Causes of CoMET’s underperformance on hyperlipidemia-specific outcome tasks

Identify the factors that lead CoMET models to underperform task-specific supervised models on hyperlipidemia-specific outcome prediction tasks, including one- and three-year risks of atherosclerotic cardiovascular disease events, heart attack, stroke, and chronic heart failure among patients receiving treatment for hyperlipidemia.

Background

The paper evaluates CoMET on multiple disease-specific outcome tasks. While CoMET generally improves with scale and often matches or exceeds supervised baselines, it underperforms on hyperlipidemia-related outcomes.

The authors explicitly note that the reasons for this underperformance are not yet understood and defer investigation to future work, making it an open problem that impacts performance optimization and clinical utility in cardiovascular risk prediction contexts.