Hardware Design for Predicting Early Signs of Sudden Cardiac Arrests from ECG Signals

Abstract

Electrocardiogram (ECG) is a non-invasive way to record the electrical activities of the heart and is recognized by its features, P-QRS-T. Any change in the amplitude or duration of these features indicates the presence of an abnormality, known as cardiac arrhythmias (CA). Among the most fatal CAs, sudden cardiac arrests (SCA) can cause death unexpectedly if left untreated, and hence, their early prediction is important to receive timely medical care. SCAs can be predicted by the detection of one of the most life threatening CAs, known as ventricular tachycardias (VT), which are characterized by the presence of three or more consecutively occurring premature ventricular contraction (PVC) beats in an ECG signal. PVCs are common among the general population, but they pose a threat to the human heart only if they occur frequently in a group of three or more with varying QRS morphologies. Based on the QRS morphology, PVC beats are classified as multifocal or unifocal, and based on the frequency of occurrences, PVCs are classified as ventricular bigeminies, ventricular trigeminies, non-sustained ventricular tachycardias (NSVT), and sustained ventricular tachycardias (SVT). However, according to medical experts, the highest risk of SCAs is related to frequent episodes of multifocal NSVTs and SVTs. On the other hand, ventricular bigeminy and trigeminy are not directly related to any fatal CAs, but they will require further medical aid if they occur frequently enough to prevent the heart from pumping the required amount of blood to all the organs of the body. In this work, a hardware-efficient FPGA-based design for predicting the early signs of SCAs is proposed by detecting PVC beats and classifying them into six major categories of ventricular arrhythmias (VAs), namely multifocal PVCs, unifocal PVCs, ventricular bigeminies, ventricular trigeminies, NSVTs, and SVTs. The SCA prediction system consists of the following stages: pre-processing, feature extraction, PVC detection, and VA classification. For pre-processing, a denoising technique using the modified lifting-based discrete wavelet transform (MLDWT) is used to combat all the ECG noises as well as enhance the QRS complexes in the ECG signal. For efficient detection of PVC beats, an accurate feature extraction stage that extracts R peaks, T peaks, and the Teager energy operator (TEO) is employed. With the extracted features, a characteristic matching algorithm is used for PVC detection, and an adaptive decision logic-based (ADL) classifier is utilized for VA classification, resulting in a detection accuracy rate of 98.2% when tested using the online ECG databases, viz., the MIT-BIH arrhythmia database (MITDB) and the MIT-BIH supraventricular database (SVDB). The complete hardware design of the SCA prediction system, when implemented on the Nexys 4 DDR Artix-7 FPGA board, outputs the number of PVCs detected and VA classified, based on which an alert on the risk of SCAs is provided while utilizing 10.4% of the total available hardware resources on the FPGA board. For future integration of the SCA prediction system into wearable healthcare devices, an ASIC implementation of the PVC detection and VA classification is performed, resulting in a place and route area of 0.02 mm2 and a power utilization of 3.47 μW at an operating frequency of 100 KHz when implemented using SCL 180 nm CMOS technology.