Artificial intelligence-enhanced electrocardiography to predict regurgitant valvular heart diseases: an international study

Yixiu Liang, Arunashis Sau, Boroumand Zeidaabadi, Joseph Barker, Konstantinos Patlatzoglou, Libor Pastika, Ewa Sieliwończyk, Zachary I. Whinnett, Nicholas S. Peters, Ziqing Yu, Xi Liu, Shuo Wang, Hongyang Lu, Daniel B. Kramer, Jonathan W. Waks, Yangang Su, Junbo Ge, Fu Siong NgEuropean 2025 · DOI

TL;DR

This study introduces AI-ECG models capable of predicting future moderate to severe regurgitant valvular heart diseases (rVHDs) like mitral, aortic, and tricuspid regurgitation. Validated on international, ethnically diverse cohorts, these models offer a promising tool for guiding echocardiographic surveillance and enabling earlier detection and intervention for at-risk patients.

What Problem Does This Paper Address?

Valvular heart disease (VHD) is a significant cause of morbidity and mortality, and early intervention can improve outcomes. However, identifying individuals at high risk of developing significant VHD before structural changes become apparent is challenging. Current clinical guidelines often recommend unselective echocardiographic surveillance for patients with mild VHD, straining healthcare resources. For individuals without any VHD signs, there are no clear recommendations for proactive monitoring. This paper aims to bridge this gap by developing AI models that can predict future VHD progression using only electrocardiogram (ECG) data, potentially enabling personalized and efficient patient management.

Methods

The researchers developed and evaluated two types of AI-ECG models: classification models for diagnosing existing VHD and survival models for predicting future VHD progression. Both models utilized a residual convolutional neural network (CNN) architecture. The survival models employed a discrete-time survival approach, which accounts for both the time to an event and censored data, making it suitable for predicting future outcomes. This approach was adapted to encode both prevalent and future disease states. For prediction, the models output a risk score, which was then used to stratify patients into risk quartiles.

Datasets

  • Derivation Cohort (Training & Internal Validation): Shanghai Zhongshan Hospital (SHZS) cohort, China. This dataset comprised 988,618 ECG and transthoracic echocardiogram (TTE) pairs from 400,882 patients.
  • External Validation Cohort: Beth Israel Deaconess Medical Center (BIDMC) cohort, Boston, USA. This dataset consisted of 34,214 patients with linked ECG and echocardiography data.

Key Insights and Ideas

  • ECG as a Prognostic Tool: The study strongly suggests that subtle ECG signatures can precede structural changes in VHD, making ECG a powerful tool for predicting future disease progression, not just diagnosing current conditions.
  • Survival Models Outperform Classification: The survival neural network models demonstrated superior performance for both diagnosis and prediction compared to standard classification models, likely due to their ability to leverage time-dependent data and censored outcomes more effectively.
  • Biological Plausibility: The AI-ECG predictions were found to correlate with subclinical chamber remodeling (e.g., left atrial and ventricular volumes), providing a biological basis for the AI's predictive capabilities. This suggests the AI is picking up on early signs of cardiac strain.
  • Generalizability Across Cohorts: The models, trained on a Chinese population, showed robust performance in an ethnically diverse US cohort, indicating good generalizability. This is a crucial step towards widespread clinical adoption.
  • Additive Value to Existing Predictors: When combined with traditional clinical and echocardiographic predictors, the AI-ECG models significantly improved the prediction of rVHD progression, particularly for severe TR.

Key Results

  • Prediction Accuracy: In the internal test set, AI-ECG survival models accurately predicted future significant MR (C-index 0.774), AR (0.691), and TR (0.793). These findings were confirmed in the external test set, albeit with a modest decrease in performance (MR C-index 0.698, AR 0.647, TR 0.702).
  • Risk Stratification: The highest risk quartile identified by the AI-ECG model had a significantly higher hazard ratio for future significant MR (HR 7.6), AR (HR 3.8), and TR (HR 9.9) compared to the lowest risk quartile.
  • Improved Prediction with AI: Combining AI-ECG with baseline clinical and echocardiographic data led to a significant improvement in prediction (Net Reclassification Index ranging from 0.299 to 0.630).
  • Explainability: ECG features associated with high-risk predictions included broader QRS complexes, left bundle branch block morphology, and specific T-wave abnormalities, which align with known indicators of cardiac strain and remodeling.

Limitations and Future Work

  • Retrospective Data & Selection Bias: The study used retrospective, hospital-based data, which may introduce patient selection bias. Future prospective studies with broader populations are needed.
  • Interobserver Variability: Echocardiographic assessments, the gold standard for VHD diagnosis, can have interobserver variability, which might have introduced minor inaccuracies.
  • Performance Disparity: The AI-ECG models showed slightly lower performance for predicting aortic regurgitation (AR) compared to mitral (MR) and tricuspid regurgitation (TR). This is attributed to AR's distinct pathophysiology and less direct electrical manifestation.
  • Clinical Utility Validation: While the models predict progression, their ultimate clinical utility in improving patient outcomes requires prospective evaluation in clinical trials.
  • Data Availability: The SHZS cohort data is not publicly available due to ethical limitations, but requests for access can be made. The underlying code is also not publicly available but can be shared under specific agreements.

The authors suggest that these AI-ECG models could be integrated into clinical workflows to guide surveillance, enabling more personalized and efficient management of patients at risk of developing significant valvular heart disease.

Summary updated

Abstract

This study developed AI-ECG models to diagnose and predict rVHDs and validated the models in a transnational and ethnically distinct cohort. The AI-ECG models could be utilized to guide surveillance echocardiography in patients at risk of future rVHDs, to facilitate early detection and intervention.
BibTeX
@inproceedings{liang2025artificial,
  author = {Yixiu Liang and Arunashis Sau and Boroumand Zeidaabadi and Joseph Barker and Konstantinos Patlatzoglou and Libor Pastika and Ewa Sieliwończyk and Zachary I. Whinnett and Nicholas S. Peters and Ziqing Yu and Xi Liu and Shuo Wang and Hongyang Lu and Daniel B. Kramer and Jonathan W. Waks and Yangang Su and Junbo Ge and Fu Siong Ng},
  title = {{Artificial intelligence-enhanced electrocardiography to predict regurgitant valvular heart diseases: an international study}},
  booktitle = {European},
  year = {2025},
  doi = {10.1093/eurheartj/ehaf448},
}
Figure