Artificial intelligence and the evolution of the electrocardiogram: from cardiovascular diagnostic tool to digital biomarker

Abstract

The electrocardiogram (ECG) is a cornerstone of cardiovascular care. Traditionally, it has relied on expert visual interpretation and rule-based systems to define the presence of disease. However, the integration of artificial intelligence (AI) has transformed the ECG into a high-dimensional biomarker capable of detecting signatures of both overt and subclinical disease. This review explores the historical progress of the technology from its inception to its diverse range of AI applications in the clinic and in research. We examine fundamental methodological advancements, including a range of deep learning methods, and the use of ECG images and wearable and portable devices for scaling these innovations globally. We also provide the full spectrum of AI-enabled care via applications for electrocardiograms, including (i) assistance to clinicians to perform interpretation of ECGs, (ii) augmenting their ability to detect latent signatures of disease from ECG, and (iii) prognostic and predictive applications of AI-ECG in cardiovascular care. Finally, we address critical challenges regarding model transparency, phenotypic selectivity, and the gap in the development of AI-ECG applications and their actual implementation. To realize the full potential of AI for ECGs, the field needs to evolve from singular AI-ECG tools evaluated in retrospective studies toward robust foundation models with broader multimodal integration and evaluation in rigorously performed randomized clinical trials. By unlocking latent physiological data, AI-ECG serves as a scalable engine for cardiovascular precision care.

Graphical Abstract

Graphical Abstract.

Introduction

The electrocardiogram (ECG) is the foundation of cardiovascular diagnostics, recording a unique 12-lead digital fingerprint of an individual’s heart.¹ Today, more than 300 million ECGs are performed annually to diagnose a spectrum of cardiovascular diseases. Current clinical guidelines mandate its use as the first-line investigation for chest pain, arrhythmia, and syncope, cementing its primacy as a cardiovascular diagnostic tool.²

Despite this ubiquity, ECG interpretation has remained largely unchanged for nearly a century. Traditional ECG analyses are constrained by reliance on visual pattern recognition (e.g. ST-segment elevation).² While these approaches have been pivotal in detecting acute consequences of disease, they treat ECGs as a diagnostic tool with a few observable patterns rather than a continuous physiological signal. Consequently, most information contained in the high-fidelity waveform is not leveraged to inform clinical care, limiting the ability to use ECGs to define subclinical signatures of structural and metabolic disease that exist long before overt symptoms appear.

Artificial intelligence (AI) represents a paradigm shift in our ability to leverage ECGs in clinical settings. By treating the ECG not as a static image but as a rich data stream, deep neural networks can detect subclinical signatures of structural and metabolic disease, phenotypes previously invisible to the human eye. This review explores the evolution of ECG analysis from visual interpretation to a new era where the ECG serves as a computational biomarker for cardiovascular health.

Historical evolution of the ECG

While the electrical nature of cardiac function was known, the first human ECG was recorded in 1887 by British physiologist Augustus Waller using a capillary electrometer.³ Because these early recordings lacked the sensitivity required for clinical use, Dutch physician Dr Willem Einthoven engineered the string galvanometer in 1903 to produce a high-fidelity tracing.¹ With this device, Einthoven captured the heart's rapid voltage fluctuations and coined the P, Q, R, S, T nomenclature that remains the global convention today.^1,3 This work set the stage for the century of development that followed (Figure 1), which can be categorized into three phases. First, the analogue electrocardiography era, where innovations in technology expanded its spatial resolution; second, the computerized rule era of standardized interpretation; and third, the deep learning and AI era, which is currently redefining the boundaries of signal processing.

Figure 1.

Historical evolution and development of the ECG. Since its inception with Dr. Willem Einthoven, the modern ECG has experienced various technical improvements tailored to improving diagnostic tasks and widespread deployment. The recent entry into the deep learning era represents a novel frontier to expand the rich electrical signal into transforming cardiovascular care.

The analogue electrocardiography era

Early ECGs remained constrained by physical immobility and a limited, two-dimensional perspective. To tackle these challenges, in 1934, Dr Frank Wilson engineered the Wilson Central Terminal to improve the spatial resolution of the electrical signal.⁴ This topology was refined in 1942 by Dr Emanuel Goldberger, who modified the circuit to yield the augmented limb leads (aVR, aVL, aVF) and completed the modern 12-lead array.⁴ These key refinements in ECG development and deployment allowed cardiologists to systematically visualize the cardiac cycle and map specific pathologies, such as myocardial infarction or arrhythmias, to distinct visual waveform changes.³

The computerized rule era

As the volume of ECGs grew in the mid-20th century, interpretation remained inherently subjective. The first step toward a solution was the codification of human logic. In the early 1960s, Dr Cesar Caceres pioneered the transition from analogue inspection to computational analysis using deterministic ‘if-then’ rules to classify waveforms.⁵ By the 1970s, commercial systems began embedding these rule-based engines directly into ECGs, democratizing baseline diagnostics to the point of care. However, these systems were constrained by their own rigidity: they could measure strict voltage criteria with precision, but lacked the nuance to recognize complex, non-linear patterns, a limitation that persisted until the emergence of modern deep learning.

