Automated analysis of wearable ECG: machine learning methods, preprocessing pipelines, and benchmark datasets.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Healthcare
• Cardiology

Background:

• Wearable and mobile electrocardiography (ECG) offers accessible rhythm monitoring but faces challenges in interpretation due to limited leads and motion artifacts.
• Machine learning (ML) presents potential solutions for analyzing these complex ECG signals.

Purpose of the Study:

• To review and synthesize recent ML approaches for wearable ECG analysis.
• To identify key challenges and limitations in current ML applications for wearable ECG.
• To provide a structured comparison of ML methods and guide future research.

Main Methods:

• Systematic review of ML approaches for wearable ECG across signal-quality assessment (SQA), single-lead, and multi-lead interpretation.
• Analysis of preprocessing pipelines, quality handling strategies, and 18 benchmark datasets.
• Evaluation of architectural paradigms, dataset characteristics, and common evaluation deficiencies.

Main Results:

• No single ML architecture consistently outperforms others; classical methods are competitive for rhythm tasks, while deep models offer better generalization with sufficient data.
• Dataset characteristics (scale, demographics, free-living conditions) are stronger predictors of performance than architectural choice.
• Recurring evaluation issues include data leakage, incomplete SQA reporting, and waveform similarity validation without diagnostic concordance.

Conclusions:

• Future progress in wearable ECG ML relies on improved evaluation rigor, representative datasets, and reproducible preprocessing.
• Architectural novelty alone is insufficient; focus must shift to robust methodology and data quality.
• Standardized evaluation protocols are crucial for reliable clinical translation of wearable ECG ML tools.