Large language models for predicting one-year major adverse cardiovascular events in acute coronary syndrome.

Effective risk stratification is crucial for managing acute coronary syndrome (ACS). This study evaluated whether general-purpose large language models (LLMs) can reliably execute the complex clinical reasoning required for cardiovascular prognosis. We quantitatively assessed three LLMs-ChatGPT 4o, DeepSeek R1, and Grok 3-for predicting one-year major adverse cardiovascular events (MACEs), using 29 guideline-recommended features from 903 participants in the LM-ACS cohort and 64 participants in the MIMIC database. All models demonstrated significant risk overestimation and substantial output variability across ten independent runs. Although Grok 3 showed the highest initial accuracy, its average performance over all runs showed no significant difference. ROC analysis indicated that none of the LLMs outperformed traditional clinical scores, with ChatGPT 4o performing below both GRACE and TIMI score. These findings demonstrate that current LLMs are not yet suitable for standalone ACS prognosis prediction; specifically, their associative reasoning poses a fundamental challenge to reliable clinical translation.