Using artificial intelligence in the development of diagnostic models of coronary artery disease based on ECG features: A scoping review

<b>Introduction</b>Coronary Artery Disease (CAD) remains the leading global cause of cardiovascular mortality, with its prevalence demonstrating a significant increase from 1811 to 2549 per 100,000 individuals between 1990 and 2019. Projections further indicate rising CAD mortality rates in low- and middle-income countries by 2050, underscoring its persistent and growing public health burden. Early diagnosis is crucial for mitigating this burden. While coronary angiography is the gold-standard diagnostic method, its invasive nature and dependency on specialized equipment and expertise create barriers to access, particularly in primary care settings and underserved regions. This limitation contributes to diagnostic delays, health inequities, and severe sequelae such as myocardial infarction and heart failure. The 12-lead electrocardiogram (ECG), being non-invasive, low-cost, and widely available, offers immense potential for early CAD detection in resource-constrained environments. However, its utility is hampered by the subtlety of ECG changes indicative of CAD, which can be difficult to discern visually and susceptible to inter-observer variability in interpretation. Artificial Intelligence (AI) emerges as a promising solution to overcome these limitations. An increasing body of research, spanning medicine, computer science, and engineering, has developed AI models using ECG features—ranging from raw signals to extracted morphological characteristics (e.g., PR interval, QRS duration, ST-segment morphology, QTc) and multidimensional features (time, frequency, non-linear domains)—for CAD detection. Initial reports demonstrate high performance metrics (Accuracy: 90%-99.21%; Sensitivity: 91.13%-98.43%; Specificity: 95.88%-100%) using diverse approaches like 1D-Convolutional Neural Networks (1D-CNN), Artificial Neural Networks (ANN) coupled with signal transforms (e.g., Wavelet, Fourier), and Support Vector Machines (SVM) with feature reduction (e.g., PCA). Despite these encouraging laboratory results, there appears to be a critical gap between development and clinical implementation. The clinical relevance, external validation robustness, and translational readiness of these models remain largely unknown, contributing to potential research redundancy. This scoping review aims to systematically map this evolving field, elucidating how AI techniques are leveraged to enhance ECG-based automated CAD diagnosis, thereby identifying the knowledge gap between promising research and practical clinical application needed to reduce the global CAD burden.<b>Methods:</b>1. Review Framework​​(1) Conducted in accordance with theJoanna Briggs Institute (JBI) Manual for Scoping Reviews.(2) Reporting follows thePRISMA-ScR checklis (Preferred Reporting Items for Systematic Reviews and Meta-Analyses Extension for Scoping Reviews).2. Review Questions​​(1) AI Methods &amp; ECG Features:Which AI techniques and ECG features are used in automated CAD diagnosis?(2) Model Performance:What is the diagnostic performance (e.g., accuracy, sensitivity, specificity) of AI-based ECG models for CAD?(3) Validation Rigor:To what extent are these models internally/externally validated?3. Inclusion Criteria​​​​Domain​​​​Criteria​​​​Population​​Patients with suspected or confirmed CAD​​Concept​​AI-driven diagnostic models using​​ECG-only features​​​Context​​All clinical or computational settings​​Evidence​​Studies reporting quantitative model performance metrics​​Language​​Full-text available in English or Chinese​​4. Search Strategy​​(1) Databases: PubMed, Embase, Cochrane CENTRAL, Web of Science, IEEE Xplore, CNKI, Wanfang.(2) Grey Literature: Google Scholar (first 100 relevant records).(3) Timeframe: Inception to June 30, 2025​​.(4) Process:① Step 1: Identify MeSH/keywords (e.g., "AI", "ECG", "coronary artery disease") via NCBI MeSH.② Step 2: Pilot PubMed search→refine logic expressions.Step 3: Execute final search→deduplicate (EndNote 20).③ Step 4: Backward citation tracing of included studies.5. Source Selection​​(1) Screening:① Phase 1: Title/abstract screening→2 independent reviewers (kappa≥0.8 after pilot testing).② Phase 2: Full-text assessment→disagreements resolved by a third reviewer.(2) Pilot Test:30 random records reviewed→revise inclusion criteria until 80% inter-rater agreement is achieved.(3) Output:PRISMA flowchart documenting exclusions.6. Data Extraction​​Tool: Structured Excel form (pilot-tested on 5 studies, refined to≥80% agreement).Extracted Fields:​​Category​​​​Variables​​Study Metadata​​Authors, year, country, design, objectivesPopulation &amp; Data Source​​Cohort characteristics (age, sex), database (e.g., PTB-XL), sample sizeAI Model Details​​Algorithm (e.g., 1D-CNN, SVM), ECG features (e.g., QTc, spectral entropy)Performance &amp; Validation​​Metrics (Acc, Sens, Spec), validation method (internal/external), generalizabilityCritical Appraisal​​Clinical relevance, limitations, implementation barriers7. Data Synthesis &amp; Presentation​​(4) Descriptive Statistics:Frequencies/percentages for study characteristics, AI methods, and feature types.(5) Thematic Analysis:① Performance Trends: Stratified comparison of metrics by algorithm type.② Validation Gaps: Proportion of externally validated models; inconsistency analysis.③ Output:Evidence tables, cross-comparison charts (e.g., forest plots for accuracy distributions), narrative summaries.