A Machine Learning Approach to Estimate Continuous Blood Pressure From Photoplethysmography

Abnormal blood pressure (BP) necessitates continuous monitoring to mitigate health risks. Traditional methods like sphygmomanometers are accurate but unsuitable for continuous monitoring. Photoplethysmography (PPG) signals offer an alternative for BP estimation. This study proposes a novel machine learning-based method for blood pressure estimation using PPG signal features. Pre-processing techniques, including normalization, baseline correction, and filtering, enhance signal quality. Fiducial points such as systolic peaks, inflection points, and diastolic peaks are automatically detected from processed PPG signals. Various morphological and timing-based features are extracted from these points. Machine learning algorithms, including Linear Regression, Support Vector Machine (SVM), Gaussian Process Regression (GPR), and Decision Tree, are trained with these features to estimate systolic and diastolic blood pressures. Evaluation on a dataset of 657 PPG signals from 219 subjects shows that the cubic support vector machine (CSVM) provides superior accuracy for systolic blood pressure (SBP) estimation, while Gaussian support vector machine (GSVM) outperforms other methods for diastolic blood pressure (DBP) estimation. The proposed method conforms to the standards set by the Association for the Advancements of Medical Instrumentation (AAMI) for BP measurement compatibility. Comparison with other studies demonstrates favorable performance. The machine learning-based method accurately estimates SBP and DBP using PPG signal features, eliminating the need for cuff-based measurement, and enabling continuous monitoring in various applications. This approach has potential for developing commercial prototypes for accurate BP estimation. It holds promise for continuous BP monitoring in research and clinical settings, facilitating baroreflex gain assessment and minimizing instrumentation requirements for physiological signal monitoring.