SHAP-Based Feature Augmentation and Stacking Ensemble Learning for ECG-Based Serum Potassium Abnormality Prediction

Electrocardiogram (ECG) signals contain important clinical information associated with serum potassium abnormalities. However, in Taiwan, raw patient data and original medical signals generally cannot be taken outside the hospital environment, thereby limiting their subsequent reuse and cross-institutional applications. The objective of this study is to transform classification-related information contained in raw ECG signals into high-level SHAP features and to evaluate their feasibility as auxiliary or alternative features to the original wave-segment model outputs. To this end, this study proposes a time-series classification framework that integrates ECG wave-segment submodels, SHAP-based feature augmentation, and stacking ensemble learning for serum potassium abnormality prediction. Submodels are first trained separately using different ECG wave segments and their combinations. SHAP is then applied to transform the contributions of the wave-segment submodel outputs to the prediction results into high-level features. In addition, PCA features are included as a comparison baseline to analyze the effects of different feature transformation methods on classification performance. The experimental results show that incorporating SHAP-based augmented features improves ECG-based serum potassium abnormality prediction performance under most settings. Even when only SHAP-based augmented features are used for training, some models still maintain performance comparable to or better than the Baseline. Although PCA provides more stable classification balance for some patients, SHAP-based augmented features can still represent classification-related model contribution information under most settings while achieving better or comparable classification performance. Overall, even without directly using the original ECG signals in the final classification stage, these features retain a certain degree of discriminative information and demonstrate the potential to serve as alternative features to the original wave-segment model outputs. Therefore, the findings of this study provide a preliminary reference for future medical data reuse, cross-institutional collaboration, and privacy risk assessment.