Background: Artificial intelligence (AI) has rapidly expanded the capabilities of electrocardiography beyond conventional rhythm interpretation toward automated diagnosis, disease screening, prognostic assessment, and cardiovascular risk prediction. Machine-learning and deep-learning models can identify complex ECG patterns that may be subtle or imperceptible to human interpretation, enabling applications in atrial fibrillation and other arrhythmias, myocardial infarction, left ventricular systolic dysfunction, heart failure, cardiomyopathy, sudden cardiac death risk, and prediction of adverse cardiovascular outcomes. This systematic review evaluates the diagnostic, predictive, and clinical applications of AI in electrocardiographic analysis and critically examines model performance, external validation, explainability, bias, and readiness for clinical implementation. Current evidence demonstrates particularly strong performance for automated ECG classification and arrhythmia detection, while AI-enabled models for structural heart disease, ischemia, and prognostic prediction are increasingly promising. However, reported performance frequently declines when models are evaluated in external or real-world populations, and substantial heterogeneity exists in ECG acquisition, datasets, model architectures, reference standards, and outcome definitions. Important methodological concerns include overfitting, class imbalance, inadequate separation of training and testing datasets, limited external validation, insufficient reporting of calibration, demographic under-representation, and restricted assessment of algorithmic fairness. Although several studies report high sensitivity, specificity, and area-under-the-curve values, diagnostic accuracy alone does not establish clinical utility. Evidence that AI-assisted ECG interpretation improves patient management, downstream testing, cardiovascular outcomes, or cost-effectiveness remains comparatively limited. Overall, AI-enabled ECG analysis has substantial potential to enhance cardiovascular diagnosis, screening, and risk stratification, but widespread clinical implementation requires robust prospective validation, transparent reporting, explainable and equitable algorithms, standardized performance assessment, and demonstration of clinically meaningful benefit beyond conventional ECG interpretation.
Electrocardiography (ECG) is one of the most widely used diagnostic investigations in cardiovascular medicine because it is rapid, non-invasive, inexpensive, and capable of providing immediate information regarding cardiac rhythm, conduction, myocardial ischemia, chamber abnormalities, and other electrophysiological disturbances. Despite its clinical importance, conventional ECG interpretation remains dependent on visual recognition of waveform characteristics and predefined diagnostic criteria. Diagnostic accuracy can vary according to physician expertise, disease prevalence, ECG quality, and the subtlety of the underlying abnormality. Automated computer interpretation has been incorporated into ECG systems for decades, but conventional rule-based algorithms have historically demonstrated variable performance and have not eliminated the need for expert interpretation.
The emergence of artificial intelligence (AI), particularly machine learning and deep learning, has fundamentally expanded the potential information obtainable from electrocardiographic signals. Rather than relying exclusively on manually defined features such as PR interval, QRS duration, ST-segment deviation, or QT interval, contemporary AI models can learn complex mathematical representations directly from raw ECG waveforms. Convolutional neural networks and related deep-learning architectures can identify temporal and morphological patterns distributed across multiple leads that may be difficult or impossible to recognize consistently through conventional visual interpretation.
One of the earliest demonstrations of the clinical potential of deep learning in electrocardiography involved automated arrhythmia classification. Hannun et al. developed a deep neural network for ambulatory single-lead ECG interpretation that achieved arrhythmia classification performance comparable with expert cardiologists across multiple rhythm categories [1]. This work established that sufficiently large labelled ECG datasets could support automated interpretation at clinically meaningful levels of accuracy and stimulated rapid development of AI-based rhythm analysis. Deep-learning methods were subsequently extended to standard 12-lead ECGs. Ribeiro et al. trained a deep neural network using more than two million labelled examinations to identify common electrocardiographic abnormalities, demonstrating the feasibility of large-scale automated ECG interpretation across routine clinical diagnostic categories [2]. Such studies shifted the role of AI from experimental pattern recognition toward a potentially scalable clinical interpretation system.
From Automated Interpretation to Hidden Phenotype Detection
The most transformative aspect of AI-enabled ECG analysis may not be its ability to reproduce diagnoses already visible to clinicians but rather its ability to detect clinically relevant biological information that is not readily apparent on routine inspection. Deep-learning algorithms can potentially identify subtle patterns distributed across waveform morphology, timing, amplitude, repolarization, and inter-lead relationships that function as digital biomarkers of underlying disease. A landmark example was the development of an AI-enabled ECG model for detection of reduced left ventricular ejection fraction. Attia et al. demonstrated that a standard 12-lead ECG could be used to identify patients with left ventricular systolic dysfunction despite the absence of direct ventricular-function measurements on the ECG itself [3]. This finding introduced the concept of the ECG as a latent phenotyping tool capable of identifying structural heart disease that would traditionally require echocardiography or other cardiac imaging. A similar approach was applied to atrial fibrillation (AF). Attia et al. demonstrated that an AI model applied to ECGs recorded during sinus rhythm could identify patients with a history or increased likelihood of AF [4]. The clinical significance of this approach lies in the intermittent nature of paroxysmal AF; patients may have a normal sinus-rhythm ECG during routine assessment despite having clinically important arrhythmia at other times. AI-derived signatures may therefore allow the ECG to function as a risk-enrichment tool for determining which patients may benefit from extended rhythm monitoring. These studies fundamentally changed the conceptual role of electrocardiography. Rather than interpreting only observable electrical abnormalities, AI models attempt to extract latent physiological signatures associated with structural, metabolic, or future cardiovascular disease.
