A SHAP-Explainable Framework for Blood Pressure Prediction Based on PPG and ECG Signal Analysis

Continuous blood pressure (BP) monitoring is crucial in managing hypertension, but current techniques are risky and uncomfortable. Extensive research has been conducted to explore the prediction of BP using features extracted from the photoplethysmogram (PPG) and the electrocardiogram (ECG). However, most of these generated features are not significantly linked with blood pressure and frequently lack a rigorous scientific explanation. This paper identifies a set of features clinically relevant to the BP prediction for predicting systolic and diastolic BP using machine learning; the process is further enhanced with Shapley Additive Explanations (SHAP) for optimal feature selection. The feature set extracted from PPG and ECG signals, and patients’ demographic data are used as the input for Support Vector Regression (SVR) and Random Forest algorithms for BP prediction. In this research, it was found that Random Forest is superior to SVR. The findings from the experiment combining ECG and demographic features revealed a decrease in mean error rate by 24.13% and 81.50%, respectively, for systolic and diastolic BP prediction compared to the analysis using only PPG. SHAP-based feature selection reduced the feature set by 50% while achieving the necessary predictive capabilities. These results offer essential insights for developing a valid and noninvasive method for BP monitoring, assisting medical teams conducting clinical research on hypertension and cardiovascular illness.