CoT-Based Clinical-BERT for Risk Stratification After Acute Coronary Syndrome

Abstract

Background: Acute coronary syndrome (ACS) represents the most severe manifestation of coronary heart disease (CHD), which is a primary cause of mortality globally. Developing a novel risk stratification system for ACS, grounded in clinical data from electronic health records (EHRs), holds significant promise for mitigating the burden of ACS. Traditional models for ACS risk assessment often struggle with handling nonlinear relationships and lack the precision required for accurate predictions. Deep learning (DL) has the potential to improve prediction accuracy; however, its “black box” nature limits clinical applicability due to insufficient interpretability. Methods: We introduce an innovative, explainable DL system that utilizes a chain-of-thought (CoT) approach in conjunction with Clinical-BERT, a model pretrained on clinical data, to predict and interpret ACS mortality risk. The system first extracts ACS-related risk factors from EHR using Clinical-BERT and then applies CoT reasoning to enhance the interpretability of these predictions. Results: The experimental findings reveal that the proposed method significantly surpasses the widely used Global Registry of Acute Coronary Events (GRACE) model in predicting ACS-related mortality. The Clinical-BERT model achieved an area under the receiver operating characteristic curve (AUROC) of 0.97, notably higher than the GRACE model’s AUROC of 0.83. Moreover, the CoT mechanism greatly improved the interpretability of the model’s predictions, providing clear insights into the underlying risk factors. Conclusions: The CoT-based Clinical-BERT model for ACS risk stratification not only improves the accuracy of ACS risk prediction but also enhances the interpretability of DL-based medical tools. This approach presents a promising avenue for incorporating advanced AI techniques into clinical practice, enabling more precise and transparent risk assessments for ACS patients.