Enhancing clinically cardiovascular machine learning model for risk prediction via sample augmentation

Background Small sample dataset and heterogeneous distributions limit the robustness and implementability of machine learning models for structured clinical data. To evaluate the value of moderate data augmentation for cardiovascular risk modeling and propose an interpretable and deployable solution within a “continuous risk to thresholding” framework. Methods The heart disease classification dataset was randomly divided into training and validation sets with an 8:2 ratio. Constrained feature space augmentation was performed within the training set, and the effects of four thresholds (0×, 1×, 2×, and 3×) on support vector regression (SVR), random forest (RF), extreme gradient boosting (XGBoost), light Gradient Boosting Machine (LightGBM), and multi-layer perceptron (MLP) were compared. Continuous risk scores were evaluated using mean absolute error (MAE), root mean squared error (RMSE), and R 2 . The output confusion matrix was then thresholded. Interpretative analysis was performed using SHapley Additive exPlanations (SHAP), and partial dependence plot (PDP). Results 2 × augmentation achieved the favorable compromise between error (reduced MAE and RMSE) and goodness of fit (increased R 2 ). Under the 2 × condition, RF performed favorable overall, achieving an accuracy of 94.0%, an F2 of 94.4%, a sensitivity of 95.9%, and a specificity of 91.8% after thresholding. SHAP/PDP shows that oldpeak, num major vessels, chest pain type, thal, exang, and max hr. are the primary driving factors, exhibiting stable patterns such as monotonic, step, and threshold. Moderate data augmentation (preferably 2×) can significantly improve robustness in small sample settings; RF achieves the favorable balance between accuracy, stability, and interpretability. Conclusion Combining SHAP/PDP’s multi-layered interpretation and thresholding approach, this study provides a reusable multiplication guidance and risk stratification scheme, providing a methodological basis for deploying interpretable cardiovascular risk models.