Machine Learning Techniques Applied to Data Analysis and Anomaly Detection in ECG Signals

摘要

In this article Tomasz Andrysiak presents the use of sparse representation of a signal based on overcomplete dictionaries of base functions and a QRS detection method using artificial neural network to detect anomalies in the analyzed ECG signals. Performance of the proposed method was tested by means of a widely available database of ECG signals, i.e., the MIT-BIH Arrhythmia Database, and the obtained experimental results confirmed its effectiveness for anomaly detection in the analyzed ECG signals. The current dynamic and intensive development of information technologies and more excellent methods of processing, analysis, and recognition of signals have enriched medicine with new quality techniques in diagnosis and therapy. At present, a special interest is put to noninvasive diagnostic methods for rapid and objective determination of vital signs and, in particular, automatic electrocardiogram performed outside medical facilities. The sparse representation of a signal was performed in an adaptive manner by means of the matching pursuit algorithm. In each step of the algorithm, there was implemented a linear signal decomposition of features belonging to the dictionary with redundancy. Two kinds of dictionaries were tested. The first type included elements that form the analytical base functions, and the second represented the essential components of the ECG signal searched with the use of the K-Singular Value Decomposition (K-SVD) algorithm. Recognition of the QRS was realized by means of a unidirectional, multidimensional neural network with backpropagation of an error. For this solution, tests were performed on a selection of appropriate neural network architectures and the impact of a particular activation function on the recognition results. Anomaly detection was realized with estimation of sparse representation parameters of the tested ECG signal within the field of the recognized QRS complex, and it was compared with the reference values. Performance of the proposed method was tested using a widely available database of ECG signals MIT-BIH Arrhythmia Database. The obtained experimental results confirmed the effectiveness of the proposed method of anomaly detection in the analyzed ECG signals.