Skip to content

Clinical Phenotype Definitions for Foundation Models

Home / Open Questions / Clinical Phenotype Definitions for Foundation Models

Background: Healthcare foundation models, particularly those trained on administrative claims, currently rely on task-specific ICD-10 groupings to support benchmarking and comparability. These groupings may not always align with biologically or clinically meaningful disease phenotypes.

Question / Future Work: There is a need to explore alternative, more clinically nuanced disease definitions—such as phenotype-driven systems like PheCodes or DRG-based families—to improve the interpretability and downstream clinical utility of healthcare foundation model representations.

Why It Matters: Using suboptimal disease definitions limits the clinical validity and precision of disease-onset predictions, potentially leading to representations that capture coding patterns rather than true pathophysiology.

Evidence: Finally, disease endpoints were defined using ICD-10 groupings to support broad-scale benchmarking and comparability with Delphi, but these groupings may not always align with clinically coherent phenotypes. Alternative definitions, including DRG-based families or phenotype-driven systems such as PheCodes, may improve interpretability and downstream clinical utility.

Metadata & Links