Automatic detection for cardiac arrhythmia based on ECG and deep learning approaches

Abstract

Electrocardiograms (ECG) provide information about the electrical activity of the heart, which is useful for diagnosing abnormal cardiac functions such as arrhythmias. Over the works of this thesis, the author aims to provide novel auto-detection approaches for arrhythmia diagnosis based on advanced deep learning methods and preprocessing techniques. The first presented auto-detection method proposed a novel frame-blocking-based preprocessing method that divides a 12-lead ECG recording into frames of a uniform length. The novel preprocessing method addressed the uneven length of clinical signals and limited the loss of valid signals. The multi-labeled classification is then split into several binary classification tasks, with each binary classifier consisting of an attention-based BiLSTM and a ResNet-based network. The advanced preprocessing technique and structure of classifiers fulfilled multi-label classification and achieved a satisfied average F1 score of 0.908. The second presented project utilized an advanced metric-learning algorithm in the training process and proposed to extract comprehensive ECG features on both morphological and temporal domains. The metric-learning-based training model and fused features contributed to more discriminative ECG features than traditional training models merely based on morphological features. With a relatively small model size and GFLOPs, the proposed algorithm achieved an average score of 0.874 and exhibits promising efficiency. Since some rare arrhythmias caused extremely insufficient training samples, the classification performance of the previous classification algorithm can be negatively affected. The third project presented a parallel multi-scale convolution based prototypical network (PM-CNN ProtoNet) for processing the few-shot learning tasks of ECG beats classification. The presented method, which represents a novel attempt to use few-shot learning in ECG auto-detection, exhibits a competitive result when compared to other state-of-the-art models and also demonstrates its potential to address the issue of limited training samples as well as real-world medical applications.

Date of Award	1 Aug 2024
Original language	English
Awarding Institution	The University of Manchester
Supervisor	Jian Lu (Supervisor) & Henggui Zhang (Supervisor)