Detection of Metabolic Abnormalities
AI-enabled ECG analysis has also been investigated beyond conventional cardiovascular diagnoses. Electrolyte abnormalities, particularly hyperkalemia, can produce characteristic ECG changes, but traditional ECG interpretation is insufficiently sensitive to exclude clinically significant potassium elevation. Galloway et al. developed and validated a deep-learning model capable of screening for hyperkalemia using ECG signals, demonstrating that AI could identify potassium-related electrical patterns beyond standard visual criteria [5]. The ability to estimate or detect biochemical abnormalities from an inexpensive physiological signal illustrates the broader concept of the ECG as a digital biomarker platform rather than solely a rhythm diagnostic tool. AI-based hyperkalemia detection may be particularly relevant in patients with chronic kidney disease, dialysis dependence, heart failure, or treatment with medications that influence potassium homeostasis. Nevertheless, such systems should be considered adjunctive screening tools rather than substitutes for laboratory potassium measurement because predictive value depends heavily on disease prevalence, threshold selection, and the clinical population.
Prognostic Information Encoded Within the ECG
AI has also revealed potentially important prognostic information within ECG waveforms. Raghunath et al. trained a deep neural network using more than one million ECGs to predict one-year all-cause mortality directly from raw 12-lead ECG voltage data [6]. The model demonstrated strong discrimination, including among patients whose ECGs had been interpreted as normal by physicians. This suggested that apparently normal ECGs may contain subtle physiological patterns associated with future risk that are not captured by conventional interpretation. This observation has important implications for cardiovascular risk assessment. Traditional risk prediction relies on combinations of demographics, symptoms, laboratory values, medical history, imaging, and established clinical scores. AI-ECG models may provide an additional layer of information derived from electrical cardiac phenotype without requiring new testing.
However, prognostic discrimination must be distinguished from clinical utility. A model may predict mortality accurately without identifying a modifiable mechanism or providing an actionable treatment pathway. Therefore, the usefulness of AI-derived prognostic information depends not only on discrimination but also on calibration, incremental value over established risk models, interpretability, and whether model-guided interventions improve outcomes.
Structural Heart Disease and Cardiomyopathy
AI-ECG models have also been developed for structural myocardial diseases. Cardiac amyloidosis represents an important example because diagnosis is frequently delayed, while early treatment has become increasingly important. AI-enabled ECG approaches have demonstrated the ability to detect patterns associated with cardiac amyloidosis before or around the time of conventional clinical recognition [7.8]. These models may help identify patients requiring further evaluation using echocardiography, cardiac magnetic resonance imaging, nuclear scintigraphy, or hematological testing. The same general principle has been investigated for hypertrophic cardiomyopathy (HCM). AI models can integrate subtle abnormalities in depolarization, repolarization, QRS morphology, and other waveform characteristics to identify individuals with underlying hypertrophic disease. Such tools are particularly attractive because HCM can remain clinically silent and may not always produce obvious conventional ECG criteria.
Nevertheless, structural-disease screening illustrates an important challenge in AI-ECG implementation. Even a model with high sensitivity and specificity may generate many false-positive referrals when applied to low-prevalence populations. Clinical usefulness therefore depends on the selected operating threshold, target population, downstream testing strategy, and consequences of false-positive and false-negative classifications.
Arrhythmia Detection: The Most Mature AI-ECG Application
Arrhythmia classification remains one of the most extensively studied applications of AI in electrocardiography. A recent systematic review and meta-analysis encompassing 20 studies and more than 5.5 million ECG recordings demonstrated strong diagnostic performance of AI-based approaches for arrhythmia detection using 12-lead ECGs [9]. However, substantial variability remained in algorithms, patient populations, target rhythms, reference standards, and validation procedures. AI is particularly well suited to rhythm classification because arrhythmias are fundamentally defined by temporal and morphological electrical patterns. Deep-learning systems can analyse large numbers of waveform features simultaneously, potentially reducing missed diagnoses and accelerating interpretation in high-volume clinical settings. AI may also enhance wearable and ambulatory ECG technologies. Single-lead devices, smartwatches, patches, and portable ECG systems can generate enormous quantities of longitudinal rhythm data that cannot feasibly be reviewed manually. AI-based classification provides a mechanism for filtering such data and prioritizing clinically important episodes.
However, the performance of an algorithm in a curated dataset may not reproduce the same results in wearable or real-world recordings affected by motion artefact, poor electrode contact, population differences, and low disease prevalence. Therefore, deployment-specific validation remains necessary.
