Explainable machine learning framework based on blood biomarkers and routine health information for assistive...

Acute myocardial infarction (AMI) presents a critical clinical challenge due to its rapid progression and high mortality, compounded by diagnostic delays from limited healthcare access. To address this, we developed a machine learning framework for assistive diagnosis and risk stratification, enabling efficient risk assessment for high-volume outpatients.

In this study, a total of 6100 electronic medical records were collected from Shanxi Cardiovascular Hospital. From these records, 31 clinical and biochemical features associated with AMI were extracted, including cardiac troponin I (cTnI), creatine kinase-MB (CK-MB), myoglobin (Myo), and N-terminal pro-B-type natriuretic peptide (NT-proBNP). Seven machine learning models were developed to perform binary classification for AMI diagnosis. Model interpretability was enhanced using the SHAP algorithm, and a risk stratification strategy was introduced to further classify samples predicted as negative by both the model and SHAP into potential high-risk categories.

Among all evaluated algorithms, the Light Gradient Boosting Machine (LightGBM) model achieved the best diagnostic performance. In the test set, the model reached an AUC of 0.949 and in the independent test set, the AUC further improved to 0.956, confirming its robustness and generalizability. SHAP-based interpretation revealed that cardiac biomarkers, particularly cTnI and CK-MB, contributed most to model predictions. The proposed risk stratification approach effectively reduced the likelihood of missed diagnoses.

The integration of machine learning with SHAP-based interpretation and risk stratification forms a robust and interpretable framework for the assistive diagnosis and risk stratification of acute myocardial infarction.