Detecting Structural Heart Disease from Electrocardiograms via a Generalized Additive Model of Interpretable Foundation-Model Predictors

Researchers have developed an interpretable AI framework that detects structural heart disease from electrocardiograms using a generalized additive model with ECG foundation model predictors. The system achieved a +0.98% improvement in AUROC and +1.01% in AUPRC over state-of-the-art deep learning baselines on the EchoNext benchmark of 80,000+ paired ECG-echocardiogram records. This transparent approach addresses clinical trust barriers by attributing risk to specific diagnostic features while maintaining performance with only 30% of training data.

Detecting Structural Heart Disease from Electrocardiograms via a Generalized Additive Model of Interpretable Foundation-Model Predictors

New AI Framework Makes Heart Disease Screening from ECGs More Transparent and Effective

A novel artificial intelligence framework that enhances the interpretability and performance of screening for structural heart disease (SHD) from electrocardiograms (ECGs) has been introduced by researchers. The method integrates clinically meaningful predictors from an ECG foundation model into a generalized additive model (GAM), creating a transparent system that attributes risk to specific diagnostic features. This addresses a critical limitation of existing "black-box" AI models and offers a scalable, cost-effective alternative to echocardiography for early SHD detection.

Structural heart disease, involving abnormalities in the heart's valves, walls, or chambers, is a widespread but often undiagnosed condition. While echocardiography (ECHO) is the gold standard for diagnosis, its high cost and limited accessibility hinder widespread screening. Recent advances have shown that AI can analyze ubiquitous ECG data to identify SHD, but the lack of interpretability in these complex deep-learning models has been a significant barrier to clinical trust and adoption.

Bridging Statistical Modeling and Modern AI

The proposed framework represents a complementary paradigm between classical statistical modeling and modern AI. Instead of treating the AI as an opaque predictor, the method uses a pre-trained ECG foundation model to generate a rich set of clinically relevant features. These features are then fed into a GAM, a type of interpretable model where the contribution of each input to the final risk score can be visualized and understood.

This design allows clinicians to see not just a prediction, but *why* the model made it. The estimated functions within the GAM provide interpretable insights into the relationships between risks of traditional ECG diagnoses—like left ventricular hypertrophy or atrial fibrillation—and the presence of underlying structural heart disease.

Superior Performance on a Large-Scale Benchmark

The team rigorously validated their approach using the EchoNext benchmark, a dataset comprising over 80,000 paired ECG and echocardiogram records. The results were compelling. The interpretable framework achieved a relative improvement of +0.98% in AUROC (Area Under the Receiver Operating Characteristic curve), +1.01% in AUPRC (Area Under the Precision-Recall Curve), and +1.41% in F1 score over the latest state-of-the-art deep learning baseline.

Remarkably, the model maintained slightly better performance even when trained on only 30% of the available data, demonstrating data efficiency. Furthermore, subgroup analyses confirmed the model's robust performance across heterogeneous patient populations, a crucial factor for equitable clinical deployment.

Why This Matters for Clinical Cardiology

  • Enhanced Trust and Adoption: By providing transparent risk attribution, this "glass-box" AI model builds the interpretability necessary for clinicians to trust and act on its predictions, paving the way for real-world clinical integration.
  • Scalable Screening Solution: It leverages low-cost, widely available ECGs to flag patients at high risk for SHD, who can then be referred for confirmatory echocardiography. This could significantly improve early detection rates.
  • Actionable Clinical Insights: The model doesn't just output a score; it highlights which specific ECG abnormalities are driving the risk, offering cardiologists actionable diagnostic clues and a deeper understanding of the patient's condition.
  • Efficient Use of Data: Its strong performance with limited training data suggests it can be effectively deployed even in settings with smaller annotated datasets, increasing its potential utility globally.

This research, detailed in the preprint arXiv:2603.02616v1, illustrates a powerful pathway to developing AI tools that are not only high-performing but also interpretable and clinically actionable. By successfully marrying the predictive power of foundation models with the transparency of statistical frameworks, it sets a new standard for responsible AI in cardiovascular medicine.

常见问题