The deep learning and AI era

The transition from rule-based algorithms to modern AI algorithms represents a paradigm shift in signal processing.⁶ The earliest efforts deployed artificial neural networks (ANNs) to ECGs in the 1990s but were constrained by a reliance on feature engineering, a process in which human-defined metrics (e.g. QRS duration) were used as inputs to models.¹ These systems were limited to optimizing established clinical criteria without expanding the diagnostic range of clinicians.

The inflection point arrived with the use of deep learning, a set of complex neural networks, which can ingest high-fidelity data directly. These models have the capacity to detect complex morphological patterns. This technological leap began with raw voltage signals, capturing high-dimensional physiological information for latent phenotype discovery,⁶ and was further catalyzed by ECG images, which allow for scaling AI applications to global settings and diverse healthcare technology infrastructures.⁷

The evolutionary arc from Einthoven’s analogue traces to contemporary digital biobanks has fundamentally redefined the ECG. The advent of deep learning has transformed the ECG into a high-dimensional data substrate, unlocking latent physiological information that serves as the foundation for modern artificial intelligence.

Methodological considerations for AI applications for ECG

The many applications of AI-enhanced electrocardiography (AI-ECG) have stemmed from advancements in machine learning. This section reviews the fundamental methodologies used to extract clinical insights from the ECG, moving from basic supervision to complex representation learning (Figure 2).

Figure 2.

Methodological framework of AI-ECG. (Left) AI models utilize supervised, unsupervised, or self-supervised learning strategies depending on data labelling and clinical goals. (Middle) Core network architectures progress from standard artificial neural networks (ANNs) to convolutional neural networks (CNNs) and advanced hierarchical models like transformers. (Right) AI deployment scales across different data input modalities, ranging from high-fidelity 12-lead signals and device-agnostic ECG images to ambulatory single-lead wearables.

Machine learning involves training algorithms to identify patterns in data without explicit programming. Most current AI-ECG applications rely on supervised learning, where the model is trained on labelled data to map an input, such as the raw ECG waveform, to a specific output, such as left ventricular ejection fraction (LVEF). In contrast, unsupervised learning operates on unlabelled data, seeking to discover inherent structures or clusters within the signal itself. This approach can reveal sub-phenotypes of disease based purely on electrical morphology, independent of explicitly defined clinical subgroups. While traditional machine learning required manual input of extracted features (e.g. QRS duration), deep learning has revolutionized the field by learning feature representations directly from raw data.⁸

Core models: neural network architectures

The simplest form of deep learning is the ANN, which consists of fully connected layers of simpler computational units, called neurons. However, standard ANNs are not designed to capture the complex temporal dependencies of signals across an ECG while analyzing local patterns. For example, if a heartbeat shifts in time, an ANN perceives it as a completely new pattern. Consequently, the Convolutional Neural Network (CNN) emerged as the key model architecture of AI-ECG due to its ability to detect local features (e.g. ST segment changes) regardless of where they appear in the heartbeat. Modern iterations employ models with multiple layers that can capture non-linear relationships without degrading signal quality during the training process.⁹

Emerging models: encoders and contrastive learning

A major limitation of traditional supervised learning is the reliance on large datasets labelled with clinical conditions of interest. To address this, encoders and contrastive learning enabled the development of robust representations of cardiac physiology before a specific diagnostic task is assigned. These approaches leverage the inherent associations within the data to develop models that can then be further refined for specific tasks. The approach is called self-supervised learning. These methods effectively align orthogonal physiological signals by projecting heterogeneous data modalities (e.g. ECGs, echocardiograms, and clinical reports) into a unified latent space, a compressed mathematical representation where the model independently organizes complex data based on shared underlying features rather than explicit human labels. Consequently, the resulting embeddings capture holistic representations that generalize robustly to downstream tasks.^10,11

Data inputs: 12-lead signal vs. 12-lead image vs. single-lead ECGs

The input format of the ECG dictates both the model's architecture and its potential for scale. Most research utilizes signal data from the 12-lead ECG, which offers high fidelity and allows the model to detect minute features invisible to the human eye. However, signal data are often stored in proprietary formats within health systems and are not directly used in clinical care settings. Therefore, most development in this domain relies on building additional infrastructure to enable interoperability across hospital systems. Recent advances have provided an alternative by building models that interpret ECG images directly.⁶ While signal-based models offer device integration, image models enable broad scalability for global deployment, enabling advanced AI screening in resource-limited settings without specialized digital infrastructure.^7,12 Furthermore, continuous innovation in sensor hardware has synergized with advances in single-lead ECGs, transforming the rhythm strip into a comprehensive diagnostic tool. This technological development currently spans two distinct hardware categories: portable devices, which provide on-demand recordings similar to clinical event monitors, and wearable biosensors, which offer continuous passive surveillance of cardiac activity.^8,13 While there have been concurrent advances in utilizing higher-dimensional hardware, such as 100-lead ECG vests, their high technical variability and logistical barriers to deployment limit their immediate clinical scalability. Conversely, advancements in low-dimensional, but highly available hardware have scaled the potential of ECG diagnostics, learning to detect complex diseases such as left ventricular systolic dysfunction (LVSD), atrial fibrillation (AF), and structural heart diseases (SHDs) using only single-lead inputs.^14–17 By uncoupling diagnostics from the hospital setting, these tools pave the way for enhanced preventative monitoring, where asymptomatic dysfunction is identified and treated before it precipitates an acute admission. The different data modalities are summarized in Table 1.

