Explainable Machine Learning for Identifying False-Negative Myocardial Infarction: A SHAP-Based Analysis

Background

The timely and accurate identification of Myocardial Infarction (MI) remains a challenge in clinical practice. Diagnostic inconsistency may result in missed or delayed diagnoses, especially in resource- limited settings. This study aimed to develop an explainable Machine Learning (ML) based framework to detect MI cases that may be overlooked by conventional biomarker-based diagnosis, particularly in patients with normal troponin and Creatine Kinase-MB (CK-MB) levels, by using SHapley Additive exPlanations (SHAP).

Materials and Methods

A retrospective dataset comprising 1,319 patient records with nine clinical variables was utilized. Multiple ML classifiers, including Random Forest, XGBoost, and Decision Tree, were trained following pre-processing steps: normalization, feature encoding, hyperparameter tuning. Model performance was assessed using accuracy, precision, recall, F1-score, and Area Under the Curve (AUC). SHAP analysis was employed to elucidate feature contributions and support model predictions.

Results

Tree-based classifiers achieved high predictive performance, with accuracies exceeding 97% and AUC values up to 0.98. SHAP analysis identified troponin, CK-MB, and age as the most influential predictors of myocardial infarction. Importantly, the model correctly identified MI in patients with normal troponin and CK-MB levels, by integrating biomarker patterns with demographic and hemodynamic features. These findings indicate that the explainable ML framework can detect MI cases at risk of false-negative classification using conventional biomarker thresholds.

Conclusion

The integration of ML with explainable Artificial Intelligence (AI) techniques shows strong potential for early and accurate MI prediction, including the identification of cases that may be falsely classified as negative using conventional diagnostic approaches. This approach supports improved clinical decision making, enhances diagnostic safety by reducing false-negative outcomes.