Machine learning models for automated interpretation of 12-lead electrocardiographic signals: a narrative review of techniques, challenges, achievements and clinical relevance

Background and Objective: Novel advances in machine learning (ML) and its subfield, deep learning (DL), as well as the recent release of large-scale electrocardiogram (ECG) databases, have driven a sharp increase in research related to automated ECG interpretation. This review aims to summarize the recent ML approaches for automatically interpreting standard 12-lead ECG signals. Methods: We searched 10 indexing databases, for original research in English, referring to the application of ML/DL techniques in 12-lead, raw ECG signal analysis. The retrieved titles were filtered based on their relevance. The results were summarized and reported. Key Content and Findings: More than 80% of studies integrated a DL approach, while fewer attempts applied a feature extraction method to obtain inputs for training a simple ML classifier. The average diagnostic accuracy was as high as 90%, while several other performance metrics, such as the area under the curve (AUC), F1-score, sensitivity and specificity, were also employed. DL models generally demanded 10-time more samples for training but were capable of better handling multi-class problems. The most frequently involved disease (49% of studies) was myocardial infarction (MI), while atrial fibrillation (AF) was encountered in more than one-third of studies. Various datasets were used for training and testing, constituting either private collections or publicly available databanks [such as the “Physikalisch-Technische Bundesanstalt” (PTB) dataset and datasets derived from the “China Physiological Signal Challenge” and the “Computing in Cardiology Challenge”]. Overall, DL and simpler ML approaches for automated ECG interpretation display unprecedented growth, reaching remarkably high performances. Conclusions: While such novel tools can support clinicians in reaching reliable diagnoses for life-threatening conditions on the spot, concerns regarding their accountability do exist. Generalizability of the developed approaches is still an issue, possibly mitigable with the extensive deployment of developed models, so as to become massively accessible and validatable. Finally, the observed heterogeneity of the various attempts underlines the need for transparency and reproducibility in the development processes.