Anterior's Anuj Iravane on Synthetic Healthcare Data

In the complex world of healthcare AI, data scarcity and privacy concerns can be significant hurdles. Anuj Iravane, Head of AI at Anterior, a clinician-led AI company focused on health plans, shared insights into how his company addresses these challenges by generating synthetic data. In his presentation titled "Don't Be Data Poor," Iravane detailed Anterior's approach to tackling high-stakes administrative workflows with AI, particularly when dealing with highly unstructured data like scanned fax bundles, where approximately 70% of medical communication still occurs.

The Challenge of Unstructured Healthcare Data

Iravane highlighted the inherent difficulties in working with medical records, which are often dense with information, present in varied formats, and contain long-tail, rare cases. He noted that while this data is rich, it's also highly protected health information (PHI), making it impossible to retain, reuse, or even derive information from due to strict contracts and regulations. This presents a paradox: the data most needed for AI development is also the least permissible to keep.

Synthetic Data Generation as a Solution

Anterior's solution lies in synthetically generating the data they need. Iravane acknowledged the growing trend and interest in synthetic data for AI training and evaluation. While Large Language Models (LLMs) are powerful tools for this purpose, they often struggle with generating diverse and realistic-looking records, especially when dealing with lengthy documents like medical records. LLMs can suffer from "mode collapse," producing data that leans towards average cases rather than the rare scenarios crucial for robust AI.

Reversing the Inference Workflow

To overcome these limitations, Anterior employs a strategy of reversing the traditional AI inference process. Instead of starting with data to derive a label, they begin by sampling a chosen outcome (label) and then work backward to generate a corresponding reasoning trace and, finally, the synthetic data itself. This approach, they believe, allows for greater diversity and better coverage of edge cases.

Leveraging Clinical Policies

A core element of Anterior's method involves modeling clinical policies as explicit decision trees. These policies, which outline conditions for approving or rejecting medical necessity reviews, for example, provide a structured way to sample diverse reasoning traces. By deterministically sampling these traces from the policy's logic, Anterior ensures that the generated data is conditioned on a wider range of scenarios than might be found in real-world datasets.

The Synthetic Data Pipeline

Iravane outlined their LLM-based pipeline, which builds medical records layer by layer. It starts with patient invariants, generates a patient journey of events and encounters, and then creates document plans for each encounter. This coarse-to-fine layering keeps prompt payloads token-efficient and allows for scaling to longer patient histories without overwhelming LLM context windows. A refinement loop with LLM-based evaluations, such as consistency checks between documents, further enhances the quality of the generated data.

Empowering Domain Experts

Crucially, Anterior emphasizes empowering their domain experts, particularly clinicians, to own the data pipeline. This is achieved through a "human-in-the-loop" mechanism, allowing clinicians to steer the generation process and provide feedback. By allowing clinicians to define and manage the pipeline's logic as skills within a generic agent harness, Anterior ensures that the synthetic data accurately reflects real-world medical scenarios and can be adapted quickly for new customer deployments.

Promising Results and Key Takeaways

The results have been promising, with approximately 90% of Anterior's datasets now being synthetic. Clinicians in blind reviews could only distinguish synthetic from real data about 60% of the time, indicating high fidelity. Furthermore, over 50% of their datasets are created just-in-time for specific customer deployments, allowing for faster and more tailored AI solutions.

Iravane concluded with key takeaways for those looking to build similar synthetic data pipelines: reverse your inference workflow, sample diversity from appropriate distributions, emulate the original data generation process, and most importantly, empower domain experts by giving them control over the data pipeline.