Artificial Intelligence-Enabled Prediction of Heart Failure Risk From Single-Lead Electrocardiograms: A Multinational Cohort Review

Highlights

• AI algorithms trained on noise-adapted single-lead ECG data can robustly predict new-onset heart failure (HF) risk across diverse populations.
• Single-lead lead I ECG-derived AI models identify left ventricular systolic dysfunction (LVSD), a key precursor to HF, using wearable device-simulated ECG signals.
• AI-ECG significantly improves discrimination and risk reclassification compared to established clinical HF risk scores like PCP-HF and PREVENT.
• Multinational validation in the YNHHS, UK Biobank, and ELSA-Brasil cohorts confirms broad generalizability and potential for scalable community-based HF risk stratification.

Background

Heart failure represents a leading global cause of morbidity and mortality, with rising prevalence despite advances in treatment. Early identification of individuals at risk for developing HF is critical to enable timely intervention with evidence-based disease-modifying therapies. Traditional risk stratification relies heavily on clinical variables and comprehensive diagnostics, which limit scalability, especially in community or resource-limited settings.

Portable and wearable ECG devices recording single-lead signals present an attractive opportunity for large-scale, low-cost HF risk screening. However, raw single-lead ECG data are often noisy and less informative than 12-lead ECGs. Artificial intelligence (AI), particularly deep learning, offers tools to extract latent features predictive of subclinical cardiac dysfunction and future HF events even from low-fidelity data.

The recently published multinational retrospective and prospective cohort study by Dhingra et al. (2025) evaluates an AI-ECG model trained on noisy lead I ECGs to predict incident HF across three diverse populations, addressing the important clinical gap of accessible scalable HF risk stratification.

Key Content

Study Design and Cohorts

Dhingra et al. analyzed a large retrospective cohort from the Yale New Haven Health System (YNHHS) comprising 192,667 patients without prevalent HF, along with prospective population-based cohorts from the UK Biobank (UKB, n=42,141) and the Brazilian Longitudinal Study of Adult Health (ELSA-Brasil, n=13,454).

Baseline clinical characteristics spanned broad age ranges and mixed sex distribution (median ages 51–65; women 52%–58%), with long-term follow-up ranging from 3.1 to 4.6 years for HF hospitalization as the primary outcome.

AI-ECG Model Development and Deployment

The model was trained to detect left ventricular systolic dysfunction (LVSD) from lead I ECG segments adapted with simulated noise characteristic of wearable ECG devices. This noise adaptation aimed to reflect real-world signal quality from consumer-grade portable monitors.

The AI model output a continuous probability score reflecting LVSD risk, which was then correlated with new-onset HF events. The model was blind-tested in all three cohorts to evaluate predictive validity and generalizability.

Main Findings

– Incidence of new HF during follow-up was 1.9% in YNHHS, 0.1% in UKB, and 0.2% in ELSA-Brasil.
– A positive AI-ECG screening for LVSD conferred a 3- to 7-fold increased hazard of subsequent HF hospitalizations.
– Each 0.1 increment in AI model probability conferred a 27% to 65% hazard increase irrespective of age, sex, comorbidities, and competing death risk.
– Harrel C-statistics for incident HF prediction by AI-ECG ranged from 0.723 to 0.828 across cohorts, consistent with good discrimination.
– When added to conventional HF risk scores (PCP-HF and PREVENT), AI-ECG improved discrimination (increase in C-statistic 0.069 to 0.107), integrated discrimination improvement (0.068 to 0.205), and net reclassification improvement (11.8% to 47.5%).

Comparative Predictive Performance

The PCP-HF and PREVENT equations, validated clinical tools based on demographic and clinical risk factors, demonstrated moderate predictive ability. Inclusion of AI-ECG probabilistic outputs encompassing subtle electrophysiological markers of LVSD and preclinical dysfunction yielded statistically significant and clinically meaningful improvements in predicting HF onset.

Multinational and Multisystem Validation

Validation cohorts from different healthcare systems and geographic regions with varying patient demographics and HF incidence rates underscore the AI-ECG approach’s robustness and potential broad applicability, supporting future implementation in diverse clinical contexts.

Expert Commentary

The study by Dhingra et al. advances the paradigm of cardiovascular risk assessment by demonstrating the feasibility of AI-guided HF risk stratification from easily obtainable single-lead ECG signals. The approach aligns with precision medicine goals, enabling early identification of individuals with subclinical LVSD who may benefit from intensified surveillance or preventative therapy.

One of the key strengths is the noise adaptation technique, which realistically simulates the signal characteristics of wearable ECG monitors, enhancing model readiness for integration with consumer-grade digital health technologies. This paves the way toward democratized HF risk screening beyond specialized cardiology centers.

Nonetheless, several challenges remain before broad clinical deployment. Prospective studies using actual wearable ECG devices are necessary to corroborate model performance in real-world noisy data beyond retrospectively extracted signals. Integration workflows combining AI-ECG outputs with clinical decision-making require definition, including cost-effectiveness, patient adherence, and management pathways triggered by positive screenings.

Pathophysiologically, the AI likely captures subtle alterations in electrophysiological waveforms caused by impaired myocardial contractility and remodeling, which evade traditional ECG interpretation. Future research should explore the biological and electrophysiologic correlates of AI-predicted LVSD insight, potentially guiding novel therapeutic targets.

Moreover, while AI improved upon established risk scores, multimodal approaches incorporating biomarkers (e.g., natriuretic peptides), imaging, and clinical data might further enhance predictive power.

Conclusion

This comprehensive multinational cohort analysis demonstrates that a noise-adapted AI model analyzing single-lead lead I ECGs robustly predicts future HF risk and identifies LVSD in asymptomatic individuals. The AI-ECG model outperforms existing clinical risk tools and shows consistent discrimination across diverse populations.

These findings suggest a transformative strategy for scalable, community-based HF risk stratification leveraging wearable ECG technologies. To realize clinical impact, prospective validation, integration into care pathways, and assessment of patient outcomes following AI-ECG-based screening are critical next steps.

Overall, AI-enabled single-lead ECG analysis represents a promising avenue to bridge current gaps in early HF detection and personalized prevention, potentially reducing HF burden globally.

References

• Dhingra LS, Aminorroaya A, Pedroso AF, Khunte A, Sangha V, McIntyre D, Chow CK, Asselbergs FW, Brant LCC, Barreto SM, Ribeiro ALP, Krumholz HM, Oikonomou EK, Khera R. Artificial Intelligence-Enabled Prediction of Heart Failure Risk From Single-Lead Electrocardiograms. JAMA Cardiol. 2025 Jun 1;10(6):574-584. doi: 10.1001/jamacardio.2025.0492. PMID: 40238120; PMCID: PMC12004248.
• Jhund PS, MacIntyre K, Simpson CR, Lewsey JD, Stewart S, Redpath A, McMurray JJ. Long-term trends in first hospitalization for heart failure and subsequent survival between 1986 and 2003: A population study of 5.1 million people. Circulation. 2009 Aug;119(4):515-523. doi: 10.1161/CIRCULATIONAHA.108.816491.
• Weng SF, Reps J, Kai J, Garibaldi JM, Qureshi N. Can machine-learning improve cardiovascular risk prediction using routine clinical data? PLoS One. 2017;12(4):e0174944. doi: 10.1371/journal.pone.0174944.
• Berdichevskaia M, Bailey K, McCarthy CP, Gupta D, Qintar M, Bhatt DL, Mentz RJ. Machine learning for risk prediction in heart failure: A systematic review. JACC Heart Fail. 2022 Oct;10(10):748-757. doi:10.1016/j.jchf.2022.06.007.