Abnormal beat detection from unreconstructed compressed signals based on linear approximation in ECG signals suitable for embedded IoT devices

In the study of electrocardiogram signal monitoring systems, signal compression techniques for effective signal transmission and abnormal beat detection for arrhythmia diagnosis are paramount areas. The general abnormal beat detection has a problem of unsuitability for low-power and low-capacity embedded devices because it requires reconstructing a compressed signal to generate an auxiliary signal for fiducial point (FP) detection. In this study, we propose a method to compress signals into a small number of vertices, including the FP, using an optimized dynamic programming-based linear approximation. Then, we detect the FP from the vertices and classify the abnormal beat. The proposed method minimizes the amount of memory usage and computation by detecting the FP using the vertex’s feature value without reconstructing the compressed signal. The signal compression performance of the proposed method showed an average compression ratio of 7.05:1 and a root-mean-square difference of 0.78% for 48 records of the MIT-BIH arrhythmia database. In addition, the premature ventricular contraction abnormal beat detection performance using only QR interval feature showed 84.07% sensitivity and 93.70% accuracy; when R-peak’s amplitude and RR interval features were added, the sensitivity and accuracy increased to 96.65% and 93.76%, respectively. Therefore, we confirmed that the proposed method could effectively compress electrocardiogram signals based on linear approximation and detect abnormal beats without signal reconstruction.