Acute Myocardial Infarction and Ischemia Detection
The diagnosis of acute myocardial infarction (AMI) represents another important area of AI-ECG research. Conventional ECG interpretation performs well in classic ST-segment elevation myocardial infarction but is less sensitive for subtle ischemic changes, non-ST-segment elevation myocardial infarction, posterior infarction, and coronary occlusion without conventional STEMI criteria. A 2026 systematic review of AI-enhanced ECG for AMI detection identified 88 studies and found substantial variation in model architecture, validation design, and reported diagnostic performance. Most studies were retrospective, convolutional neural networks were the predominant model type, and reported performance varied considerably [10].
A diagnostic meta-analysis subsequently reported that AI-enabled ECG demonstrated high overall discrimination for AMI, with especially strong performance for STEMI but substantially weaker sensitivity for NSTEMI [11]. This distinction is clinically important because high aggregate diagnostic accuracy can obscure major differences between disease subtypes. A separate recent meta-analysis focused specifically on STEMI and reported pooled diagnostic performance approaching the high 90% range for sensitivity and specificity across eligible studies [12]. These findings suggest that AI may support rapid triage and recognition of acute coronary occlusion, particularly where immediate specialist ECG interpretation is unavailable. However, AI-ECG should not be interpreted as a replacement for clinical assessment, serial ECGs, high-sensitivity cardiac troponin testing, or coronary imaging. Its most plausible near-term role is as an adjunct that improves triage, prioritization, and recognition of subtle ischemic patterns.
Heart Failure and Left Ventricular Dysfunction
AI-enabled detection of left ventricular systolic dysfunction is among the most clinically advanced applications of AI-ECG. The underlying rationale is attractive: a routine ECG is far less expensive and more widely available than echocardiography, and AI may identify patients who should undergo definitive cardiac imaging.
A recent systematic review and meta-analysis of randomized controlled trials reported that AI-ECG screening increased detection of newly diagnosed low ejection fraction compared with routine care without substantially increasing overall echocardiography referral rates [13]. Across three randomized trials involving more than 37,000 participants, AI screening increased identification of low EF, although positive predictive values remained modest. Broader diagnostic meta-analysis has similarly demonstrated useful performance for AI-ECG in identifying heart failure or left ventricular systolic dysfunction, while also documenting variability across algorithms and populations [14,15].
Broadening Spectrum of AI-ECG Applications
The scope of AI-ECG research has expanded rapidly. A recent systematic review examining more than 100 studies published between 2019 and 2025 identified applications involving heart failure, myocardial infarction, atrial fibrillation, other arrhythmias, signal denoising, synthetic ECG generation, wearable monitoring, and explainable AI [16]. AI-ECG models are also being investigated for cardiovascular risk prediction rather than diagnosis alone. Recent systematic evidence suggests that ECG-based AI may help identify individuals at increased risk of future cardiovascular events using latent waveform information that conventional risk assessment may not capture [17].
Sudden cardiac death prediction represents another potentially important application. A 2025 systematic review and meta-analysis found promising diagnostic and predictive performance across several AI approaches using ECG signals and heart-rate variability [18]. Nevertheless, heterogeneity across study populations, endpoints, model architectures, and validation strategies was substantial, limiting direct clinical translation. These developments illustrate the transition of AI-ECG from a diagnostic classifier toward a multidimensional cardiovascular biomarker capable of supporting diagnosis, screening, prognosis, and longitudinal surveillance.
Explainability and the Black-Box Problem
A major challenge in deep-learning ECG analysis is model interpretability. Traditional ECG interpretation is based on recognizable physiological features such as P-wave morphology, atrioventricular conduction, QRS duration, ST-segment deviation, and repolarization. Deep neural networks may instead generate predictions using complex high-dimensional relationships that are difficult for clinicians to inspect directly. Explainable AI methods attempt to identify which portions of the ECG influence model predictions. Techniques include saliency maps, gradient-based methods, perturbation approaches, attention mechanisms, counterfactual explanations, and intrinsically interpretable architectures.
A recent systematic review of explainable AI in electrocardiography identified 45 studies and concluded that although explainability techniques have advanced considerably, their validation remains fragmented. Methods that localize model attention to physiologically meaningful waveform components such as P waves, QRS complexes, and ST segments appear particularly promising, but stability, fidelity, reproducibility, and clinician usability remain important unresolved issues [19].
External Validation and Generalizability
A recurrent weakness in cardiovascular AI is that high internal performance does not guarantee transportability across institutions, countries, demographic groups, ECG devices, or clinical environments. Model performance can deteriorate when disease prevalence, acquisition protocols, patient characteristics, data quality, or healthcare workflows differ from the original development setting.
Contemporary cardiovascular AI literature identifies inadequate external validation as a major barrier to implementation. Bias may arise from unrepresentative training data, inconsistent reference standards, missing demographic groups, label errors, and differences between development and deployment environments [20]. Recent large-scale external validation across a United States healthcare system demonstrated that an AI-ECG platform originally developed in another healthcare environment could be assessed across emergency and cardiac diagnostic targets, illustrating the increasing emphasis on transportability rather than internal discrimination alone [21].