Table 1.

Comparison of AI-ECG input modalities

Modality	Advantages	Limitations	Scalability	Representative Applications
Raw 12-lead digital signal	High temporal and amplitude fidelity; enables detection of subtle morphological features; direct waveform input for deep learning architectures including CNNs and transformers	Requires digital infrastructure and EHR interoperability; vendor-specific formats create data silos; limited utility in resource-constrained or legacy settings	Moderate. Well-suited for integrated health systems with digital ECG repositories; constrained by device and format heterogeneity	LVSD detection11; multilabel arrhythmia classification7; AF from sinus rhythm18; mortality prediction19 ; foundation model pre-training.11
ECG image (scanned or photographed 12-lead ECG)	Device-agnostic; compatible with legacy and paper records; enables AI deployment without raw signal access; facilitates global deployment via computer vision	Loss of raw voltage precision; susceptible to image artefacts (resolution, skew, lighting); variable layout across institutions requires preprocessing	High. Functions across diverse healthcare infrastructures; enables retrospective analysis of historical archives; impactful for low- and middle-income settings	LVSD screening20; HCM detection20,21; structural heart disease screening22,23; heart failure risk stratification24; cardio-oncology risk prediction.24
Single-lead wearable or portable ECG	Enables ambulatory and continuous monitoring; captures paroxysmal events missed by clinic-based snapshots; supports longitudinal physiological tracking; consumer devices lower access barriers	Reduced spatial resolution vs. 12-lead; susceptible to motion artefact and noise; models require adaptation from 12-lead training data; dependent on user adherence	High. At the population level via consumer wearables; equity concerns if restricted to commercial devices, image-based adaptations extend reach to lower-resource settings	AF detection16,2 526; single-lead LVSD screening15; decompensation surveillance in heart failure; population-level cardiovascular screening

AF, atrial fibrillation; CNN, convolutional neural network; ECG, electrocardiogram; EHR, electronic health record; HCM, hypertrophic cardiomyopathy; LVSD, left ventricular systolic dysfunction.

Table 2.

Model- and system-level challenges and limitations of AI-ECG

Domain	Domain	Description	Clinical consequence	Mitigation strategies
Model-level	Transparency & interpretability trade-offs	Many high-performing AI-ECG models are not readily explainable; some interpretability constraints can reduce performance.	Lower clinician trust; harder adjudication of discordant predictions; limited mechanistic insight.	Use purpose-fit explainability (e.g. saliency/attribution) as a debugging and communication aid; prospectively evaluate human-AI teamwork.
Model-level	False positives/low PPV in low-prevalence settings	Even high-specificity models can produce substantial false alarms when disease prevalence is low; some false positives may reflect preclinical risk signals.	Downstream testing burden; patient anxiety; alert overload.	Deploy in enriched-risk populations; staged screening (AI-ECG →confirmatory test); calibrate thresholds by setting/actionability.
Model-level	Phenotypic nonspecificity	Models trained for a single target can capture shared cardiovascular vulnerability and correlate with nontarget outcomes or related phenotypes.	Ambiguity in interpretation; cross-identification; unclear whether a positive output reflects the index lesion vs. broader risk.	Phenome-wide association and negative control testing; multimodal grounding (ECG/ECHO/EHR); consider reframing outputs as risk markers with defined downstream pathways.
System-level	Implementation	Retrospective performance does not guarantee improved outcomes since adoption and clinician behaviour are not captured in silico.	Limited real-world benefit; unintended harms; uncertain value proposition.	Pragmatic trials; measure downstream actions/outcomes; implementation science frameworks; define actionable outputs.
System-level	Equity in access	Training data and deployment platforms can underrepresent vulnerable groups, risking performance gaps and access barriers.	Widened disparities; unequal false positives/negatives; uneven adoption.	Subgroup audits; representative training/validation; device-agnostic pathways (including image-based ECG); deploy in safety-net settings; governance for fairness.
System-level	Care models	AI outputs that are not embedded into clear care pathways can create ambiguous tasks and excessive notifications	Alert fatigue, missed high-risk cases, increased workload, and variable adherence.	Tiered alerting; codesign with frontline users; define ownership and confirmatory testing pathways; monitor alert acceptance/override rates.
System-level	Regulations	AI models may evolve post-deployment; oversight is needed for versioning, retraining, and safety monitoring.	Risk of silent performance degradation; unclear accountability	Total product lifecycle approach; predetermined change control plans; post-market surveillance; clear labelling of intended use and limits.
System-level	Reimbursements	If payment mechanisms do not reflect clinical value, health systems may lack incentives to adopt and sustain AI-ECG programs. Furthermore, in nationalized models (e.g. European systems) where care is free to patients, high system-level procurement and IT installation costs remain a primary barrier to adoption.	Unsustainable deployment; limited spread beyond early adopters.	Budget impact and cost-effectiveness analyses; align with value-based programs; pursue reimbursement pathways for software/decision support; evaluate downstream resource use.

