An explainable machine learning model for predicting chronic coronary disease and identifying valuable text features.

Area of Science:

• Cardiology
• Medical Informatics
• Machine Learning

Background:

• Chronic Coronary Disease (CCD) presents a significant global health burden.
• Current Pre-test Probability (PTP) models often depend on in-hospital data and subjective clinical judgment.
• There is a need for more accessible and objective methods for CCD prediction.

Purpose of the Study:

• To develop machine learning (ML) models for predicting CCD using readily available text data and baseline patient characteristics.
• To assess the impact of text data on the diagnostic performance of ML models for CCD.
• To interpret the ML models using explainability techniques.

Main Methods:

• Text mining was applied to structure patient data, including chief complaints and medical history.
• Customized text processing techniques were developed for cardiovascular medicine.
• Random Forest models were trained and evaluated using patient data, with SHAP used for interpretation.

Main Results:

• A Random Forest model achieved an AUC of 0.93, indicating excellent performance in predicting CCD.
• Key predictive features identified included text data like "chest pain" and "chest tightness", alongside structured data such as age.
• The SHAP algorithm provided insights into how these features influenced model predictions.

Conclusions:

• Text data can be effectively utilized to build accurate prediction models for Chronic Coronary Disease (CCD).
• The SHAP approach aids clinicians in understanding the factors driving CCD predictions.
• This methodology offers a valuable tool for enhancing CCD diagnosis and management.