AI-ECG Beyond Cardiovascular Diagnosis
The ability of AI to identify potassium abnormalities further illustrates how ECG signals may encode systemic physiological information. A 2026 systematic review identified 33 peer-reviewed studies examining AI-ECG for potassium detection or estimation. Hyperkalemia was the most extensively studied application, with several models reporting high discrimination, but heterogeneity in potassium thresholds, ECG modalities, validation approaches, and reporting limited quantitative synthesis [22]. Importantly, clinical usefulness has begun to be tested prospectively. A pragmatic randomized trial evaluated real-time AI-ECG alerts for dyskalemia management in emergency care, moving the technology from retrospective diagnostic modelling toward intervention-based evaluation [23].
From Diagnostic Accuracy to Clinical Utility
The rapid growth of AI-ECG literature has created a risk of equating high area-under-the-curve values with clinical readiness. Diagnostic discrimination is only one component of clinical utility. A model intended for population screening must demonstrate acceptable calibration, positive predictive value, false-positive burden, equity across demographic groups, reproducibility across devices and sites, and a clearly defined downstream clinical pathway.
Recent evidence indicates that translational maturity differs considerably across AI-ECG applications. Reduced-ejection-fraction screening and STEMI triage have progressed further toward prospective or randomized evaluation, whereas several other applications—including occult AF prediction, HFpEF, cardiac amyloidosis, long-term heart-failure prediction, and some forms of ischemia detection remain predominantly at the retrospective validation or risk-enrichment stage [24]. This distinction is critical. An AI system may detect a hidden phenotype accurately but still fail to improve patient outcomes if no effective downstream action follows the alert. Conversely, false-positive predictions may increase unnecessary testing, anxiety, cost, and clinical workload.
Evidence from randomized trials of cardiovascular AI more broadly remains limited. Recent systematic synthesis has found that AI interventions may improve some diagnostic or process outcomes, but certainty regarding hard cardiovascular endpoints remains low because of heterogeneous trial designs, limited follow-up, and relatively few outcome-driven randomized evaluations.
Study Design and Search Strategy
This systematic review was conducted according to the PRISMA 2020 reporting framework [25], with literature-search reporting guided by PRISMA-S [26]. PubMed/MEDLINE, Embase, Scopus, Web of Science, and the Cochrane Library were searched for studies evaluating artificial intelligence in electrocardiographic analysis.
Search terms included combinations of “artificial intelligence,” “machine learning,” “deep learning,” “neural network,” “electrocardiography,” “ECG,” “arrhythmia,” “atrial fibrillation,” “myocardial infarction,” “heart failure,” “left ventricular dysfunction,” “cardiomyopathy,” “mortality,” “risk prediction,” and “wearable ECG.” Reference lists of eligible studies and relevant reviews were additionally screened.
Eligibility and Outcomes
Studies evaluating AI- or machine-learning-based analysis of 12-lead, single-lead, ambulatory, or wearable ECG recordings were eligible. Randomized trials, prospective and retrospective validation studies, diagnostic-accuracy studies, and prognostic-model studies were considered. Studies without clinical ECG data, animal studies, editorials, case reports, conference abstracts without sufficient data, and purely technical studies without clinically relevant validation were excluded.
Primary outcomes included sensitivity, specificity, area under the receiver-operating-characteristic curve (AUC), positive and negative predictive values, and discrimination for cardiovascular diagnosis or prediction. Secondary outcomes included calibration, external validation, mortality or cardiovascular-event prediction, clinical workflow outcomes, explainability, fairness, and patient-level clinical benefit.
Risk of Bias and Evidence Synthesis
Diagnostic-accuracy studies were evaluated using QUADAS-2 [27], while studies developing or validating prognostic and diagnostic prediction models were assessed using PROBAST [28]. Certainty of the principal diagnostic and clinical findings was considered using the GRADE framework where applicable [29]. Because of substantial heterogeneity in ECG modality, disease target, model architecture, reference standards, datasets, and performance measures, evidence was primarily synthesized according to clinical application: arrhythmia, myocardial ischemia/infarction, heart failure and ventricular dysfunction, cardiomyopathy, prognostic prediction, metabolic abnormalities, and wearable ECG analysis.
PRISMA Study Selection
A provisional search identified 1,742 records. After removal of 416 duplicates, 1,326 records underwent title and abstract screening. Of these, 1,109 were excluded. Full texts of 217 reports were assessed, and 151 were excluded, leaving 66 studies for inclusion.
Overall Evidence and Maturity of AI-ECG Applications
The included evidence demonstrated substantial expansion of artificial intelligence-enabled electrocardiography across arrhythmia classification, myocardial ischemia and infarction, left ventricular systolic dysfunction, heart failure, cardiomyopathy, mortality prediction, metabolic abnormalities, and wearable monitoring [30]. However, the maturity of evidence differed considerably between applications. Arrhythmia classification and detection of reduced left ventricular ejection fraction (LVEF) had the strongest combination of large datasets, external validation, and prospective evaluation, whereas HFpEF, cardiomyopathy, prognostic prediction, and metabolic applications remained predominantly retrospective.
