1. Introduction

Cardiovascular disease (CVD) remains one of the leading causes of death and disability worldwide, posing a substantial public health burden^[1,2]. Among its major manifestations, arrhythmias are particularly challenging because they often have silent onset, intermittent occurrence, and potentially rapid clinical deterioration, making continuous monitoring critical for timely detection and intervention^[3]. Although conventional 12-lead electrocardiography (ECG) and Holter monitoring are fundamental tools for cardiac assessment, their utility is limited by short monitoring duration, restricted portability, dependence on expert interpretation, and limited suitability for long-term out-of-hospital management^[4]. With the advent of smartwatches and ECG patches, wearable ECG monitoring has become an important approach for continuous CVD management^[5,6].

Recent advances in artificial intelligence, particularly artificial intelligence-enabled ECG (AI-ECG), are reshaping the conventional use of ECG^[7]. Driven by deep learning, AI-ECG not only improves the accuracy of arrhythmia diagnosis but also shows promise for early warning of high-risk cardiac events and risk assessment across multiple noncardiac systems^[8]. In wearable settings, these capabilities are transforming ECG from a signal-recording tool into a platform for continuous health sensing and risk prediction, creating new opportunities for out-of-hospital monitoring, chronic disease management, and personalized intervention^[8]. In this setting, model performance is tightly coupled to single- or few-lead acquisition, motion artifacts, electrode contact, sampling differences, and longitudinal individual variability, which makes wearable AI-ECG a distinct real-world modeling problem rather than a direct extension of hospital-based ECG analysis^[9].

As summarized in Figure 1, wearable AI-ECG monitoring can be organized as a workflow that begins with wearable ECG sensors, proceeds through data acquisition and mobile/cloud transmission, and then connects AI analysis with clinical tasks and output feedback. Wearable AI-ECG can support a closed-loop management framework in which ECG signals are acquired by wearable devices, transmitted to mobile or cloud-based systems, analyzed by AI algorithms, and returned to users or clinicians as risk estimates, alerts, or follow-up information. This human-in-the-loop workflow may improve long-term monitoring, triage, and individualized management, but also raises higher requirements for signal quality control, model reliability, and clinically actionable outputs.

Display Full Size

Download

Figure 1. Workflow of AI-assisted wearable ECG monitoring. The figure summarizes data acquisition, mobile or cloud platform, AI analysis, clinical tasks, and user or clinician feedback. ECG: electrocardiography; AI: artificial intelligence.

Unlike previous AI-ECG reviews that mainly focus on diagnostic algorithms or cardiovascular applications, this review emphasizes wearable monitoring, where device constraints, model design, clinical tasks, and validation requirements are tightly coupled. As wearable AI-ECG continues to evolve, its scope is extending from arrhythmia diagnosis to more predictive and integrative applications, including early warning of impending arrhythmic events and disease prediction beyond rhythm disorders. These developments make it valuable to review wearable AI-ECG within a unified framework that connects signal characteristics, analytical pipelines, and downstream clinical tasks. In this framework, arrhythmia diagnosis, arrhythmia early warning, and selected cross-system risk prediction represent a progression from event detection to longitudinal health assessment.

Against this background, this review focuses on the development of AI-ECG in wearable monitoring, from arrhythmia diagnosis to arrhythmia early warning and multi-disease prediction. We first summarize wearable ECG devices, data-collection workflow, signal characteristics, and the general analytical pipeline of wearable AI-ECG. We then review recent progress in three major application domains: arrhythmia diagnosis, early warning of impending arrhythmic events, and broader disease prediction beyond rhythm disorders. Here, multi-disease prediction refers to selected ECG-based assessments of cardiac structural and functional abnormalities, metabolic or electrolyte disorders, physiological aging and prognosis, and systemic risk indicators. Finally, we discuss the key barriers to clinical translation, including data limitations, generalizability, interpretability, and real-world deployment, and outline future directions for wearable AI-ECG research and practice. Particular attention is given to standardized datasets, cross-device and cross-population validation, prospective evaluation, foundation models, longitudinal risk modeling, and different reliability requirements for wellness monitoring and clinical decision support.

