Using PPG and Machine Learning to Measure Blood Pressure

Measurement of blood pressure is critical for patients with cardiovascular disease. Using a cuff sphygmomanometer with non-invasive registration is currently the most common practice since its fast and does not require an experienced operator. However, it does not allow continuous pressure monitoring and the discomfort of the procedure can discourage consistent use. Photoplethysmography (PPG) is a technique increasingly used for non-invasive, portable devices to monitor arterial oxygen saturation (SpO\(_2\)), heart rate and, more recently, for glycemic control. In this study, we evaluate different methods to estimate blood pressure using PPG. Two methods presented are adapted from the literature, while the third is an improvement proposal. Multiple linear regression (MLR), artificial neural network (ANN), support vector machine regression (SVR) and decison tree regression (DTR) using temporal and spectral PPG features are evaluated. Principal component analysis (PCA) is used in order to reduce dimensionality. The MIMIC (Multiparameter Intelligent Monitoring in Intensive Care) database is used to train and evaluate the approaches. Results indicate that the proposal improves diastolic (DBP) and systolic (SBP) blood pressure estimation with mean absolute errors (MAE) of 6.52 ± 5.75 mmHg and 13.19 ± 11.90 mmHg, respectively.