Model performance was generally highest in internally validated datasets and more variable across external populations. Differences in disease prevalence, ECG acquisition systems, reference standards, patient demographics, and model architecture contributed substantially to between-study heterogeneity [31]. Consequently, high discrimination in a development cohort was not considered equivalent to demonstrated clinical utility.
Arrhythmia and Atrial Fibrillation Detection
Arrhythmia detection represented the most mature AI-ECG application. A recent meta-analysis of 20 studies encompassing more than 5.5 million ECGs reported a pooled sensitivity of 94.0% (95% CI 90.8–96.2), specificity of 98.7% (95% CI 97.3–99.3), and AUC of 0.982 for AI-based arrhythmia detection [32]. For atrial fibrillation specifically, pooled sensitivity was 92.6%, specificity 99.1%, and AUC 0.988. Importantly, performance remained high in the subset of externally validated models, with sensitivity of approximately 96.9%, specificity of 95.6%, and AUC of 0.983 [33]. Nevertheless, heterogeneity was substantial for both sensitivity and specificity, reflecting differences in datasets, arrhythmia definitions, model architectures, and validation procedures. AI applied to sinus-rhythm ECGs also demonstrated the ability to identify individuals with occult or future atrial fibrillation [34]. This application differs from direct rhythm classification because it functions as a risk-enrichment strategy rather than diagnosis of AF on the index ECG. Its potential clinical role is therefore primarily to identify patients who may benefit from extended ambulatory monitoring rather than to establish AF independently. Wearable and single-lead systems further extended AI-based rhythm detection into ambulatory settings [35]. However, signal artefact, lower disease prevalence, consumer-device variability, and false-positive alerts remain important implementation concerns.
Myocardial Infarction and Ischemia
AI-ECG demonstrated strong diagnostic performance for acute myocardial infarction, particularly ST-segment elevation myocardial infarction (STEMI). Meta-analytic evidence involving more than half a million ECGs reported pooled sensitivity and specificity of approximately 95% each, with an AUC of approximately 0.98 for STEMI detection [36]. Performance was less consistent for acute myocardial infarction considered more broadly. A diagnostic meta-analysis of 10 studies involving 94,510 participants reported pooled sensitivity of 89.4%, specificity of 96.0%, negative predictive value of 98.7%, and AUC of 0.97 [37]. However, important differences emerged by infarction phenotype. For STEMI, sensitivity and specificity were approximately 94.4% and 97.5%, respectively, whereas sensitivity for NSTEMI fell to approximately 65%, with an AUC of only 0.71.
Thus, aggregate AMI performance may overstate clinical capability if STEMI and NSTEMI are not analysed separately. Current evidence supports AI-ECG more strongly for rapid STEMI recognition and triage than for exclusion or diagnosis of NSTEMI.
Left Ventricular Systolic Dysfunction
Detection of reduced LVEF represented one of the strongest examples of AI-ECG identifying a clinically important phenotype not directly measurable from the ECG. Across diagnostic studies, AI-ECG demonstrated reproducible discrimination for heart failure or left ventricular systolic dysfunction [38]. A recent diagnostic meta-analysis including five studies and more than 100,000 participants reported pooled sensitivity of 0.86, specificity of 0.79, and an HSROC AUC of 0.89 for HF/LV systolic dysfunction. Restriction to studies at low risk of bias produced similar estimates, suggesting that the signal was not solely driven by poorer-quality studies.
More importantly, evidence has progressed beyond retrospective diagnostic accuracy. A meta-analysis of three randomized trials including 37,467 participants found that AI-ECG screening increased detection of previously unrecognized low ejection fraction by approximately 36% compared with usual care (RR 1.36, 95% CI 1.12–1.64) [39]. The number needed to screen was approximately 204. Notably, overall echocardiography referral rates did not significantly increase, although the positive predictive value of AI alerts remained modest at approximately 20–34%.
Prediction of Future Heart Failure
AI-ECG has also been evaluated for predicting incident heart failure rather than detecting existing ventricular dysfunction. A meta-analysis involving five studies, 11 cohorts, and approximately 1.73 million participants reported a pooled AUROC/C-statistic of 0.76 (95% CI 0.74–0.78) [40]. Although this represents moderate-to-good discrimination, heterogeneity was high (I² approximately 89%), and several studies had important risks of bias. Predictive performance was therefore materially weaker and less consistent than diagnostic performance for arrhythmia or reduced LVEF.
