Automated computation of the HEART score with the GPT-4 large language model

Abstract

Background

Automated computation of the HEART score has the potential to facilitate clinical decision support and safety interventions. The goal of this study was to assess the performance of the GPT-4 large language model (LLM) in computation of the HEART score and prediction of 60-day major adverse cardiac events (MACE).

Methods

In this retrospective cohort study from February 2022 to September 2023, patients admitted to a chest pain observation unit were identified. HEART scores were calculated by a physician assistant or nurse practitioner (APP) as part of routine care. Separately, the LLM calculated a HEART score utilizing an iteratively developed prompt from deidentified chart documentation. Any cases of disagreement with the APP score were adjudicated by an emergency physician blinded to clinical outcomes. Agreement on HEART score was assessed, and 60-day MACE was obtained via linkage to an institutional registry.

Results

Of the 601 participants, 50 were utilized for prompt development. Among the remaining 551 participants, agreement by Cohen's weighted kappa between the LLM and adjudicators was 0.67 which was similar to the agreement of 0.66 between the APP and adjudicators. The LLM predicted a higher average HEART score (mean 5.06) compared to the adjudicators (mean 4.69) or APP (mean 4.23). No significant difference was seen in diagnostic performance for 60-day MACE by DeLong pairwise comparison (all p > .05).

Conclusions

Automated risk score computation with language models has the potential to power interventions such as clinical decision support but has systematic differences from physician judgment. Prospective investigation is needed.