ECG-CL: A Comprehensive Electrocardiogram Interpretation Method Based on Continual Learning

TL;DR

This research confirms the potential for developing comprehensive ECG interpretation algorithms based on single-lead ECGs but also fosters progress in intelligent wearable applications and introduces an innovative ECG continual learning approach based on parameter isolation, designed to enhance data usage effectiveness and facilitate inter-task knowledge transfer.

Abstract

Electrocardiogram (ECG) monitoring is one of the most powerful technique of cardiovascular disease (CVD) early identification, and the introduction of intelligent wearable ECG devices has enabled daily monitoring. However, due to the need for professional expertise in the ECGs interpretation, general public access has once again been restricted, prompting the need for the development of advanced diagnostic algorithms. Classic rule-based algorithms are now completely outperformed by deep learning based methods. But the advancement of smart diagnostic algorithms is hampered by issues like small dataset, inconsistent data labeling, inefficient use of local and global ECG information, memory and inference time consuming deployment of multiple models, and lack of information transfer between tasks. We propose a multi-resolution model that can sustain high-resolution low-level semantic information throughout, with the help of the development of low-resolution high-level semantic information, by capitalizing on both local morphological information and global rhythm information. From the perspective of effective data leverage and inter-task knowledge transfer, we develop a parameter isolation based ECG continual learning (ECG-CL) approach. We evaluated our model's performance on four open-access datasets by designing segmentation-to-classification for cross-domain incremental learning, minority-to-majority class for category incremental learning, and small-to-large sample for task incremental learning. Our approach is shown to successfully extract informative morphological and rhythmic features from ECG segmentation, leading to higher quality classification results. From the perspective of intelligent wearable applications, the possibility of a comprehensive ECG interpretation algorithm based on single-lead ECGs is also confirmed.

Figures (7)

• Figure 1: A schematic diagram detailing the analysis of an ECG. The central section illustrates the morphology of standard ECG waveforms and the standard interval information between two consecutive heartbeats. Circling this central depiction are examples of various cardiac arrhythmias. The arrhythmias encapsulated by the red dashed lines signify rhythm abnormalities, characterized by fluctuations in the RR intervals immediately preceding and following them. The irregularities enveloped by the green dashed lines denote morphological abnormalities, as indicated by variations in the QRS complex and other sub-waves. The intersection of the two boxes highlights instances wherein both rhythm and morphological abnormalities coexist.
• Figure 2: ECG-CL Processing: An Overview of the Block Diagram.
• Figure 3: The entire ECG processing backbone comprises the multi-resolution architecture (a) along with two decoders for segmentation (a.3) and classification (a.4). The multi-resolution architecture is divided into four stages, with each stage composed of modularized blocks. The branch partition module splits the current filters into two branches, with half retained for the current resolution and the other half allocated for a lower resolution. Following a convolution block (a two-layer residual module), the branch merging module (a.2) employs either strided convolution or deconvolution methods to integrate features of differing resolutions. ECG segmentation requires a high-resolution feature that interpolates low-resolution information. Conversely, ECG classification can be executed on low-resolution features by progressively transferring information from other resolutions to the lowest.
• Figure 4: Illustration of continual learning on model parameters. Empty cells in the training matrix are available for training, whereas grey cells are utilized primarily for forward training, and black cells are frozen. In the compact & retrain matrix, $t$-filled cells indicate the weights for Task $t$, zero indicates released parameters, and the gradient color indicates the weights are utilized for more than one task.
• Figure 5: Confusion Matrix Analysis for Nine-Class Classification on ICBEB Database.