AI-Enhanced ECG Predicts Future Regurgitant Valvular Disease: Large International Study Shows Promise for Earlier Detection and Targeted Echocardiography

Highlight

• Artificial intelligence applied to standard 12‑lead ECGs predicted future moderate–severe regurgitant valvular heart disease (rVHD) — mitral (MR), tricuspid (TR) and aortic regurgitation (AR) — with good discrimination and large hazard gradients between risk quartiles.

• Models trained on nearly one million ECG–echocardiogram pairs (400,882 patients) in China maintained performance in a transnational external cohort from Beth Israel Deaconess Medical Center (34,214 patients).

• AI‑ECG predictions correlated with subclinical chamber remodelling on imaging, supporting biologic plausibility and potential use for guiding surveillance echocardiography.

Background and disease burden

Regurgitant valvular heart diseases (rVHDs) — mitral, tricuspid, and aortic regurgitation — contribute substantially to heart failure, arrhythmia, hospitalizations, and premature mortality. Timely recognition of progressive valve disease and ventricular or atrial remodelling is essential because targeted surveillance and timely intervention can alter outcomes. However, population‑level echocardiographic screening is resource-intensive and often impractical. Standard 12‑lead electrocardiography (ECG) is ubiquitous, inexpensive and routinely captured in many clinical settings. Recent work demonstrates that deep learning applied to ECGs can detect left ventricular dysfunction and other structural and rhythm abnormalities earlier than clinical recognition. This study by Liang et al. extends the paradigm to predicting future clinically significant rVHD, aiming to provide a low‑cost triage tool to guide echocardiography surveillance and early intervention planning.

Study design

The authors developed AI‑ECG models using a large, longitudinal paired database of ECGs and transthoracic echocardiograms assembled at Zhongshan Hospital (Shanghai): 988,618 paired ECG‑echo records from 400,882 unique patients. The modelling approach employed a residual convolutional neural network (CNN) architecture trained with a discrete‑time survival loss function, enabling time‑to‑event prediction rather than cross‑sectional classification alone.

External validation used a transnational cohort from Beth Israel Deaconess Medical Center (BIDMC), Boston — a secondary care outpatient data set of 34,214 patients with linked echocardiography. Primary outcomes were future development of moderate or severe MR, TR, and AR as defined by echocardiography. Performance metrics included the concordance (C‑index) for time‑to‑event discrimination and Cox proportional hazards models adjusted for age and sex to compare risk strata.

Key findings

Model discrimination and risk stratification

In the internal (Shanghai) test set the AI‑ECG models achieved the following discrimination for future moderate–severe rVHD:

• Mitral regurgitation (MR): C‑index 0.774 (95% CI 0.753–0.792)
• Aortic regurgitation (AR): C‑index 0.691 (95% CI 0.657–0.720)
• Tricuspid regurgitation (TR): C‑index 0.793 (95% CI 0.777–0.808)

When stratified by predicted risk quartiles, adjusted Cox models showed sharply increased hazards in the highest quartile versus the lowest:

• MR: hazard ratio (HR) 7.6 (95% CI 5.8–9.9; P < .0001)
• AR: HR 3.8 (95% CI 2.7–5.5)
• TR: HR 9.9 (95% CI 7.5–13.0)

These gradients indicate strong risk concentration in a subgroup flagged by the AI‑ECG models.

External (transnational) validation

Crucially, the authors report confirmation of these findings in the BIDMC cohort, an ethnically and geographically distinct outpatient population. Preserved discriminatory performance across populations supports robustness and potential generalizability of the approach beyond the original training environment.

Imaging associations and biological plausibility

Secondary analyses linked AI‑ECG predictions to imaging markers of subclinical chamber remodelling (for example, left atrial or ventricular enlargement or dysfunction), suggesting that the ECG‑based signal relates to early structural changes known to precede clinically overt valve dysfunction. This strengthens plausibility that the models detect real pathophysiological substrates, not merely dataset artefacts.

Safety and interpretability

The study focuses on prognostic discrimination and does not raise direct safety concerns about the ECG modality. However, the clinical deployment of predictive algorithms has downstream implications: false positives can increase echocardiography workload and anxiety; false negatives could delay surveillance. The authors used time‑to‑event modelling rather than simple binary labels, which supports more nuanced risk communication and scheduling of surveillance.

Expert commentary and context

These findings align with prior work demonstrating that deep learning can reveal structural heart disease signatures in standard ECGs. Notable examples include AI‑ECG detection of reduced left ventricular ejection fraction and prediction of arrhythmic risk. The current study extends those applications to valvular regurgitant lesions and, importantly, trains models to predict future events rather than contemporaneous disease only. This temporal element is essential if AI‑ECG is to be used for anticipatory surveillance.

