Domain and Patient Adversarial Multi-Task Learning for Arrhythmia Classification.

Area of Science:

• Artificial Intelligence
• Cardiology
• Machine Learning

Background:

• Manual screening of electrocardiograms (ECGs) for heart arrhythmias is time-consuming.
• Automated diagnosis models are limited by small clinical datasets.
• Training models with multiple datasets is needed for improved arrhythmia classification.

Purpose of the Study:

• To propose adversarial multi-task learning (AMTL) for extracting domain and patient invariant features from ECG databases.
• To investigate the influence of beat segmentation and normalization on domain invariance.
• To enhance the accuracy and F1 score of heart arrhythmia classification using AMTL.

Main Methods:

• Utilized adversarial multi-task learning (AMTL) to train models on two distinct ECG databases (MIT-BIH Arrhythmia and St Petersburg INCART).
• Investigated the impact of beat segmentation location and beat normalization techniques on achieving domain invariance.
• Compared the performance of domain adversarial models against non-adversarial counterparts.

Main Results:

• Domain adversarial models demonstrated higher accuracy and average F1 scores compared to models without domain adversarial learning.
• The proposed patient and domain adversarial model significantly improved F1 scores from 70% and 74% to 77% on both tested databases.
• Beat segmentation location and normalization influenced the degree of domain invariance achieved.

Conclusions:

• Adversarial multi-task learning with multiple datasets and adversarial tasks effectively improves the F1 score for arrhythmia classification.
• AMTL offers a promising approach to overcome data limitations in training automated diagnostic models for heart arrhythmias.
• The findings establish the clinical relevance of AMTL for enhancing ECG analysis and patient care.