2. Overview of AI-ECG for Wearable Monitoring

2.1 Wearable ECG devices and data-collection workflow

ECG records cardiac electrical activity through body-surface electrodes, and conventional 12-lead ECG remains the clinical standard for evaluating rhythm disorders, conduction abnormalities, ischemia, and related cardiac conditions, as shown in Figure 2a. Wearable ECG monitoring is shifting ECG assessment from short-term in-hospital testing to continuous out-of-hospital monitoring. Wearable devices, including Holter monitors, ECG patches, smartwatches, chest straps, and flexible electronic systems^[10], emphasize portability, wearing comfort, and remote transmission, thereby changing ECG from a one-time clinical recording into a continuous real-world data source for AI-driven analysis^[11-13].

Display Full Size

Download

Figure 2. Clinical and wearable ECG acquisition. (a) Standard multi-lead clinical ECG; (b) Wearable ECG, where electrode contact, body motion, and device settings affect signal quality and AI analysis. ECG: electrocardiography; AI: artificial intelligence; BW: baseline Wander; EM: electrode motion; MA: motion artifact.

For AI development, the usefulness of wearable ECG data depends not only on recording volume but also on the acquisition workflow. Key factors include electrode position, skin-electrode contact, device fixation, sampling rate, filtering settings, recording duration, recording context, and symptom or activity logs. Because annotations may come from expert review, device-generated events, clinical reports, symptom-triggered recordings, or linked outcomes, wearable ECG datasets should clearly report device type, lead configuration, acquisition settings, population characteristics, labeling procedure, and signal-quality criteria.

2.2 Characteristics and key challenges of wearable ECG signals

Compared with standard clinical ECG, wearable ECG is characterized by limited lead information, stronger noise contamination, and greater temporal and inter-individual variability^[14,15]. Most wearable devices acquire single-lead or few-lead signals, which provide less spatial information than 12-lead ECG and therefore constrain some diagnostic tasks^[16,17]. In addition, wearable monitoring is performed in uncontrolled environments and is therefore highly vulnerable to multiple sources of signal contamination, as illustrated in Figure 2b, including baseline wander, motion artifacts, electromyographic interference, and other noise sources that may distort waveform morphology and compromise beat detection and interval measurement^[18-20]. As long-term recordings also reflect diverse physiological states and marked individual heterogeneity, wearable ECG analysis is fundamentally a real-world modeling problem involving noisy, dynamic, and heterogeneous signals.

The main wearable-specific constraints and their modeling implications are summarized in Table 1 below. This mapping provides the technical basis for later discussions of arrhythmia diagnosis, early warning, and cross-system risk prediction.

Common wearable artifacts include baseline wander related to respiration or posture change, electrode-motion artifacts caused by unstable skin-electrode contact, and muscle artifacts generated by skeletal muscle activity. Because these components may overlap with ECG waveforms in both time and frequency domains, they can distort P wave-QRS complex-T wave (P-QRS-T) morphology, interfere with R-peak detection and interval measurement, and increase false-positive or false-negative AI outputs.

2.3 General analytical pipeline of wearable AI-ECG

As shown in Figure 3, wearable AI-ECG generally involves signal acquisition, preprocessing and quality control, feature representation learning, downstream task modeling, and risk output. After ECG representation, the pipeline separates into two major application branches that correspond to the structure of this review: arrhythmia diagnosis, discussed in Section 3, and arrhythmia early warning or broader disease prediction, discussed in Section 4. This organization links wearable signal constraints to model design and then to clinically different evaluation requirements.

Display Full Size

Download

Figure 3. Analytical pipeline of wearable AI-ECG. ECG acquisition and quality control lead to ECG representation, task-specific modeling, and validation for diagnosis, early warning, or disease prediction. ECG: electrocardiography; AI-ECG: artificial intelligence-enabled ECG.

