Exploiting Sparsity for Efficient Compression and Analysis of ECG and Fetal-ECG Signals

Over the last decade there has been an increasing interest in solutions for the continuous monitoring of health status with wireless, and in particular, wearable devices that provide remote analysis of physiological data. The use of wireless technologies have introduced new problems such as the transmission of a huge amount of data within the constraint of limited battery life devices. The design of an accurate and energy efficient telemonitoring system can be achieved by reducing the amount of data that should be transmitted, which is still a challenging task on devices with both computational and energy constraints. Furthermore, it is not sufficient merely to collect and transmit data, and algorithms that provide real-time analysis are needed. In this thesis, we address the problems of compression and analysis of physiological data using the emerging frameworks of Compressive Sensing (CS) and sparse signal processing. In particular, we develop new methods and propose specific applications for compression and real-time analysis with a special focus on electrocardiogram (ECG) and fetal electrocardiogram (fECG) signals. Moreover, the proposed frameworks and results could potentially be extended to a much wider class of physiological signals. To improve the performance of current CS frameworks, we introduce a novel sparsifying dictionary, which, when used in combination with the existing reconstruction algorithms, allows for accurate recovery of ECG signals at high compression ratios. While the CS compression is a low-complex procedure, signal recovery can be computationally expensive, and very often we are only interested in extracting certain information without necessarily needing the full reconstructed signal. This is the case of clinical evaluation based on analyzing beat-to-beat timing variation, which is calculated from the time occurring between two consecutive beats, identified by the R-peak in the ECG signal. Thus, we consider the possibility of avoiding signal recovery and directly performing beat detection in the compressed domain. To this end we propose a new method capable to provide a real- time detection of R-peaks with a limited complexity with respect to typical reconstruction procedures. Increasing the compression ratio is the main objective, but, due to distortion, the signal quality is an issue that should always be kept under control. In particular, due to distortion induced by the sampling process it is essential to guarantee that all clinically relevant information for a given task is preserved, in order to prevent significant degradation in the performance of any standard or novel clinically relevant algorithm. Thus, in order to assess the effectiveness of the proposed dictionary and beat detector, we verify the impact of CS at different compression ratios on Atrial Fibrillation (AF) detection. We demonstrate the possibility of accurately detecting episodes of atrial fibrillation (AF) directly on the compressed measurements which has enormous potential for extending long term monitoring of transient AF and other episodic phenomena, which requires long term monitoring and processing of data on energy-constrained devices. Moreover, the proposed dictionary allows to increase the accuracy of AF detection for a given compression ratio with respect to standard method. We also design a framework for the compression of abdominal fECG and to obtain real time information of the fetal heart rate, providing a suitable solution for real-time, very low power fECG monitoring. Taking advantage of the sparse representation with the proposed dictionary, it is possible to increase the quality of the compressed signals, and, at the same time, perform fetal and maternal beat detection/classification. The detection scheme uses Independent Component Analysis (ICA), which we propose to compute directly in the compressed domain before signal reconstruction. The need for fast and robust reconstruction algorithms inspired us to modify an existing reconstruction algorithm and make it error-tolerant. The proposed method guarantees better immunity against inaccuracy caused by noisy original signals and possibly ill-conditioned reconstruction procedures. Finally, we compare fECG compression using CS to a standard compression scheme using wavelets in terms of energy consumption, reconstruction quality and, more importantly, performance of fetal heart beat detection on the reconstructed signals. An actual implementation on a commercial device prove the suitability of CS as an ultra-low power compression technique for fECG signals. Indeed, CS allows for significant reductions in energy consumption in the sensor node and the detection performance is comparable to that obtained on original signals for compression ratios of up to 75%.