Clinical applications

The clinical maturation of deep learning has transformed the ECG from a traditional qualitative tracing into a quantitative digital biomarker, an objective, digitally measured indicator of disease risk (Figure 3).²⁵ Formally establishing AI-ECG as a digital biomarker requires navigating a rigorous validation framework: analytical validity to ensure technical robustness across diverse hardware, clinical validity to confirm accurate correlation with target phenotypes, and clinical utility to demonstrate improved, actionable patient outcomes. The following sections explore this spectrum of real-world utility, beginning with the technology's foundational role in clinician assistance and diagnostic automation.

Figure 3.

The interconnected clinical continuum of AI-ECG. The applications of AI-ECG function as an integrated ecosystem where foundational automation of routine diagnostic tasks in acute settings, such as the emergency department, provides the infrastructure for broader opportunistic screening. This screening capacity, increasingly driven by wearable and single-lead platforms, enables the identification of latent diseases and long-term prognostic risk stratification. Ultimately, these interconnected stages move the field toward new frontiers in biological discovery, transforming the standard tracing into a dynamic digital biomarker for population health.

AI for clinician assistance: diagnostic automation

The most immediate clinical application of deep learning focused on improving the workflow of busy clinicians. In this application, the primary objective has been to develop models that match human expert performance, alleviating the interpretation burden on clinicians and reducing the variability inherent in human reading.

Automated rhythm classification

Deep learning models have achieved parity with cardiologists in classifying cardiac arrhythmias. Among the first applications of AI to ECG, a deep neural network developed on single-lead ambulatory ECG monitoring data could identify 12 distinct rhythm classes (AF, atrial flutter, and AV blocks) with a diagnostic accuracy comparable to that of board-certified cardiologists.¹⁸ By automating the analysis of various arrhythmias, these models can dramatically reduce the time required to filter through thousands of ECGs and identify the early onset of cardiac pathologies.

Automated morphologic diagnosis

Beyond rhythm analysis, AI has been successfully scaled to interpret 12-lead ECGs.²⁶ A range of these models has been proposed, and demonstrated performance metrics that exceeded those of cardiologists in training and matched the accuracy of experienced specialists.⁷ This capability is now further enhanced by a new class of models that can generate free-text reports matching expert-level interpretations directly from ECG images.²⁷ These can serve as a highly reliable second reader, particularly impactful in settings where access to cardiologists is limited.

Point-of-care deployment and emergency triage

The emergency setting represents the highest-stakes application of diagnostic automation, where AI functions as a real-time triage tool. This represents a much-needed upgrade as current triage relies on computer reads and ad hoc manual screening that frequently yield incorrect clinical conclusions.^26,28,29 By instantly flagging critical ECGs, such as those with acute ST-elevation, AI algorithms can reprioritize the reading queue so that the highest-risk tracings reach attending physicians first, independent of arrival order. Crucially, AI can further expand this safety net by detecting latent signatures of acute myocardial infarctions even in the absence of ST-segment elevation. This has direct clinical consequences: AI-assisted triage has been associated with reduced door-to-balloon times for STEMI patients, linking algorithmic prioritization to improvements in time-sensitive outcomes.^28,30 However, these models require evaluation against more rigid standards than are available in routinely available clinical data.

AI for augmenting clinicians: detection of latent disease

While diagnostic automation aims to replicate human expertise, the greatest potential of AI-ECG lies in identifying pathology that is invisible to standard manual interpretation. Rather than relying on discrete interval thresholds or morphological abnormalities, deep learning models interrogate high-dimensional, non-linear patterns embedded within the voltage–time signal. In doing so, they enable detection of structural and electrophysiologic remodelling that may precede clinical recognition, conventional imaging abnormalities, or classical ECG criteria.

Structural and functional mechanical disease

One of the earliest and illustrative applications of this paradigm is LVSD detection. Traditionally, identification of reduced ejection fraction requires echocardiographic imaging. However, deep learning models trained on ECG-echocardiogram pairs have demonstrated that an ECG has enough latent information to identify patients with reduced left ventricular ejection fraction.³¹ This application effectively transforms the ubiquitous, low-cost ECG into a screening tool for subclinical heart failure, while also potentially serving as a gatekeeper to prioritize patients for more expensive and resource-intensive echocardiography. Beyond systolic dysfunction, similar approaches have been extended to other common structural abnormalities, including valvular heart disease and left ventricular hypertrophy. These applications reinforce a central biological principle: electrical remodelling reflects mechanical stress and structural change, often before abnormalities are clinically apparent or formally diagnosed. AI-ECG can function as a scalable triage tool by identifying individuals who warrant expedited echocardiographic evaluation and thereby optimizing allocation of imaging resources.