Preprocessing and quality control aim to make wearable signals usable under noisy conditions. Artifact-reduction approaches include conventional filtering and adaptive filtering^[19], wavelet or empirical-mode decomposition^[36], nonlocal-means and related signal-processing methods^[37], deep denoising models such as denoising autoencoders^[38], convolutional neural network (CNN)/recurrent neural network (RNN)-based networks^[39], generative adversarial network-based methods^[40], and diffusion-based models^[41]. Beyond denoising, artifact-aware AI systems may use noise augmentation, signal-quality gating, robust loss functions, missing-segment handling, and uncertainty estimation to reduce unsafe predictions from low-quality segments.

Modeling strategies can be organized by input representation. Feature-based models^[42-46] use extracted physiological parameters, including R-R intervals, heart rate variability (HRV), ECG intervals, waveform morphology, ST-T features, ectopic burden, frequency-domain HRV, and nonlinear indices, followed by conventional machine learning or shallow neural networks. These methods are interpretable and data-efficient but depend strongly on signal quality and accurate fiducial-point detection. Full-signal deep learning models instead learn from raw or minimally processed ECG segments. CNNs^[47] capture local waveform and beat-level morphology, RNNs and long short-term memory (LSTMs)^[16] model rhythm dynamics, and Transformer-based^[22], self-supervised^[48], or foundation-model^[49,50] approaches may improve long-range representation learning and transferability.

After representation learning, model objectives should be defined by the downstream clinical task. Arrhythmia diagnosis mainly requires reliable identification of rhythm events or heartbeat categories from noisy long-term recordings. In contrast, arrhythmia early warning and disease prediction require models to capture pre-event, longitudinal, or latent physiological patterns rather than only overt abnormalities. Across these branches, validation should include patient-level data separation, external and cross-device testing, calibration, and clinically interpretable outputs; a more detailed staged validation framework is discussed in Section 5.

3. AI-ECG for Arrhythmia Diagnosis

3.1 Clinical need for arrhythmia diagnosis in wearable settings

Arrhythmia diagnosis is the most direct and currently the most mature application of wearable AI-ECG. Its main value lies in overcoming the limited ability of static ECG and short-term monitoring to capture paroxysmal abnormalities. This is particularly important for conditions such as atrial fibrillation, premature atrial or ventricular contractions, and intermittent atrioventricular block, which are often brief or clinically subtle and therefore prone to missed diagnosis during routine short-duration ECG assessment^[51-53]. Among these applications, AF screening is the most representative because of its high prevalence, major association with stroke and heart failure, and strong suitability for long-term wearable monitoring. Large-scale studies such as the Apple Heart Study and related wearable AI investigations have demonstrated the feasibility of AF screening in both general and high-risk populations^[54].

Compared with many structural or ischemic ECG tasks, arrhythmia analysis is particularly compatible with wearable ECG because rhythm disorders are often expressed through temporal patterns, R-R irregularity, ectopic burden, and beat morphology, which can be captured by single- or few-lead long-term recordings. However, the clinical meaning of arrhythmia outputs differs across use cases. Consumer-facing screening mainly aims to identify possible abnormal rhythms for follow-up, whereas clinical decision support requires higher reliability, lower false-alarm burden, and clearer evidence for actionable intervention.

3.2 Main methods for arrhythmia diagnosis

Arrhythmia models can be broadly divided into feature-based and full-signal approaches. Feature-based methods use fiducial points and physiological parameters such as RR intervals, HRV, PR/QRS/QT measures, P-wave or QRS morphology, ST-T patterns, ectopic burden, and nonlinear rhythm indices, followed by rule-based classifiers, support vector machines (SVMs), random forests, gradient boosting, logistic regression, or shallow neural networks^[44-46]. These methods are interpretable and data-efficient, but their performance may degrade when wearable noise, poor electrode contact, or motion artifacts impair fiducial-point detection.

Full-signal deep learning models learn diagnostic representations directly from raw or minimally processed ECG and have achieved strong performance in ambulatory single-lead and large clinical 12-lead benchmarks^{[16,17,52,54-56]}. CNNs are commonly used to capture local waveform and beat-level morphology, RNNs or LSTMs model sequential rhythm dynamics, and Transformer or pretrained representation models may improve long-range context modeling and transferability. Representative studies are summarized in Table 2.

