Direct Feature Extraction and Diagnosis of ECG Signal in the Compressed Domain

Recently, Internet of Things (IoT) is becoming an important enabler to achieve smart health care for patients, where electrocardiograph (ECG) analyzing has been widely studied. In order to realize long-term detection of ECG under the limited hardware and network resources especially in the power/computing constrained IoT scenarios, compressive sensing (CS) has attracted tremendous research efforts thanks to its high efficiency in sample collection and information retrieval. Most existing researches focus on reconstructing the original ECG signal from the received compressive samples, which however not only leads to high energy consumption but also suffers from unnecessarily huge computational complexity for full-dimensional recovery. To solve such problems, this paper aims at a task-cognizant sparse signal processing technique for retrieving the necessary information only. Specifically, a correlation-assist compression with adaptive length (CCAL) algorithm is developed, where ECG signal is compressed individually within each pseudo-period to retain the information of each separate heartbeat. In doing so, the compression length is dynamically varying, which is determined by calculating the correlation coefficients between adjacent fragments. In addition, a preprocessing filter is adopted to offset the performance loss resulted from compression. With the compressed heartbeat acquired through CCAL, we extract autoregressive (AR) coefficients directly in the compressed domain. Afterwards, an SVM classifier is used to identify the conditions of the ECG signal. Finally, real data from MIT-BIH arrhythmia database is used for validating the proposed techniques that work efficiently in moderate compression ratios, while achieving desired classification accuracy as the non-compression benchmark at much reduced computational complexity.