Rare hypertrophic and infiltrative diseases

Rare cardiac conditions, such as hypertrophic cardiomyopathy (HCM) and transthyretin amyloid cardiomyopathy (ATTR-CM), are frequently underdiagnosed due to their subtle early-stage presentation and low specificity of standard ECG criteria. Deep learning has emerged as a unique tool for learning subtle ECG signatures to identify rare diseases, successfully flagging sarcomere-positive patients for HCM even when their wall thickness was normal on echocardiography.^20,21 Similarly, AI models are reshaping the screening workflow for infiltrative diseases such as ATTR-CM, often using AI-ECG models as a pre-screen to significantly improve the positive predictive value of subsequent imaging and avoid indiscriminate imaging in low-prevalence populations.^32,33 In these rare-disease contexts, AI-ECG is best conceptualized not as a standalone diagnostic replacement, but as a probabilistic enrichment layer within staged care pathways.¹⁰

Occult rhythm and conduction disorders

Beyond mechanical disease, AI-ECG has shown promise in identifying concealed electrical substrates that may not meet conventional diagnostic thresholds. Certain channelopathies and conduction disorders are ‘latent’ not because electrical abnormalities are absent, but because they are subtle, intermittent, or expressed only under stress. Congenital long QT syndrome exemplifies this phenomenon. A substantial proportion of genotype-positive individuals have normal or borderline QTc intervals on resting ECG. Deep learning models have demonstrated the ability to distinguish patients with concealed LQTS from controls using features beyond simple QT measurement. This reframes the ECG from a tool that categorizes interval prolongation to one that probabilistically detects repolarization vulnerability, prioritizing patients for genetic testing or specialist evaluation.³⁴ By identifying subthreshold morphological signatures invisible to manual interpretation, these models suggest that electrical remodelling may be detectable even when classic criteria are not met.

Composite and multilabel SHD screening

Moving beyond single-disease detection, recent efforts have embraced multilabel architectures that simultaneously screen for multiple structural abnormalities from a single 12-lead ECG.^22,23 Rather than training isolated classifiers for LVSD, valvular disease, or hypertrophy independently, these models generate composite predictions that reflect broader patterns of mechanical remodelling. This multilabel approach has important conceptual implications. First, it aligns with real-world clinical practice, where structural abnormalities frequently coexist. Second, it suggests that AI-ECG may function less as a collection of disease-specific detectors and more as a generalized sensor of cardiovascular vulnerability. Electrical signatures of mechanical stress, fibrosis, and chamber remodelling may overlap across conditions, enabling the ECG to serve as a comprehensive, low-cost structural screening platform.

AI for prognostic evaluation: detection of future risk

Beyond identifying latent structural or electrical disease, AI-ECG has demonstrated the capacity to predict future clinical events. By capturing subclinical patterns of electrical remodelling, deep learning models can stratify patients according to their risk of adverse outcomes, shifting the role of the ECG from diagnostic assessment toward anticipatory risk modelling.

Prediction of heart failure risk

Prognostic applications have identified individuals at risk for developing HF before overt structural abnormalities appear on imaging. In studies of AI-ECG for LVSD, patients with a positive screen but without concurrent HF or LVSD were found to have a 4-fold increased risk of developing clinical heart failure in the future.³⁵ This suggests that these models detect subclinical signatures of LVSD that precede measurable dysfunction. Risk stratification for HF has been demonstrated across all ECG modalities, suggesting that this prognostic capability can be scaled across different resource settings.³⁵

AF from sinus rhythm

One of the earliest demonstrations of prognostic AI-ECG involved identifying patients with paroxysmal AF using ECGs recorded in sinus rhythm.³⁶ This suggests that the atrial electrical remodelling is detectable on ECGs independent of arrhythmia events. Clinically, such signatures may inform decisions regarding extended rhythm monitoring or earlier preventive strategies in high-risk individuals.

Sudden cardiac death (SCD) and mortality risk

Beyond arrhythmia prediction, AI models have demonstrated the capacity to stratify outcomes, such as risk for SCD and all-cause mortality. A deep neural network predicted 1-year all-cause mortality directly from ECG traces with an AUC of 0.85¹⁹ with high accuracy even in the subset of patients whose ECGs were interpreted as normal by physicians. This indicates that AI can detect subtle markers of frailty and dysfunction invisible to human readers.