Together, these studies show that arrhythmia diagnosis is the most mature application of wearable AI-ECG. However, Table 2 also highlights a translational distinction: strong retrospective performance should be interpreted together with lead configuration, device type, annotation quality, and validation setting, especially when models are transferred from standard 12-lead ECG to noisier single- or few-lead wearable recordings.

3.3 Current challenges

Despite its expert-level performance, AI-ECG for arrhythmia diagnosis still faces several important limitations. Deep learning models’ performance often declines across devices, lead configurations, sampling settings, and patient populations, indicating limited generalizability^[23,30]. In real-world monitoring, noise, motion artifacts, and low-quality segments remain major sources of false positives and false negatives. More complex tasks, including fine-grained classification of supraventricular and ventricular arrhythmias, multilabel recognition, and rare-event detection in long recordings, also remain challenging because of class imbalance and annotation burden^[56]. In addition, although deep models frequently outperform conventional methods, limited interpretability continues to hinder clinical trust^[59-61].

For wearable deployment, these limitations should be translated into task-specific evaluation requirements. Beyond accuracy or area under the receiver operating characteristic curve (AUC), wearable arrhythmia systems should report signal-quality handling, patient-level validation, external and cross-device testing, false alarms per monitoring time, calibration, and clinician-interpretable evidence according to the intended use. These diagnostic challenges also define the boundary of event recognition: most current systems identify arrhythmias during or after their occurrence, whereas continuous wearable monitoring creates the opportunity to model pre-event risk states and broader longitudinal health signals, which motivates the frontier applications discussed in the next section.

4. Frontier Extensions of AI-ECG in Wearable Monitoring

4.1 Risk prediction: Expanding the capability boundary of AI-ECG

Whereas arrhythmia diagnosis focuses on detecting abnormalities after they occur, recent AI-ECG research has increasingly shifted toward risk prediction. Building on the diagnostic foundation discussed in Section 3, this shift asks whether long-term wearable ECG can reveal pre-event or latent physiological patterns before overt clinical abnormalities are recognized. Two directions have emerged: early warning of impending arrhythmic events and prediction of disease beyond rhythm disorders. Together, these developments mark a transition in wearable AI-ECG from passive detection to proactive risk assessment.

This shift reflects more than an expansion of task categories. It also indicates a change in analytical focus, from segment-level classification to longitudinal modeling, from overt waveform abnormalities to latent risk representations, and from single-disease detection to broader digital biomarker discovery^[62]. Because wearable devices support continuous monitoring in real-world settings, they are particularly well suited for these emerging applications. Accordingly, this section discusses two frontier directions: arrhythmia early warning, which aims to identify pre-event risk states, and selected cross-system prediction, which explores ECG-derived digital biomarkers beyond rhythm diagnosis.

4.2 AI-ECG for arrhythmia early warning

Arrhythmia early warning is based on the assumption that clinically overt arrhythmia may be preceded by a transitional pre-event state rather than occurring instantaneously. Early studies of paroxysmal atrial fibrillation (PAF) suggested that ECG may contain detectable prodromal information before onset^[63-65]. As shown in Figure 4, this transition may involve progressive or fluctuating changes in rhythm organization, ectopic activity, conduction, repolarization, and autonomic regulation.

Display Full Size

Download

Figure 4. Pre-event evolution before arrhythmia onset. A conceptual transition from stable rhythm to arrhythmia may involve subtle changes in rhythm, ectopy, conduction, repolarization, and autonomic modulation.

Long-term wearable ECG is therefore suited to early warning because weak and nonlinear pre-event patterns are difficult to identify by visual inspection but may be captured through continuous monitoring and AI-based modeling (Figure 5). In practice, an early-warning model must define the target event, observation window, pre-event window, prediction horizon, baseline or control window, and acceptable false-alarm burden^[66], because these definitions directly affect training, reported performance, and clinical usefulness.

