. No AI engineer ever would. You want your domain experts to be the ones telling you what's good and what's not good, which is why we believe that it is of great value to empower your domain experts to own your whole data pipeline. Specifically, we do this in two ways, right? We enable our clinicians to interject at each point in the generation process with a human-in-the-loop mechanism. At any point, a clinician can steer the generation process to make a medical record the way they want it. We often see our clinicians use this to first look at cases that happen in production, get some interesting ideas, and then use those ideas along with the steering in this pipeline to make cases that look similar to what we might see in production or what they have seen in production. This is what makes the data generated from this really useful, right? You can actually model your failure cases beforehand or even after you see them in production. Secondly, and I think most importantly, we let our clinicians also own the whole logic of the pipeline. We do this by modeling the whole pipeline as a skills-based workflow running on a genetic agent harness that we built internally. Every section you see here, all the way from the patient journey to the document generation to the document enrichment to the evals, all of these things are skills that run on our agent harness. As an example, if a clinician wanted to add support for a new document type, let's say for a new customer, and they wanted their intake forms to look a certain way, they could easily just make a new skill file for it, attach it to the pipeline, and voila, there wouldn't be any engineering changes required. So it's completely clinician-owned from that perspective. Just as an aside generally, I feel like skills are really an amazing interface between AI engineers and domain experts, especially in vertical AI. We see this being modeled in several of our other workflows, both for internal use cases and in production as well. So some results from this. Even though we only really use synthetic data for evaluation at the moment, there are already a lot of merits that we get from it. Roughly 90% of our data sets are already made of synthetic data. This helps us maintain a very high production accuracy score across many customer deployments. The pipelines that I just showed you are already able to achieve very high fidelity on this generated data. In a blind review, clinicians were only able to distinguish synthetic from real about 60% of the time. So there is room for improvement, but it's close, and it's quite a promising avenue for us to invest more here. The fact that is most interesting to me, and what I'm really excited about, is that most of our data sets today are created just in time for these customer deployments, right? When you have the ability to create data from scratch so quickly, you don't need to depend on waiting for data from your customer. You can just model all your edge cases, simulate them, and test your workflows before you go live in production. So some takeaways. If you're looking to build your own synthetic data pipeline in healthcare or even another domain, try reversing your inference workflow. Diversity should always be sampled from an appropriate distribution for your use case. Try to emulate the process in which the data was actually generated. Like I showed you, we were using LLMs while trying to emulate how our medical records might actually be generated during patient encounters. I would highly recommend you try doing that. The fourth, most important thing, I think, is when you're making a data pipeline like this, it's really important to give your domain experts the keys, because these are the people who know about your data, and they will help you drive toward recursive self-improvement, not the AI engineers. Cool. So you don't need a PHI problem for this. Anywhere the data you need is ephemeral, sensitive, or even expensive to label, you can think about generating data yourself, and hopefully you won't be data poor. Thank you, everyone. Thank you. Thank you. Thank you. Thank you. Thank you. You follow this reasoning trace through it. And you arrive at some sort of an outcome, which is your label. So this is our forward task. And the idea we had was to try and reverse this process. Can we actually start by sampling a random label, figuring out a reasoning trace for that label, and then trying to generate data backwards from that? The idea here being that if you can actually sample these two things with enough diversity, we will be able to generate data that's conditioned on diverse set of inputs, allowing us to kind of circumvent the diversity problem a little bit. So just a quick aside on policies. We've talked about policies a bit, but let me just clarify what these really mean, right? So this is an example policy we have for a CPAP device for patients. This particular one is for a medical necessity review workflow. And it sort of outlines all these diverse set of conditions that a patient might have in which a CPAP device should be approved or rejected. So in this policy, as well as many other policies, you can think of these as essentially decision trees that outline all these sorts of conditions that dictate how some outcomes are met. And at NTIR, actually, we spend a lot of time and energy in trying to model these policies explicitly as decision trees. We work with symbolic representation similar to decision trees, and it helps us achieve a better accuracy and consistency score when executing them in LLM-based workflow. And the reason why I'm bringing this up is that by having this sort of symbolic representation of a policy, you actually have a way to kind of deterministically sample different reasoning traces for a given outcome. So back to the idea of, like, reversing the process, right? This sampling of reasoning traces from the policies is what helps us get that diverse conditioning input to then generate medical records from. And the key idea here is that the distribution here that we sample from is a much more uniform and effective prior distribution than what you'd normally get from an LLM. One added benefit of sampling this way is that, in theory, you're able to test for far more scenarios than you would likely get from production data sources. So what I mean by that is, like, say you get a sample of 200 cases from your customer, and you try to, like, have an eval that measures performance against that, and you get a 95% score. It doesn't really tell you about what your performance would be in those rare edge cases that are not in that data set. There will always be rare edge cases that are outside that distribution just because of the fact that our data is so highly variant. So for those family with Cynthia, like, they follow a similar pattern of sampling scenarios from a symbolic causal state representation. There's a few of the folks in the space who are working with these symbolic representations to generate diversity in synthetic data generation. So let me walk you through the rest of the pipeline. All right? So once we have this diverse set of samples as a conditioning input, what we did was we built an LLM-based pipeline that follows a course-to-find pattern to progressively build up a medical record layer by layer. So here we first start with creating some patient invariants like the biological sex, the birth date, the blood group. We use that, along with the reasoning trace, with