Blood pressure estimation using smartphone

This diploma thesis describes an experimental cuffless measurement of systolic and diastolic blood pressure (SBP and DBP) using only a smartphone. Photoplethysmographic signal (PPG), which was measured by the smartphone camera and phonocardiographic signal (PCG), which was captured by an external microphone in the mitral valve area, were used to estimate blood pressure (BP). The BP estimate was divided into 2 parts. The first was the regression estimation of BP and comparison of several machine learning (MLR) and convolutional neural network (CNN) methods according to the use of the color channel to extract PPG from the video. The second objective was the classification of BP into the category of normotension or hypertension. The PPG signal was filtered for extraction using a bandpass filter with finite impulse response (FIR) (0,5–5 Hz) followed by feature extraction. For estimation of vascular transit time (VTT) and cardiac ejection time (ET), two main cardiac echoes S1 and S2 were detected in the PCG signal. Furthermore, raw sections of PPG signals (RGB video channels) and PCG signals of different time lengths were used for the CNN models. Among the regression approaches, the most accurate algorithm was the support vector regression (SVR) method, which achieved acceptable error according to the standards of the Association for the Advancement of Medical Instrumentation (AAMI) and the British Hypertension Society (BHS) in experimental work. The mean absolute error (MAE) was determined to be 7.64±6.87 mmHg for SBP and 5.07±4.88 mmHg for DBP. The red channel was evaluated as the most appropriate color channel for extracting PPG from the video recordings for regression estimation. For classification, the AdaBoost algorithm with weak random forest (RF) classifiers proved to be the most reliable. It achieved an F1 score of almost 80 % on the PPG features extracted from the combination of red and green color channels of the video