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

Area of Science:

• Biomedical Engineering
• Machine Learning in Healthcare
• Clinical Informatics

Background:

• Electrocardiogram (ECG) signals are crucial for diagnosing serum potassium abnormalities.
• Restrictions on raw medical data sharing in Taiwan limit its reuse and cross-institutional applications.
• Developing methods to extract clinical information without raw data is essential.

Purpose of the Study:

• To transform classification information from raw ECG signals into high-level SHAP features.
• To evaluate the feasibility of SHAP features as auxiliary or alternative inputs for serum potassium abnormality prediction.
• To propose a framework integrating ECG submodels, SHAP features, and ensemble learning.

Main Methods:

• A time-series classification framework was developed, integrating ECG wave-segment submodels.
• SHAP (SHapley Additive exPlanations) was used to generate high-level features from submodel contributions.
• Stacking ensemble learning was employed for serum potassium abnormality prediction, with PCA features used for comparison.

Main Results:

• Incorporating SHAP-based augmented features generally improved ECG-based serum potassium abnormality prediction.
• Models trained solely on SHAP features achieved performance comparable to or exceeding baseline models.
• SHAP features retained discriminative information, demonstrating potential as alternative features to raw ECG data.

Conclusions:

• SHAP-based features offer a viable method for extracting predictive information from ECG signals without direct use of raw data.
• This approach supports medical data reuse, cross-institutional collaboration, and privacy risk assessment.
• The findings provide a foundation for future applications in remote diagnostics and federated learning.