Heart Failure With Preserved Ejection Fraction
AI-ECG models for HFpEF demonstrated promising but substantially less mature evidence. A recent systematic review identified 10 studies involving more than 270,000 participants; seven studies contributed quantitative data from 11 independent cohorts [41]. The pooled AUROC was 0.84 (95% CI 0.78–0.88). Despite apparently good discrimination, between-study heterogeneity was extreme (I² = 100%). Several studies also had moderate-to-high risk of bias because of selective populations, inconsistent HFpEF reference standards, and incomplete reporting. No prospective outcome-based study had established that AI-ECG-guided HFpEF detection improves diagnosis, treatment, hospitalization, or mortality. Accordingly, AI-ECG for HFpEF should currently be regarded as an investigational screening or risk-enrichment approach rather than a validated standalone diagnostic strategy.
Cardiomyopathy and Structural Heart Disease
AI-ECG models demonstrated encouraging performance for hypertrophic cardiomyopathy, cardiac amyloidosis, and other structural myocardial diseases [42]. These algorithms appear capable of identifying subtle combinations of depolarization, repolarization, conduction, and waveform abnormalities that may not meet conventional diagnostic ECG criteria. Cardiac amyloidosis is a particularly attractive target because delayed diagnosis is common and early treatment may alter disease progression. AI-based screening models have identified patients with cardiac amyloidosis before conventional clinical recognition in retrospective cohorts [43,44]. However, the evidence remains less mature than for arrhythmia or reduced LVEF. Most structural-disease studies used retrospective case-control or enriched cohorts, and prospective screening studies evaluating downstream diagnostic yield and false-positive testing remain limited.
Electrolyte and Metabolic Abnormalities
AI-ECG demonstrated clinically relevant performance for detecting hyperkalemia and other potassium abnormalities [45]. Such applications may be particularly useful in patients with CKD, dialysis dependence, heart failure, or emergency presentations where rapid biochemical assessment is desirable. However, study populations, potassium thresholds, ECG modalities, and validation strategies varied considerably. Laboratory testing remains the diagnostic standard, and AI-ECG should therefore be interpreted as a screening or alert system rather than a replacement for serum electrolyte measurement.
Explainability, Bias, and External Validation
Explainability methods increasingly accompanied AI-ECG models, particularly saliency mapping, attention mechanisms, SHAP, LIME, and gradient-based approaches [46]. These methods frequently identified waveform regions that appeared physiologically plausible, including P waves, QRS complexes, and ST-T segments. However, most explainability approaches remained post hoc and were rarely validated for stability, fidelity, or clinician comprehension. A visually plausible heatmap does not establish that the algorithm has learned a biologically causal feature.
External validation also remained inconsistent. Although some large arrhythmia and LV dysfunction models retained strong performance in independent populations, many AI-ECG studies relied on single-centre datasets or random splits from the same underlying healthcare system. Differences in sex, age, ethnicity, comorbidities, devices, acquisition protocols, and disease prevalence may therefore reduce transportability.
Clinical Utility and Randomized Evidence
The largest gap in the evidence concerned whether AI-ECG improves clinically meaningful patient outcomes. Randomized evidence is currently strongest for increasing detection of reduced LVEF, where AI-ECG screening improved identification without substantially increasing overall echocardiography use. In contrast, broader randomized evidence across cardiovascular AI remains limited. Systematic reviews of randomized cardiovascular AI trials have found that many interventions improve diagnostic or process measures, but evidence for reductions in cardiovascular events, hospitalization, or mortality remains low or very low certainty [47]. This distinction materially changes the interpretation of the evidence. AI-ECG has demonstrated that it can identify patterns and diseases with high accuracy; it has not yet demonstrated consistently that deployment improves hard cardiovascular outcomes.
Table. AI-ECG Performance and Evidence Maturity
|
Application |
AUC/Effect |
Evidence |
Limitation |
|
Arrhythmia/AF |
0.98–0.99 |
High |
Heterogeneity |
|
STEMI/AMI |
0.97–0.98 |
High |
Validation |
|
NSTEMI |
0.71 |
Low |
Sensitivity |
|
Reduced LVEF |
0.89 |
High |
Low PPV |
|
HF prediction/HFpEF |
0.76–0.84 |
Emerging |
Bias |
|
Cardiomyopathy |
Promising |
Emerging |
Selection bias |
|
Metabolic/mortality |
Promising |
Emerging |
Validation |
|
Clinical outcomes |
Few RCTs |
Low |
Limited evidence |
Principal Findings
This systematic review demonstrates that artificial intelligence has substantially expanded the clinical capabilities of electrocardiography beyond conventional automated waveform interpretation. The strongest evidence currently supports AI-assisted arrhythmia classification, atrial fibrillation detection, STEMI recognition, and screening for reduced left ventricular systolic function, whereas applications involving HFpEF, cardiomyopathy, cardiac amyloidosis, electrolyte abnormalities, mortality prediction, and broader cardiovascular risk stratification remain at earlier stages of translation.
The findings also reveal an important distinction between technical performance and clinical maturity. Many AI-ECG models achieve high sensitivity, specificity, or area under the receiver-operating-characteristic curve in retrospective datasets, but fewer have undergone independent external validation, prospective implementation, or randomized evaluation. Contemporary reviews therefore argue that translational maturity should be judged not merely by discrimination, but by reference-standard quality, transportability, clinical actionability, workflow impact, and effects on patient outcomes [48]. Recent systematic evidence similarly indicates that arrhythmia detection remains the most mature domain, while structural-disease screening and prognostic modelling show greater variability in external validation and risk of bias [49].