Strengths of the study include the enormous paired dataset enabling learning of subtle ECG–echo relationships, use of a survival‑oriented loss function, and transnational validation. Demonstration of association with subclinical remodelling supports biological validity rather than overfitting to site‑specific practice patterns.

Key limitations and considerations:

• Population selection: Both development and test datasets are echocardiography‑linked cohorts. ECGs performed in contexts that prompt an echo may differ from general population ECGs. Performance in truly unselected community screening populations will require prospective evaluation.
• Case mix and prevalence: Prevalence of progressive rVHD and echo ordering thresholds differ across health systems, which affects positive predictive value and operational utility when transposed to new settings.
• Black‑box behaviour and calibration: Deep networks can be poorly calibrated; clinicians need clear calibration metrics, decision thresholds, and guidance on how AI scores map to surveillance intervals. Explainability methods can help but do not replace prospective validation.
• Clinical impact: Demonstrating improved early detection is an important first step; however, showing that AI‑guided surveillance reduces morbidity, delays progression, or improves patient‑centred outcomes requires prospective implementation trials or pragmatic randomized studies.

Clinical and implementation implications

Potential use cases for validated AI‑ECG rVHD prediction models include:

• Risk‑stratified surveillance: Prioritizing echocardiography for patients in higher AI‑predicted risk strata may efficiently detect progressive regurgitation earlier than routine practice.
• Primary care triage: In primary care or preoperative settings where access to echocardiography is limited, AI‑ECG could identify patients meriting expedited referral.
• Integration with electronic health records: Automated scoring of routine ECGs could trigger guideline‑aligned workflows (e.g., earlier echo scheduling when thresholds are exceeded), but such pathways require careful design to avoid overtesting.

Operational questions that need addressing before wide deployment include choice of actionable risk thresholds, follow‑up intervals, clinician notification strategies, and cost‑effectiveness — particularly the balance of earlier detection against additional imaging resource use.

Next steps and research priorities

To translate these models into practice, the following are priorities:

• Prospective validation in unselected ambulatory and primary care cohorts to measure real‑world performance and calibration.
• Randomized or pragmatic implementation studies evaluating whether AI‑guided surveillance reduces time to detection, changes management (e.g., timing of intervention), or improves clinical outcomes.
• Health‑economic analyses to determine the cost per case detected and downstream effects on resource utilization.
• Regulatory and equity assessments to ensure algorithms perform across diverse demographic and socioeconomic groups; transparency about training data and monitoring for performance drift is essential.
• Development of clinician‑facing decision aids that translate AI risk scores into concrete surveillance recommendations consistent with valve disease management guidelines.

Conclusion

Liang et al. present compelling evidence that AI applied to routine 12‑lead ECGs can predict future moderate to severe regurgitant valvular disease with good discrimination and large risk gradients. The combination of a very large training set, survival‑aware modelling, and transnational external validation strengthens confidence in the signal. If validated prospectively in unselected populations and shown to improve clinically meaningful outcomes and workflow efficiency, AI‑ECG could become a practical, low‑cost tool to prioritize echocardiography and identify patients for earlier intervention. For now, these models should be viewed as promising decision support tools that require prospective implementation studies and careful integration with clinical pathways and guideline recommendations.

Funding and clinicaltrials.gov

Funding and trial registration details are reported in the primary publication (Liang et al., Eur Heart J. 2025). Readers should refer to the source article for grant and registry information.

Key references

1. Liang Y, Sau A, Zeidaabadi B, et al. Artificial intelligence‑enhanced electrocardiography to predict regurgitant valvular heart diseases: an international study. Eur Heart J. 2025 Nov 21;46(44):4823–4837. doi:10.1093/eurheartj/ehaf448.

2. Attia ZI, Noseworthy PA, Lopez‑Jimenez F, et al. Screening for cardiac contractile dysfunction using an artificial intelligence–enabled electrocardiogram. Nat Med. 2019;25(1):70–74. doi:10.1038/s41591-018-0240-2.

3. Raghunath S, Feduscak S, Kwon D, et al. Deep neural networks can predict cardiovascular risk from electrocardiograms. Nat Med. 2020;26(9):1334–1339. doi:10.1038/s41591-020-0970-7.

4. Otto CM, Nishimura RA, Bonow RO, et al. 2020 ACC/AHA Guideline for the Management of Patients with Valvular Heart Disease. Circulation. 2020;142(24):e352–e454.