ADAPT-CEC is an AI-based clinical endpoint adjudication (CEA) tool designed to classify multiple cardiovascular events and adapt to new endpoint definitions without rederivation from scratch. Derived on MI, stroke, and heart failure from ODYSSEY OUTCOMES and externally validated in EUCLID for MI, stroke, bleeding, and CV death after adapting with a limited set of EUCLID suspected events, it was evaluated against human adjudication and alternative AI methods. In EUCLID, among 13,885 suspected primary endpoints, ADAPT-CEC correctly classified 86.4% of all endpoints and 99.4% of non-endpoints; a hybrid approach—with the bottom 30% of AI-prediction certainty subjected to human adjudication—achieved 95.6% overall accuracy and highest F1 metrics across components (CV death, MI, stroke, bleeding). By contrast, GPT-4.0 alone classified 76.3% of endpoints with 99.8% non-endpoint accuracy, and the AI-hybrid approach outperformed both AI-alone methods for several measures. The EUCLID treatment effect was similar across adjudication strategies (HRs near 1.0). The authors conclude that brief adaptation enables a single AI model to reproduce primary outcomes in a new trial and that a human-in-the-loop hybrid approach enhances performance.
Circulation, Ahead of Print. Background: Clinical endpoint classification (CEC) is the gold standard for cardiovascular endpoint measurement in clinical trials, but adds time and cost. We developed and validated an artificial intelligence (AI) algorithm (ADAPT-CEC) that adjudicates multiple cardiovascular endpoints and adapts to new definitions.Methods: ADAPT-CEC was derived on myocardial infarction (MI), stroke, and heart failure from the ODYSSEY OUTCOMES trial and externally validated on MI, stroke, bleeding and CV death from the EUCLID trial after adaptation with 20 EUCLID suspected events per endpoint. ADAPT-CEC was compared via F1 score with direct generative pretrained transformer (GPT) 4.0 adjudication and a hybrid approach where the 30% of suspected events with the lowest AI prediction certainty used human adjudication. The EUCLID primary endpoint of CV death, MI, or stroke was re-estimated for all three adjudication strategies.Results: Amongst 13,885 suspected EUCLID primary endpoint events, ADAPT CEC, hybrid, and GPT 4.0 strategies correctly classified 86.4%, 95.6%, and 76.3% of all endpoints and 99.4%, 99.6%, and 99.8% of all non-endpoints compared with human adjudication, respectively. Hybrid adjudication F1 metrics were the highest [CV death (0.94, 95% CI 0.92 – 0.96), MI (0.80, 95% CI 0.77 – 0.82), stroke (0.82; 95% CI 0.78 – 0.86), bleeding (0.83, 95% CI 0.82 – 0.85)]. ADAPT-CEC F1 metrics were lower for CV death, MI, and stroke but similar to GPT 4.0 while bleeding (0.78, 95% CI 0.77 – 0.79) was superior to GPT 4.0. The EUCLID primary treatment effect was similar by human adjudication (HR 1.02, 95% CI 0.93 – 1.13), hybrid (HR 1.04; 95% CI 0.94 – 1.15) ADAPT-CEC (HR 0.98, 95% CI 0.88 – 1.09) and GPT 4.0 (1.06, 95% CI 0.95 – 1.19) adjudication.Conclusions: After brief adaptation, a single trial derived AI algorithm can adjudicate similar (MI and stroke) and new endpoints (CV death and bleeding) in a second trial and replicate the EUCLID primary outcome treatment effect. A hybrid approach with humans adjudicating those suspected events with the lowest 30% of ADAPT-CEC prediction certainty was superior to ADAPT-CEC alone or GPT 4.0 alone and replicated the EUCLID primary outcome treatment effect. Prospective studies of adaptive AI adjudication are needed to determine future trial implementation.