4Why Arrhythmia Detection Has Progressed Furthest
Arrhythmia classification is particularly suitable for deep learning because rhythm disorders are fundamentally encoded within temporal and morphological ECG patterns. AI models can simultaneously evaluate rhythm regularity, atrial activity, QRS morphology, conduction intervals, and beat-to-beat variation without depending entirely on hand-engineered criteria. The very high pooled sensitivity and specificity reported for arrhythmia detection in the Results therefore represent biologically plausible performance rather than an unexpected algorithmic phenomenon. Large datasets containing millions of labelled ECGs have also facilitated model development and external validation.
However, high diagnostic performance does not eliminate implementation challenges. Class imbalance, rhythm prevalence, recording quality, and differences between 12-lead and ambulatory recordings may alter real-world performance. Algorithms trained on hospital-grade recordings may not reproduce the same accuracy when deployed through wearable or portable devices. Recent clinical reviews of wearable AI-ECG systems emphasize additional concerns involving device-specific preprocessing, user-generated recordings, signal artefact, and domain shift [50].
Reduced LVEF: The Strongest Example of Clinical Translation
Detection of reduced left ventricular ejection fraction represents one of the most important achievements of AI-enabled ECG because the model predicts a structural cardiac phenotype that cannot be directly measured from the surface ECG. The evidence is particularly compelling because this application has progressed beyond retrospective diagnostic accuracy. A recent meta-analysis of three randomized trials including 37,467 participants demonstrated that AI-ECG screening increased detection of previously unrecognized low ejection fraction by approximately 36%, with a number needed to screen of approximately 204, without significantly increasing overall echocardiography utilization [51]. This represents a higher level of translational evidence than that available for many AI-ECG applications. It demonstrates that algorithmic predictions can alter diagnostic yield within routine clinical practice.
Heart Failure Prediction and HFpEF
The evidence for predicting future heart failure is less mature than that for detecting existing ventricular dysfunction. The pooled discrimination observed in this review was moderate rather than exceptional, and heterogeneity remained high across studies. This is not surprising because incident heart failure is a complex clinical outcome influenced by age, blood pressure, coronary disease, diabetes, kidney function, obesity, valvular disease, and many other factors. An ECG-based signal alone cannot be expected to capture the entire causal pathway. HFpEF presents an even greater challenge. Although pooled discrimination appears promising, extreme between-study heterogeneity indicates that models may be learning population-specific features rather than a universally transportable HFpEF phenotype. Current translational assessments therefore place HFpEF AI-ECG primarily at the validation or risk-enrichment stage rather than routine clinical implementation [52].
A major research priority is determining whether AI-ECG can identify patients who would otherwise remain undiagnosed and whether subsequent echocardiography, natriuretic peptide testing, or clinical intervention improves meaningful outcomes.
AI-ECG as a Digital Biomarker Platform
A broader interpretation emerging from this review is that AI transforms the ECG from a conventional electrophysiological test into a potential digital biomarker platform. Deep neural networks can extract high-dimensional waveform information related not only to rhythm but also to ventricular function, myocardial disease, electrolyte disturbances, mortality risk, and future cardiovascular events. Recent conceptual work has similarly proposed that AI-ECG should be viewed as a digital biomarker spanning screening, diagnosis, and prognosis across the cardiovascular disease continuum [53]. This perspective helps explain why models can predict conditions that are not explicitly visible on conventional ECG interpretation. Electrical activity reflects structural, autonomic, metabolic, and hemodynamic properties of the heart, and subtle combinations of waveform features may encode information about underlying biological state.
However, the term “digital biomarker” should not imply biological causality. A model may identify a statistical signature associated with disease without revealing its physiological mechanism. Clinical use should therefore remain anchored to validated outcomes rather than presumed biological interpretation.
Prognostic Prediction: Strong Signal, Limited Actionability
AI-ECG models have demonstrated substantial potential for predicting mortality and major adverse cardiovascular events. Recent systematic evidence covering approximately 5.9 million patients reported C-statistics generally ranging from approximately 0.70 to 0.89 for mortality prediction [54]. Such results indicate that raw ECG waveforms contain prognostic information beyond conventional visual interpretation. Nevertheless, prognostic prediction poses a different clinical question from diagnosis.
An algorithm that predicts increased one-year mortality does not necessarily identify:
Performance may also decline when models are transported outside the development environment. Recent pooled work suggests that mortality prediction is materially attenuated in externally validated cohorts, reinforcing the need for local validation before deployment [55]. Therefore, prognostic AI-ECG should currently be considered a risk-stratification tool rather than a stand-alone therapeutic decision system.
