MACRO: A Multi-Head Attentional Convolutional Recurrent Network for The Classification of Co-Occurring Diseases in 12-Lead ECGs

Vanessa Borst, Robert Leppich, Samuel Kounev

Keywords: ECG signal, Ensemble, Gradient boosting, Healthcare, Multi-head attention, Multi-label classification

TL;DR: We propose a novel method of deep feature extraction and machine learning ensemble for classifcation of co-occurring diseases in 12-lead ECG data, achieving SOTA performance on the CPSC 2018 benchmark dataset with reduced trainable parameter count.

Abstract:

Cardiovascular diseases (CVDs) are a significant global health concern, causing more deaths than several types of cancer combined. Early detection and proper treatment are crucial for better clinical outcomes. Electrocardiograms (ECGs) offer valuable insights into the presence of CVDs. However, extracting powerful features from raw ECG signals for reliable automated diagnostics remains challenging due to high inter- and intra-patient variability, diversity of rhythmic and morphological abnormalities, and noise distributions.
This paper proposes a deep learning architecture for the automated detection of diseases in 12-lead ECG data, capable of recognizing concurrent irregularities through multi-label classification.
Moreover, we present a novel method that combines deep feature extraction with binary machine learning classifiers. To account for the distinct characteristics of various ECG leads, we use a multi-loss optimization strategy.
Our methodology is rigorously evaluated through 10-fold cross-validation using the publicly available CPSC 2018 dataset. With macro $F_1$ and AUC scores reaching up to 85.2\% and 98.0\%, respectively, our approach demonstrates advantage over existing state-of-the-art methods. At the same time, our architecture remains lightweight with approximately 1.7 million trainable parameters, which represents a reduction in the number of parameters of up to 68\% compared to previous methods.
Assessing the generalizability of our approach, we further evaluated it on the PTB-XL dataset and achieved macro $F_1$, AUC, and accuracy scores comparable to existing methods, demonstrating the robustness of our model across diverse datasets.
This advancement holds promise for enhanced automated diagnosis and improved patient care in the context of CVDs. Our code is available at: https://github.com/VanessaBorst/MACRO.

Track: 11. General Track

Registration Id: K2NCHYRX665

Submission Number: 17