Explainable artificial intelligence in electrocardiography: A systematic review. Biomed Signal Process Control 2026 Apr 01;114
Date
04/09/2026Pubmed ID
41953749Pubmed Central ID
PMC13056371DOI
10.1016/j.bspc.2025.109325Scopus ID
2-s2.0-105024192349 (requires institutional sign-in at Scopus site) 7 CitationsAbstract
Electrocardiography (ECG) is a cornerstone of cardiac diagnostics, detecting cardiac pathologies ranging from arrhythmias to myocardial infarction. To enhance diagnostic accuracy and efficiency, deep learning models have been developed that now match or surpass human performance in ECG interpretation. However, their opaque reasoning hinders clinical trust and regulatory approval. This challenge is particularly acute for ECG signals because, unlike structured feature data, they are sequential, variable, and noise-prone, making interpretability both more difficult and more essential for clinical adoption. This review systematically evaluates ECG-specific explainable AI techniques using PRISMA guidelines. We screened 380 records across six databases and included 45 peer-reviewed studies examining diverse explainability methods including perturbation-based, gradient-based, intrinsically interpretable, sequence-aware, and counterfactual approaches. Our analysis reveals that perturbation-based techniques designed for structured data prove suboptimal for ECG signals because they treat features as independent rather than temporally dependent. In contrast, methods that transparently reveal model attention to physiologically meaningful ECG intervals such as the P wave, QRS complex, and ST segment demonstrate superior performance across localization accuracy, fidelity, and robustness metrics. While explainable AI in ECG interpretation has advanced substantially, it remains fragmented and insufficiently validated for clinical deployment. The most promising methods reveal what the model attends to in relation to known physiologic features, yet significant challenges persist in stability, computational efficiency, and regulatory readiness. Accelerating clinical translation requires open, multi-institutional benchmarks with cardiologist-annotated explanations and clinician-in-the-loop validation studies that can strengthen trust and support meaningful integration of explainable AI into patient care.
Author List
Taleban A, Sparapani R, Noffke P, Zlochiver S, Lu Q, Widlansky ME, Luo JAuthors
Jake Luo Ph.D. Associate Professor; Director, Center for Biomedical Data and Language Processing (BioDLP) in the Health Informatics & Administration department at University of Wisconsin - MilwaukeeRodney Sparapani PhD Associate Professor in the Data Science Institute department at Medical College of Wisconsin
Michael E. Widlansky MD Center Director, Interim Chief, Professor in the Medicine department at Medical College of Wisconsin