Cancer therapy-related cardiac dysfunction (CTRCD)

In the field of cardio-oncology, AI-ECG is addressing the critical need to both predict and monitor cardiotoxicity related to chemotherapy. Models applied to ECG images have stratified patients receiving anthracyclines or trastuzumab according to subsequent risk of cardiotoxicity, identifying individuals with a markedly higher incidence of dysfunction before measurable declines in ejection fraction.²⁴ This offers a scalable, cost-effective surveillance strategy that could reduce the reliance on frequent echocardiograms during cancer treatment.³⁷

Other unique applications

Beyond standard diagnostics and risk prediction, deep learning is opening new frontiers in applying the ECG as a health data stream. These applications move away from simple classification tasks toward complex simulation, unsupervised discovery, and scalable global health solutions.

Unsupervised disease profiling

While most AI-ECG models are supervised, unsupervised deep learning allows the ECG to reveal new biological insights. By utilizing autoencoders to compress ECG data into a latent space, recent efforts have demonstrated that these unsupervised features correlated with over 1600 different diseases across the phenome.³⁸ This suggests that the ECG contains rich, systemic health information that human definitions have yet to categorize, allowing the data to give us insight about non-cardiac disease phenotypes.

Genetic discovery

Deep learning has altered the landscape of cardiac genetics by transforming the ECG into a rich phenotypic data stream. Traditional genome-wide association studies (GWAS) rely on binary disease labels (e.g. AF positive vs. negative). In contrast, AI models can quantify subtle, continuous electrical traits such as the probability of having AF, which serve as powerful phenotypes for genetic discovery. Recent work has demonstrated that conducting GWAS on these AI-derived phenotypes yields significantly more genetic hits than standard clinical labels.³⁹

Biological ageing (ECG age)

A unique extension of mortality prediction is the concept of ECG age, a biomarker of biological ageing.⁴⁰ Models trained to predict a patient's chronological age from their ECG often make errors. However, these errors are clinically significant. Patients whose predicted ECG Age exceeds their chronological age have a significantly higher hazard ratio for death and cardiovascular disease.⁴⁰ Conversely, having an ECG age younger than one's actual age is a marker of longevity. This serves as a strong, non-invasive surrogate of global cardiovascular health, potentially useful for tracking the efficacy of anti-ageing or cardioprotective therapies.

Challenges of AI-ECG

While AI-ECG offers a scalable and sensitive strategy to detect and prognosticate a range of diseases, the technology faces limitations that must be recognized to ensure optimal safety and efficacy prior to clinical adoption. This section explores the inherent tradeoffs in interpretability and phenotypic selectivity, and the practical barriers to effective clinical integration (Table 2).

A central tension in medical AI is the lack of model interpretability. To address this, researchers employ explainability methods, such as saliency maps (e.g. Grad-CAM) or attention mechanisms, to highlight parts of the ECG the model is using. However, the clinical applicability of these tools is limited because they frequently provide only technical interpretability (indicating where the model looked) rather than true clinical interpretability (explaining the underlying pathophysiology). For instance, a saliency map may highlight the ST segment, but it cannot articulate the medical logic of whether the model is detecting acute ischemia or benign early repolarization.

Bridging this gap to achieve true clinical transparency often comes at a quantitative price. A study that compared standard deep learning models with feature-constrained models demonstrated that constraining models to be interpretable resulted in a significant degradation in diagnostic performance compared to pure black box models.⁴¹ This suggests a cost of interpretability—the features that make AI robust are precisely the ones that are opaque to a human reader.

Critically, this opacity also presents a challenge to evaluating biological plausibility. In cases where AI models appear to predict non-cardiac or systemic conditions, the lack of interpretability often masks confounding. Rather than isolating disease-specific electrical signatures, these models frequently rely on latent proxies for chronological age, sex, or generalized frailty. Because overall vulnerability increases with age and frailty, the ECG may act as a non-specific marker of broad systemic sickness.⁴² Therefore, advancing explainability is not merely about building clinical trust, it is a scientific necessity to distinguish genuine electrophysiological pathology from generalized markers of ageing.

Even with high validation accuracy, clinical deployment of AI-ECG is hindered by the challenge of false positives, particularly when applied to rare diseases.⁴² In these settings, clinical utility is governed by disease prevalence, which dictates the mathematical balance between sensitivity and positive predictive value (PPV). Because PPV is inextricably linked to baseline prevalence, applying even a highly specific algorithm can yield a low PPV, generating a volume of alerts that outnumbers true cases. This false positive problem is a challenge in implementation as it can trigger a cascade of resource-intensive confirmatory testing (e.g. echocardiograms) for healthy individuals. However, increasing evidence suggests that many of these false positive flags represent the detection of subclinical or molecular-level changes that precede overt structural disease, as evidenced by future risk of disease.¹²

Furthermore, a fundamental challenge in evaluating AI-ECG is the field's historical overreliance on discrimination metrics, such as the AUROC. While AUROC demonstrates a model's ability to rank relative risk, it provides an incomplete picture of clinical utility. Real-world deployment is heavily dependent on calibration, the degree to which a model's predicted probability of disease matches the actual observed risk in the target population. Even with a high AUC, poor calibration can lead to significant over- or under-estimation of risk when models are deployed in novel clinical environments.

Beyond global accuracy, a growing challenge focuses on the phenotypic selectivity of AI-ECG models. A fundamental assumption in clinical AI is that a model trained to detect a specific pathology (e.g. LVSD) learns features unique to that condition. However, a recent phenomewide association study demonstrated that AI-ECG models trained for distinct SHD targets functioned less as condition-specific classifiers and more as broad risk markers. Specifically, the authors observed that the signatures of several distinct models were highly correlated. For example, a model trained to detect valve disease would cross-identify individuals with other SHDs and predict a broad range of future adverse events. This implies that deep learning identifies a shared cardiovascular identification of disease vulnerability rather than isolated anatomical defects.⁴³

Challenges & limitations of AI-ECG: clinical implementation

A formidable challenge in AI-ECG is the implementation gap between retrospective validation and real-world utility. High diagnostic accuracy in a dataset does not inherently translate to improved patient outcomes. To bridge this divide, the field must move beyond in-silico validation to rigorous prospective clinical studies. While many AI models describe robust performance, few have been subjected to randomized intervention studies to demonstrate improvements in changing clinical management. Retrospective data cannot account for whether clinicians will trust and adopt these models, how they will act on the information, and whether those actions yield a net benefit.

Furthermore, a critical ethical dimension of AI-ECG deployment lies in the risk of exacerbating existing healthcare disparities. Deep learning models are only as robust as the data on which they are trained, yet many foundational datasets, such as those from the UK Biobank or large academic medical centres, often lack racial, ethnic, and socioeconomic diversity. Consequently, models trained primarily on homogenous cohorts may fail to generalize to underrepresented populations, leading to variable performance and potential misdiagnosis in minority groups.⁸ Beyond algorithmic bias, there are challenges in hardware access and systemic implementation. The ultimate impact of AI-ECG on health equity will be heavily dictated by the healthcare systems in which it is deployed. In universally accessible or nationalized health systems, AI-ECG could serve as a scalable safety net, democratizing expertlevel screening to under-resourced community clinics. Conversely, in privatized systems, many AI-ECG tools are increasingly integrated into expensive consumer wearables. If these technologies remain accessible only to affluent populations, AI-ECG risks becoming a tool of privilege rather than a mechanism for population health. To mitigate these risks, concrete strategies must be enforced across the algorithm life cycle: prioritizing globally diverse cohorts during development, mandating local calibration to target populations during deployment, and requiring demographic-stratified performance audits prior to regulatory approval.

The transformative potential of AI-ECG cannot be realized without redesigning clinical workflows to accommodate the influx of new diagnostic data. Current healthcare systems are reactive and lack the bandwidth to manage continuous streams of risk data from asymptomatic individuals. Integrating AI-ECG screening into primary care or remote monitoring requires establishing downstream pathways, such as discernible flags that can be referred for confirmatory testing (e.g. echocardiography). Without structural guardrails, AI alerts may risk information overload for already overburdened care providers.

The long-term sustainability of AI-ECG hinges on the establishment of rigorous regulatory frameworks and viable reimbursement structures. Robust oversight is essential to safeguard patient welfare, ensuring that the rapid pace of innovation does not outstrip safety standards. Regulatory bodies, such as the FDA, must enforce strict validation criteria to guarantee that these tools are not only accurate but also robust across diverse clinical populations, preventing the deployment of algorithmic biases or unvalidated models.

However, even a safe and compliant tool will fail to achieve widespread adoption without a clear economic pathway. Currently, there is a disconnect between the clinical value of AI and the financial realities of healthcare systems. In traditional fee-for-service models, there is often no financial mechanism to bill for algorithmic risk stratification, meaning hospitals incur the cost of the technology without a corresponding revenue stream. For AI-ECG to become a sustainable standard of care, payers must align financial incentives with patient outcomes, potentially by introducing specific CPT codes for AI-derived interpretation or by shifting toward value-based care models that reward the early, cost-saving detection of asymptomatic disease.⁴⁴

The next generation of AI-ECG

The next era of AI-ECG will entail addressing the critical challenges and making this field more interpretable, specific, and actionable (Figure 4). As the field matures, the focus must shift from isolated algorithms to integrated, scalable systems that directly enhance patient care. The future of AI-ECG will be defined by innovations in technical development and deployment.

Figure 4.

The future landscape of AI-ECG can be conceptualized along three broad pathways. 1.) The development of robust tools to interpret the black box of AI-ECG architectures. 2.) The development of large-scale foundation models that integrate diverse and informative data modalities to improve generalizability and specificity. 3.) Rigorous investigation related to clinical implementation to ensure the ability of AI-ECG to improve clinical outcomes and patient care.

Current AI-ECG tools are narrow, developed via supervised learning to detect isolated pathologies. A paradigm shift is now underway toward foundation models, which leverage self-supervised learning on vast datasets to learn generalizable representations of cardiac physiology.⁴⁵ Several efforts have established this frontier, including ECG-FOUNDER, ECG-FM, and TARGET-AI,^10,45,46 though they employ different development strategies optimized for different tasks.