Display Full Size

Download

Figure 5. Long-term ECG changes before malignant rhythm onset. Weak and nonstationary pre-event patterns may require continuous monitoring and AI-based analysis for detection. ECG: electrocardiography; AI: artificial intelligence.

For PAF, ventricular tachycardia, ventricular fibrillation, and related events, early-warning studies therefore aim to identify pre-event ECG changes before post-onset diagnosis is too late to guide timely intervention.

Existing methods can be divided into two broad groups. Feature-based studies usually extract R-R interval dynamics^[67], ectopic activity^[68], QT or repolarization indices^[69], morphology descriptors^[70], frequency-domain features^[71], or nonlinear dynamics^[72], and then use SVM, k-nearest neighbor, decision trees, XGBoost, Markov models, or related classifiers. These approaches are interpretable and data-efficient, but they depend on accurate beat detection and are sensitive to wearable noise and window definition. Full-signal or sequence-based approaches use CNNs, RNN/LSTMs, transfer learning, or self-supervised representations to learn warning patterns from raw or long-term ECG. These models may better capture nonlinear and long-range dependencies, but they require larger event datasets and careful prospective validation.

Table 3 summarizes representative patterns across early-warning studies rather than listing all reports individually. The key message is that prediction windows, feature sets, model families, and validation strategies vary substantially, making direct comparison difficult.

Recent deep learning studies have begun to revisit this problem by learning latent warning patterns directly from long-term ECG recordings. As shown in Figure 6a, using recordings longer than 20 hours, Fiorina et al.^[82] and Gavidia et al.^[83] applied convolutional neural networks to pre-event ECG and demonstrated the potential of deep models to identify prodromal signatures before arrhythmias. Related work further suggests that these warning patterns are nonlinear, exhibit long-term dependency, and jointly involve rhythm and morphological changes^[81], shown in Figure 6b. Nevertheless, this field remains less mature than arrhythmia diagnosis because of limited event samples, inconsistent pre-event definitions, inter-individual variability, motion-related false alarms, uncertain clinical actionability, and lack of prospective real-world testing.

Display Full Size

Download

Figure 6. Long-term ECG modeling for arrhythmia early warning. (a) Pre-event ECG can be used to estimate near-term arrhythmia risk; (b) Deep sequence models can learn nonlinear rhythm-morphology changes. ECG: electrocardiography; CNN: convolutional neural network; LSTM: long short-term memory.

4.3 AI-ECG for multi-disease prediction and cross-system health assessment

Another extension of AI-ECG is the prediction of selected diseases beyond rhythm abnormalities. Because ECG reflects conduction, repolarization, structural remodeling, autonomic regulation, metabolic status, and electrolyte balance, AI methods may uncover latent physiological information not readily accessible through visual interpretation. Recent AI-ECG digital biomarker platforms further reflect this shift from single-task diagnosis toward systematic ECG-derived phenotype exploration. Here, multi-disease prediction refers to selected ECG-based assessments of cardiac structural or functional abnormalities, metabolic and electrolyte disturbances, physiological aging or prognosis, and systemic risk indicators, rather than unrestricted diagnosis of all diseases.

Current studies include reduced left ventricular ejection fraction^[62], heart failure risk^[84], myocardial hypertrophy^[85], and ischemia-related states^[86-88]. These tasks are clinically attractive, but much of the supporting evidence still relies on standard clinical ECG rather than wearable single- or few-lead signals.

AI-ECG has also been explored for cross-system tasks, including hyperkalemia^[89], physiological age estimation^[90], and prediction of hospitalization or mortality risk^[35]. Some studies further use explainability methods to relate model outputs to candidate waveform or frequency-domain patterns. As summarized in Figure 7, these studies suggest that AI-ECG may serve as a digital biomarker platform, but the evidence strength differs markedly across tasks. Recent wearable digital medicine studies combining ECG with other biosignals, such as PPG, also indicate broader interest in physiological state assessment, although such multimodal applications require separate validation and are not the main focus of this ECG-centered review.

