AI Meets Human Expertise: A Hybrid Approach to Cardiovascular Event Adjudication

Meet ADAPT-CEC: an adaptive AI algorithm that nails cardiovascular event classification with the same accuracy as expert humans—and a hybrid AI-human approach could slash the time and cost of this labor-intensive process.

Background

Cardiovascular event classification in clinical trials has long demanded painstaking work from expert committees. ADAPT-CEC changes that. This algorithm intelligently categorizes myocardial infarction, stroke, heart failure, and other cardiac endpoints across multiple trials, even when endpoint definitions vary from study to study.

Key Findings

The results are striking. ADAPT-CEC correctly identified 86.4% of events and 99.4% of non-events—matching expert adjudication exactly. But here’s where it gets interesting: when researchers added human review of the 30% of cases the AI felt least confident about, performance jumped. F1 scores climbed to 0.80–0.94 and event classification accuracy hit 95.6%. For comparison, direct GPT 4.0 adjudication fell short, especially on bleeding endpoints (F1: 0.78 vs. 0.56). What really matters? All methods—human, AI, hybrid, or GPT—produced the same primary trial endpoint results, preserving the integrity of the studies.

Why It Matters

Endpoint adjudication drains time and resources in clinical trials. If this hybrid approach lives up to its promise, it could free up those resources while keeping accuracy intact—potentially opening the door to faster cardiovascular research.

Limitations

One catch: this work relied on retrospective data with simulated hybrid adjudication. Real-world prospective studies are needed to figure out how best to actually use this in practice before rolling it out widely.

Original paper: Adaptive AI for Cardiovascular Event Adjudication: Cardiovascular Event Adjudication Across Different Definitions in the ODYSSEY OUTCOMES and EUCLID Trials. — Circulation. 10.1161/CIRCULATIONAHA.126.080072

🎧 Listen to the podcast