Key result
ECG deep learning performance varies widely, heavily biased by data splitting and external validation.
Why the study?
Deep learning has advanced automated ECG interpretation, but progress remains difficult to compare across studies due to heterogeneous datasets, preprocessing pipelines, and evaluation protocols, along with underreported external validation and reproducibility details.
Systematic Review
Answers are grounded in this paper and cite their sources.
This systematic review synthesizes 2016-2025 ECG deep learning architectures into three modeling paradigms and highlights that protocol choices like patient-wise splitting and external validation heavily impact reported diagnostic metrics, necessitating standardized reporting checklists.
Deep learning has rapidly advanced automated electrocardiogram (ECG) interpretation by enabling end-to-end learning of diagnostic patterns from raw or minimally processed signals. Nonetheless, progress remains challenging to compare across studies due to heterogeneous datasets, preprocessing pipelines, and evaluation protocols, with details on external validation and reproducibility often underreported. This review systematically surveys ECG deep learning research published between 2016 and 2025 across major biomedical and engineering databases, complements it with reference screening, and synthesises evidence across various model architectures, input representations, diagnostic tasks, and evaluation practices. The literature is organised into three predominant modelling paradigms: 1D time-series models, 2D transformed representations, and multimodal fusion frameworks. The influence of design choices, such as lead configuration, window length, augmentation, and training objectives, on performance and deployability is summarised. In applications including arrhythmia detection, atrial fibrillation screening, myocardial infarction identification, and multi-label ECG diagnosis, it is observed that reported performance metrics are heavily affected by protocol choices, particularly patient-wise versus segment-wise splitting and the utilisation of external test sets, which constrains the interpretability of primary metrics. Furthermore, emerging directions are highlighted, including efficiency-aware modeling for edge and wearable settings, reliability-aware learning, and privacy-preserving training. Finally, recurring limitations are identified, along with a reproducibility-oriented checklist and reporting recommendations to foster more reliable benchmarking and translation. Overall, while deep learning demonstrates considerable potential.
No takes yet. Share an insight, caveat, or question.
Yang et al. (2026) conducted a systematic review in Cardiovascular diagnosis via electrocardiogram (including arrhythmia, atrial fibrillation, and myocardial infarction). Deep learning models for automated ECG interpretation (1D time-series models, 2D transformed representations, and multimodal fusion frameworks) was evaluated on Diagnostic performance metrics and deployability across ECG-based cardiovascular diagnostic tasks. Deep learning for ECG diagnosis published from 2016-2025 spans 3 main paradigms, with reported performance heavily affected by patient-wise vs segment-wise splitting and external validation.
Synapse has enriched 10 related papers. Consider them for comparative context: