ECG Guided Automated Diagnostic Intervention of Cardiac Arrhythmias with Extra-Cardiac Noise Detection using Neural Network

The electrocardiogram's (ECG) cyclic activity provides information about a person's emotional, behavioural, and cardiovascular health. Noise that occurs during acquisition and symptomatic patterns produced by pathologies are two examples of irregular behaviours that affect the analysis of these signals. This paper presents a Deep Neural Network algorithm that learns the normal behaviour of an ECG when identifying irregular events, which is studied in two different settings: noise detection and symptomatic events triggered by multiple pathologies. Two noise detection algorithms were developed using an auto-encoder and Convolution Neural Networks (CNN), with the binary class model achieving 98.18 percent accuracy and the multi-class model achieving 70.74 percent accuracy in distinguishing between base wandering, muscle artefact, and electrode motion noise. Recurrent Neural Networks and an autoencoder with Gated Recurrent Units (GRU) configuration were used to create the arrhythmia detection algorithm. With a 56.85 percent accuracy and a 61.13 percent overall sensitivity for a 7-class model. It was determined that the machine's learning mechanism learned characteristics of a regular ECG signal, sacrificing precision for greater generalisation at the moment. In the ECG, the frequency of sporadic events is more discriminatory than the classification of different types of events.