Photoplethysmography-Based Biomedical Signal Processing

In this dissertation, photoplethysmography-based biomedical signal processing methods are developed and analyzed. The developed methods solve problems concerning the estimation of the heart rate during physical activity and the monitoring of cardiovascular health. For the estimation of heart rate during physical activity, two methods are presented that are very accurate in estimating the instantaneous heart rate at the wrist and, at the same time, are computationally efficient so that they can easily be integrated into wearables. In the context of cardiovascular health monitoring, a method for the detection of atrial fibrillation using the video camera of a smartphone is proposed that achieves a high detection rate of atrial fibrillation (AF) on a clinical pre-study data set. Further monitoring of cardiovascular parameters includes the estimation of blood pressure (BP), pulse wave velocity (PWV), and vascular age index (VAI), for which an approach is presented that requires only a single photoplethysmographic (PPG) signal. Heart rate estimation during physical activity using PPG signals constitutes an important research focus of this thesis. In this work, two computationally efficient algorithms are presented that estimate the heart rate from two PPG signals using a three axis accelerometer. In the first approach, adaptive filters are applied to estimate motion artifacts that severely deteriorate the signal quality. The non-stationary relationship between the measured acceleration signals and the artifacts is modeled as a linear system. The outputs of the adaptive filters are combined to further enhance the signal quality and a constrained heart rate tracker follows the most probable high energy continuous line in the spectral domain. The second approach is modest in computational complexity and very fast in execution compared to existing approaches. It combines correlation-based fundamental frequency indicating functions and spectral combination to enhance the correlated useful signal and suppress uncorrelated noise. Additional harmonic noise damping further reduces the impact of strong motion artifacts and a spectral tracking procedure uses a linear least squares prediction. Both approaches are modest in computational complexity and especially the second approach is very fast in execution, as it is shown on a widely used benchmark data set and compared to state-of-the-art methods. The second research focus and a further major contribution of this thesis lies in the monitoring of the cardiovascular health with a single PPG signal. Two methods are presented, one for detection of AF and one for the estimation of BP, PWV, and VAI. The first method is able to detect AF based on a smartphone filming the finger placed on the video camera. The algorithm transforms the video into a PPG signal and extracts features which are then used to discriminate between AF and normal sinus rhythm (NSR). Perfect detection of AF is already achieved on a data set of 326 measurements (including 20 with AF) that were taken at a clinical pre-study using an appropriate pair of features whereby a decision is formed through a simple linear decision equation. The second method aims at estimating cardiovascular parameters from a single PPG signal without the conventional use of an additional electrocardiogram (ECG). The proposed method extracts a large number of features from the PPG signal and its first and second order difference series, and reconstructs missing features by the use of matrix completion. The estimation of cardiovascular parameters is based on a nonlinear support vector regression (SVR) estimator and compared to single channel PPG based estimators using a linear regression model and a pulse arrival time (PAT) based method. If the training data set contains the person for whom the cardiovascular parameters are to be determined, the proposed method can provide an accurate estimate without further calibration. All proposed algorithms are applied to real data that we have either recorded ourselves in our biomedical laboratory, that have been recorded by a clinical research partner, or that are freely available as benchmark data sets