Display Full Size

Download

Figure 7. AI-ECG for cross-system prediction and interpretation. (a) AI-ECG may identify selected cardiac, metabolic, prognostic, and systemic risk signals; (b) Explainability methods can localize candidate ECG patterns but remain hypothesis-generating. ECG: electrocardiography; AI-ECG: artificial intelligence-enabled ECG.

However, this field remains less mature than arrhythmia diagnosis. Labels often rely on imaging, laboratory tests, follow-up data, or clinical endpoints, and it remains uncertain whether wearable single- or few-lead signals are sufficient for stable cross-disease prediction. These indirect physiological pathways also make interpretation and appropriate clinical use more challenging.

For wearable translation, cross-system AI-ECG should therefore be evaluated with stricter attention to label quality, lead configuration, external validation, calibration, and clinical actionability. These requirements are especially important because many cross-disease predictions depend on indirect ECG patterns rather than visually obvious diagnostic signs.

5. Challenges and Future Perspectives

5.1 Challenges

Despite the rapid expansion of AI-ECG in wearable monitoring, several challenges still limit its clinical translation.

First, data availability and annotation remain major bottlenecks. Although wearable devices can generate large volumes of continuous ECG data, high-quality datasets with standardized acquisition protocols and reliable clinical labels are still limited. This problem is especially pronounced in arrhythmia early warning and multi-disease prediction, where event samples are sparse, pre-event definitions are inconsistent, and labels often depend on additional examinations or long-term follow-up. For wearable ECG, dataset quality depends not only on sample size but also on lead configuration, sampling rate, electrode placement, recording context, signal-quality criteria, annotation source, and patient-level metadata. Without these details, model performance is difficult to compare or reproduce across studies.

Second, model generalizability remains insufficient. Wearable ECG data vary substantially across devices, lead configurations, sampling frequencies, wearing positions, and use scenarios. Patient populations also differ in age, comorbidity burden, and physiological status. Such heterogeneity can lead to marked performance degradation when models are transferred across datasets or deployed in practice. In addition, motion artifacts, noise, and low-quality segments remain common in wearable recordings and continue to challenge model stability. These factors make cross-device and cross-population generalization a central barrier for wearable AI-ECG.

Third, interpretability and clinical trustworthiness remain unresolved. Although deep learning models often outperform conventional methods, their decision logic is still not sufficiently transparent^[60,91]. In wearable monitoring, where outputs may influence long-term management and clinical decision-making, accuracy alone is not enough. This concern is even more critical in early warning and multi-disease prediction, where the underlying mechanisms are less direct and the consequences of uncertainty may be greater. Clinically useful explanations should therefore link model outputs to interpretable ECG evidence, signal quality, temporal context, and uncertainty, rather than relying only on post-hoc heatmaps.

Finally, real-world validation and deployment remain limited. Many current studies are still based on retrospective datasets or offline analyses and have not been tested in long-term, real-world monitoring workflows. This limitation is particularly relevant to ECG denoising and artifact-reduction studies, which are often evaluated by signal-level metrics rather than downstream clinical utility. Important translational issues include false-alarm burden, handling of low-quality segments, clinician review workload, privacy protection, regulatory approval, continual model updating, and integration with wearable hardware and clinical workflows.

5.2 Future perspectives

Looking ahead, progress in wearable AI-ECG will depend not only on better algorithms, but also on advances in data infrastructure, validation, robustness, personalized risk assessment, and clinically usable intelligent systems. The most actionable priorities are standardized wearable ECG datasets, unified validation frameworks, artifact- and uncertainty-aware modeling, foundation and longitudinal ECG models, and clinical workflow integration.

First, large-scale, multicenter, multi-device real-world datasets will be essential for improving model generalizability and enabling prospective validation. This is particularly important for early warning and multi-disease prediction, where event samples are sparse and labels may depend on follow-up, imaging, or laboratory tests. A practical dataset protocol should report device type, lead placement, sampling frequency, filtering, recording context, signal-quality rules, annotation workflow, outcome linkage, and patient-level train-test separation.