For instance, models like ECG-FM predominantly employ masked signal modelling, learning to reconstruct hidden segments of the waveform. This approach excels at capturing the intrinsic electrophysiological properties of the heart and offers a scalable, open-source architecture ideal for general feature extraction across diverse hardware. Conversely, approaches like TARGET-AI use multimodal contrastive learning, aligning the ECG signal directly with longitudinal EHR data. While this approach is limited by its reliance on a complex, highly integrated digital health infrastructure, it successfully grounds the electrical representations in a real clinical context. Understanding these architectural trade-offs, between intrinsic signal reconstruction and extrinsic clinical alignment, is critical, as they dictate whether a model is best suited for broad biological discovery or specific, targeted clinical deployment.

Despite these divergent architectures, these models share the goal of creating robust, generalized representations, offering a powerful mechanism to address and actively limit existing disparities. By training on massive, diverse, and unlabelled datasets, foundation models can overcome the biases inherent in smaller, human-annotated cohorts, such as demographic homogeneity, while simultaneously bypassing the subjective, error-prone human labels that limit traditional supervised learning. Future iterations should focus on greater multimodal integration, such as broadening the information leveraged by ECG, echocardiography, and clinical reports. This cross-modal synthesis aims to significantly reduce false positives by grounding electrical signals in anatomical ground truth, thereby enhancing both specificity and clinical utility.

Secondly, as portable and wearable ECG technologies mature, the application of AI-ECG will expand beyond the clinical encounter to individuals directly in their communities, allowing for population-level screening and longitudinal physiological tracking. In the context of screening, the growing availability of consumer wearables offers an unprecedented opportunity to detect occult pathology in asymptomatic populations.⁴⁷ Algorithms for these data will likely evolve from simple arrhythmia detectors into comprehensive digital biomarkers capable of identifying a range of cardiac disorders, using single-lead inputs acquired at home.²⁹ Simultaneously, longitudinal tracking will shift the paradigm from reactive management to proactive surveillance.³³ Rather than relying on sporadic snapshots taken during clinic visits, future AI models will ingest continuous data streams to construct personalized physiological baselines. Deviations from these patient-specific norms could serve as early warning systems for decompensation in heart failure or drug-induced adverse cardiac events. By anchoring diagnostics in the community, AI-ECG has the potential to improve the gap between detection and intervention.

Ultimately, the clinical impact of AI-ECG will not be determined by performance observed in historical ECGs, but in rigorous prospective deployment, confirming both their performance in real-world settings and their equitable integration into the standard of care.⁴⁸ To bridge the current implementation gap, the field must prioritize randomized trials that evaluate clinically meaningful endpoints.^49,50 Recent landmark trials have begun to establish this translational evidence base. For instance, the ARISE⁵⁰ trial demonstrated that real-time AI-ECG triage significantly reduced door-to-balloon times for myocardial infarction, while the pragmatic CARDIOLOGIST¹⁸ trial showed that AI alerts improved atrial fibrillation detection and oral anticoagulant adoption among non-cardiologists. Despite these well-conducted studies, the number of randomized trials for AI-ECG remains limited. While there is an emphasis on translating these technologies into clinical care, it is critical that large-scale randomized clinical trials that prioritize clinically meaningful endpoints are embedded in this translation.

Once clinical utility is established, the final frontier is integrating these tools into clinical guidelines. This would require robust evidence, along with a commitment to ensuring any endorsed tool is free of bias and broadly accessible. Specifically, before AI-ECG can be endorsed as a standard component of clinical care, each tool requires a rigorous audit of performance across diverse racial, ethnic, and socioeconomic subgroups. Moreover, ensuring equitable uptake means designing implementation frameworks that do not privilege only those with access to better hospitals or consumer devices, but are available broadly. Only through this dual commitment to rigorous evidence and broad accessibility can AI-ECG reach its potential as a transformative public health tool.

Conclusion

The evolution of the ECG from Einthoven's string galvanometer to a substrate for deep learning represents one of the most significant digital transformations in modern cardiology. By unlocking latent physiological information hidden within the waveform, AI has redefined the ECG from a static snapshot of rhythm to a comprehensive biomarker of structural health, biological ageing, and systemic risk. However, the transition from retrospective validation to prospective clinical utility remains the field's defining challenge. Realizing the full potential of AI-ECG will require understanding its role in the growing landscape of AI technologies through rigorous evaluation of novel multimodal care pathways. With continued progress, AI-ECG holds the promise of democratizing precision medicine, turning a century-old, ubiquitous test into a scalable engine for global cardiovascular health.

Author contributions

Ryan Choi (Conceptualization, Investigation, Writing—original draft, Writing—review & editing) and Rohan Khera (Conceptualization, Investigation, Writing—original draft, Writing—review & editing)

Funding

R.K. receives support from the National Institutes of Health (under awards R01AG089981, R01HL167858, and K23HL153775) and the Doris Duke Charitable Foundation (under award 2022060).

Data availability

No new data were created or analyzed in this study. Data sharing is not applicable to this article as it is a review of existing literature.