A study published in Cardiology Research and Practice evaluated an explainable deep learning (DL) system for mortality prediction in acute coronary syndrome (ACS). A Clinical-BERT framework enhanced with chain-of-thought (CoT) reasoning was developed to improve both prediction accuracy and interpretability.

Electronic health records (EHRs) from ACS patients were analyzed, and the Clinical-BERT + CoT system was compared with the Global Registry of Acute Coronary Events (GRACE) score.

The AI model achieved an AUROC of 0.97, significantly higher than the GRACE model (AUROC 0.83). The CoT approach improved interpretability by identifying clinical features such as age, Killip class, blood pressure, and biomarkers that contributed to predictions.

The study concluded that a CoT-based Clinical-BERT system can enhance ACS mortality prediction while providing transparent insights into model outputs. Broader validation is required, but this approach shows promise for integration into decision support tools to enable more accurate and interpretable risk assessments.