Second, validation should move from internal testing with patient-level separation to external, cross-device, cross-lead, subgroup, and prospective real-world evaluation. In wearable monitoring, reliability should be judged not only by sensitivity, specificity, AUC, and calibration, but also by false alarms per monitoring time, handling of low-quality segments, clinician review burden, and clinical actionability.

Third, future models should treat signal quality and uncertainty as part of the inference process rather than as only preprocessing issues. Signal-quality assessment, artifact-aware training, quality-gated prediction, calibration, and uncertainty output may reduce false alarms and prevent overconfident predictions from low-quality wearable segments.

Fourth, foundation and longitudinal ECG models may support more scalable and personalized wearable AI-ECG. Self-supervised pretraining, masked autoencoders, and ECG foundation models can reduce reliance on task-specific labels and improve transferable representation learning, while longitudinal modeling may help distinguish stable individual baselines from clinically meaningful temporal changes. These models should still be evaluated for cross-device transfer, calibration, failure modes, subgroup performance, and interpretability before clinical use, especially when transferred from standard 12-lead ECG to wearable single- or few-lead recordings.

Finally, clinical workflow integration will determine whether wearable AI-ECG outputs can be translated into actionable care. Deployment should define alert thresholds, escalation rules, explanation formats, privacy safeguards, and clinician review pathways. LLM-based agents may help summarize ECG findings and support interactive decision-making^[92], but their outputs should remain auditable and should support, rather than replace, clinician judgment.

Taken together, these priorities define a practical path from dataset construction to model development, validation, and deployment. They are summarized in Table 4 as an actionable research agenda for wearable AI-ECG.

6. Conclusion

Wearable AI-ECG has rapidly evolved from arrhythmia screening toward broader risk prediction and cross-system health assessment. This progression reflects a fundamental shift in ECG analysis, from detecting manifest abnormalities to uncovering latent physiological states and dynamic risk trajectories. In wearable settings, where continuous and real-world monitoring is feasible, these capabilities may substantially expand the clinical value of ECG. However, wearable AI-ECG is not simply an extension of hospital-based ECG analysis; its performance and clinical value are shaped by single- or few-lead acquisition, motion artifacts, device differences, individual variability, and long-term monitoring context.

At present, the maturity of wearable AI-ECG remains uneven across tasks: arrhythmia diagnosis is relatively established, whereas arrhythmia early warning and multi-disease prediction are promising but still limited by data scarcity, weak standardization, and incomplete clinical validation. Future progress will require advances not only in algorithms, but also in data, interpretability, deployment, and integration into clinical workflows. In particular, standardized wearable ECG datasets, artifact-aware modeling, cross-device and cross-population validation, prospective real-world evaluation, multimodal fusion, and clinically interpretable outputs will be essential for translation. If these barriers can be overcome, wearable AI-ECG may become a key platform for continuous, proactive, and personalized cardiovascular and systemic health management.

Acknowledgements

This manuscript was edited for language clarity using Grammarly. Qwen-Image (version 2.0) was used to generate the sub-figures of Figure 1 and Figure 7 illustrating the background of AI-ECG applications in cross-system diseases. The authors take full responsibility for the integrity, originality, and accuracy of the work.

Authors contribution

Li Z: Conceptualization, writing-original draft.

Tian Y, Jin Y: Writing-review & editing.

Liu C: Supervision, writing review & editing.

Conflicts of interest

Zhiyuan Li is a Youth Editorial Board of BME Horizon. The other authors declare no conflicts of interest.

Ethical approval

Not applicable.

Consent to participate

Not applicable.

Consent for publication

Not applicable.

Availability of data and materials

Not applicable.

Funding

This paper was partly supported by the National Natural Science Foundation of China (Grant No. 62406190), National Postdoctoral Program for Innovative Talents (Grant No. BX20230215), China Postdoctoral Science Foundation (Grant No. 2023M732219), Shanghai Municipal Science and Technology Major Project (Grant No. 2021SHZDZX0102).

Copyright

© The Author(s) 2026.