an LLM again, to produce an ordered list of events and provider encounters that a patient might have had, and we call this the patient journey. So this is a high-level, you can think of it as a high-level overview of what a patient might have gone through in their lifespan, captured by a list of events in natural language. And in the real world, it is actually only during these encounters, provider encounters, that documentation is really generated, at least for the data that we get. Most of our data source data is generated during these provider encounters. So we model exactly that in our pipeline. We first generated a document plan for each encounter, and then based on that and the preceding history of the patient, we fan out into generating the actual documents to hydrate them with actual synthetic information. And this course-to-find layering is actually what allows us to keep the different prompt payloads in the pipeline very token-efficient from both input and output perspective. This also helps us enable to scale across longer patient journeys. So you can scale this pipeline, you can have a much longer patient journey, and you can just fan out and generate documents that way without overloading the context windows of your LLMs. Finally, we have this sort of refinement loop in the end that uses a set of evals to provide feedback to improve specific parts of the generated documents. For example, one of the evals we have is an LLM-based check for consistency between all documents. So this makes sure that there's no contradictions or inaccuracies or conflicting information between two documents that are generated. And this is important because we have a parallel fan-out process that is used to generate these documents independently. And because we started with the labels for this particular pipeline run, what we actually also have is an ability to kind of use those labels and compare those against the generated medical records to see if the tasks that we originally used actually matches and the data is in concordance with the task inputs and outputs. So we can do this sort of round-trip check to ensure that our data is actually in sync and by default get correct labels by construction. So in theory, this is a really nice property to have. You can basically skip the expensive ground-truthing process you need for your data. One thing to clarify here is that so far, all the generation has been happening just in plain text and markdown text. It is possible to go from that to a rendered PDF, but we don't really see much value in doing that because we have state-of-the-art PDF parses today. They are available to everyone, and they just allow you to convert any sort of complex PDF into a nice smart-down representation. So all of this synthetic generation and evaluation happens in the text domain. So this is just an example of a pipeline that we created from scratch, and it's very easy to build. It's largely fully LLM-based. But who came up with this, right? Like, who am I to know anything about what a good medical record looks like? So how do we know if this is any good? And I think this has been mentioned a few times today already, but, like, you really don't. Like, no AI engineer would ever would. Like, you want your domain experts to be the ones telling you what's good, what's not good, which is why we believe that it is of great value to empower your domain experts to own your whole data pipeline. And specifically, we do this in two ways, right? We enable our clinicians to kind of interject at each point in the generation process with a human-in-the-loop mechanism. So at any point, a clinician can steer the generation process to make a medical record in the way they want it. We often see our clinicians use this to first look at cases that happen in production, get some interesting ideas, and then use those ideas along with the steering in this pipeline to make cases that look similar to what we might see in production or they have seen in production. And this is what makes the data generated from this really useful, right? Like, you can actually model your failure cases beforehand or even after you see them in production. And secondly, I think most importantly, we let our clinicians also own the whole logic of the pipeline. We do this by modeling the whole pipeline as a skills-based workflow running on a genetic agent harness that we build internally. So every kind of section here you see, all the way from the patient journey to the document generation to the document enrichment to the evals, all of these things are skills that run on our agent harness. As an example, if a clinician wanted to, say, maybe add support for a new document type, let's say for a new customer, they wanted their intake forms to look a certain way, they could easily just make a new skill file for it, attach it to the pipeline, and voila, there wouldn't be any engineering changes required. So it's completely clinician-owned from that perspective. And just an aside generally, I feel like skills are really an amazing interface between AI engineers and domain experts, especially in vertical AI. We see this being modeled in several of our other workflows, both for internal use cases and in production as well. So some results from this, right? So even though we only really use synthetic data for evaluation at the moment, there's already a lot of merits that we get from it. Roughly 90% of our data sets are already made of synthetic data. This helps us maintain a very high production accuracy score across many customer deployments. The pipelines that I just showed you, already we're able to achieve a very high fidelity on this generated data. In a blind review, clinicians were only able to distinguish synthetic from real about 60% of the time. So room for improvement, but it's close. And it's a quite promising avenue for us to invest more here. And the fact that is the most interesting to me and what I'm really excited about is that all of these data sets, well, most of our data sets today then are created just in time for these customer deployments, right? When you have the ability to create data from scratch so quickly, you don't need to depend on waiting for data from your customer. You can kind of just model all your edge cases, simulate them, and test your workflows before you go live in production. So some takeaways, if you're looking to build your own synthetic data pipeline in healthcare or even another domain, try reversing your inference workflow. Diversity should always be sampled from an appropriate distribution for your use case. Try to emulate the process in which the data was actually generated. So like I showed you, we were trying to sort of like, we were using LLMs where we were trying to emulate how our medical records might actually be generated during patient encounters. So I would highly recommend you try doing that. And the fourth most important thing, I think, is when you're making a data pipeline like this, it's really important to give your domain experts the keys because these are the people who know about your data and they will help you drive towards a recursive self-improvement and not the AI engineers. Cool. So you don't need a PHI problem for this. Anywhere the data you need is ephemeral, sensitive, or even expensive to label, you can think about generating data yourself and hopefully you won't be data poor. Thank you, everyone. Thank you. Thank you. Thank you. Thank you. Thank you.