Explainability and Clinical Trust
The “black-box” nature of deep learning remains an important barrier to clinician acceptance. Traditional ECG interpretation is physiologically interpretable: clinicians can identify atrial activity, conduction delay, ventricular hypertrophy, ischemic ST changes, or repolarization abnormalities. Deep-learning models may generate a probability without revealing an equally intuitive explanation. Explainable AI techniques attempt to address this limitation. A recent systematic review of 45 studies found increasing use of gradient-based methods, perturbation approaches, attention mechanisms, counterfactual explanations, and intrinsically interpretable architectures [56]. Methods that identify physiologically meaningful ECG intervals, including the P wave, QRS complex, and ST segment, appear more clinically intuitive than generic feature-importance maps. However, explainability methods remain fragmented and insufficiently validated.
Recent reviews therefore emphasize that explainability should be evaluated for fidelity, robustness, reproducibility, and clinician usefulness rather than visual plausibility alone [57].
External Validation, Domain Shift, and Generalizability
External validation emerged as one of the most important methodological weaknesses across the literature. Models frequently achieve outstanding performance using random train-test splits from the same institutional dataset, but such evaluation may underestimate the effect of domain shift. A recent systematic survey found that external validation was absent in approximately 72% of evaluated AI-ECG studies, while only about 38.6% demonstrated low overall risk of bias [58]. The same review found substantial geographic concentration of datasets in North America and Europe and limited demographic fairness assessment.
ECG-AI models may be sensitive to differences in:
An algorithm intended for widespread clinical deployment should therefore undergo validation in geographically and demographically independent populations using a frozen model and prespecified decision thresholds.
Wearables and Portable ECG Systems
Wearable and portable ECG systems represent an important opportunity for AI because continuous or repeated monitoring generates far more data than can be interpreted manually. Evidence is currently strongest for atrial fibrillation detection, whereas applications involving other arrhythmias, structural disease, and prognostic assessment remain less mature [59].
However, wearable deployment introduces new challenges. Consumer-generated recordings may be affected by motion artefact, inconsistent recording technique, skin contact, device-specific filtering, and incomplete lead information. An AI model trained primarily on conventional 12-lead ECGs should therefore not be assumed to perform equivalently on single-lead wearable recordings.
Clinical Utility Versus Diagnostic Accuracy
Perhaps the most important conclusion of this review is that diagnostic accuracy is not equivalent to clinical utility.
A model with an AUC of 0.98 may have little clinical value if:
Conversely, an algorithm with somewhat lower discrimination may have high clinical value if it identifies a treatable condition earlier and directs an effective intervention. Randomized evidence across cardiovascular AI remains limited. A 2026 systematic review and meta-analysis of randomized cardiovascular AI trials found potential benefits for some intermediate and process outcomes, but certainty of evidence generally ranged from low to very low, with limited evidence for hard cardiovascular outcomes [60,61]. Excessive alerts can produce alert fatigue, while poorly calibrated thresholds may increase unnecessary testing. Implementation should therefore consider not only model accuracy but also workload, cost-effectiveness, interpretability, medico-legal responsibility, cybersecurity, monitoring for model drift, and post-deployment performance.
Regulatory oversight is particularly important for adaptive or continuously updated algorithms because performance can change after deployment. Transparent model documentation, version control, and ongoing surveillance should therefore form part of routine AI governance [62-64]. AI-ECG should consequently be viewed not as a replacement for cardiologists or conventional electrocardiography, but as an additional layer of cardiovascular decision support. Widespread adoption should follow the same evidentiary principles applied to other diagnostic technologies: independent validation, prospective evaluation, transparent reporting, safety monitoring, and demonstration of meaningful patient benefit [65].
Artificial intelligence has substantially expanded the diagnostic and predictive potential of electrocardiography. The strongest current evidence supports AI-enabled ECG for arrhythmia detection, atrial fibrillation screening, STEMI recognition, and identification of reduced left ventricular systolic function. These applications demonstrate that AI can detect clinically relevant electrical patterns with high diagnostic accuracy and, in selected settings, improve disease detection in routine care. However, the maturity of evidence varies considerably across applications. HFpEF, cardiomyopathy, cardiac amyloidosis, electrolyte abnormalities, mortality prediction, and broader cardiovascular risk stratification remain promising but are supported predominantly by retrospective or heterogeneous studies. High sensitivity, specificity, or AUC should therefore not be interpreted as proof of clinical effectiveness. The principal barriers to widespread implementation include limited external validation, domain shift, inconsistent calibration, algorithmic bias, insufficient demographic representation, uncertain explainability, and limited randomized evidence demonstrating improvements in hard cardiovascular outcomes. Future AI-ECG research should prioritize multicentre external validation, prospective and randomized evaluation, standardized reporting, fairness assessment, clinician-centred explainability, and post-deployment monitoring.
Overall, AI-enabled ECG should currently be viewed as a clinician-supervised decision-support technology rather than an autonomous diagnostic replacement. Its long-term clinical value will depend not simply on how accurately algorithms classify ECGs, but on whether their predictions are generalizable, actionable, equitable, and capable of improving patient outcomes.