Abstract Body (Do not enter title and authors here):
Background: Medical record review by a physician clinical events committee (CEC) is the gold standard for adjudicating heart failure hospitalizations (HFH) in clinical trials but is labor-intensive. Automated adjudication by natural language processing (NLP) could improve efficiency, but has not been validated in global trials.

Methods: We developed a model for automated NLP-based HFH adjudication (“HF-NLP”) that integrates two complementary approaches: a Clinical Longformer model trained in the INVESTED, PARADISE-MI, and PARAGON-HF trials, and a GPT4o prompt that identifies HF criteria (signs, symptoms, and treatment) and quotes supporting sentences. The HF-NLP model generates a binary HFH adjudication and an ordinal five-category score indicating increasing probability of HFH. In the DELIVER randomized trial, which compared dapagliflozin to placebo in 6063 patients with heart failure with mildly reduced or preserved ejection fraction, we evaluated agreement between HF-NLP and human CEC adjudications and compared the estimated effect of dapagliflozin on HFH using HF-NLP versus CEC adjudication.

Results: The DELIVER CEC adjudicated HFH in 1233/1597 (77%) of investigator-reported events. The HF-NLP adjudication agreed with the CEC in 83% of events (Figure 1). The proportion of events adjudicated as HFH by the CEC increased across the HF-NLP ordinal categories: 17% in confident non-HFH, 32% in probable non-HFH, 66% in uncertain, 85% in probable HFH, and 96% in confident HFH (Figure 2). A strategy of human review for events that HF-NLP deemed uncertain would have achieved 91% agreement with the CEC while reducing adjudication workload by 84%. The effect of dapagliflozin on HFH was similar with HF-NLP adjudication (hazard ratio [HR] 0.76 [95% CI 0.66-0.88]) compared to CEC adjudication (HR 0.77 [95% CI 0.67-0.89]) (Figure 3).

Conclusion: Automated adjudication of HFH by the HF-NLP model demonstrated substantial agreement with the physician CEC in the DELIVER trial and yielded a similar estimate of dapagliflozin efficacy. NLP adjudication could improve the efficiency of global clinical trials while preserving accuracy and interpretability.