Improving Systolic Blood Pressure Prediction from Remote Photoplethysmography Using a Stacked Ensemble Regressor

Abstract

Hypertension is a serious health risk, and early diagnosis is key to start treatment and avoid fatal complications. We present a stacked ensemble model to predict systolic blood pressure from remote photoplethysmogramy, which enables cuffless measurements. To train the stacked ensemble model, a large dataset with facial remote photoplethysmogramy signals and ground truth values for blood pressure was collected by trained clinicians. From over 4500 measurements 1410 were selected for training following quality control. Over 100 different features were derived from these signals, including statistical features, time domain and frequency domain features. Nine of these features were selected using a forward feature selector. We verified the accuracy of the model on a separately collected validation set. Using a multi-layer perceptron regressor, linear support vector regressor, radial support vector regressor, and ElasticNet for the base models combined with a support vector machine classifier in the stacked ensemble and a RidgeCV model for the final layer, the mean error of the model is reduced to 1.1 mmHg, mean absolute error to 9.5 mmHg and the standard deviation to 12.3 mmHg. Critically, 79% of the hypertensive patients are correctly identified as hypertensive with a prediction over 140 mmHg.