AHA Scientific Sessions 2026 Abstract · Nov 2, 2026
Alternative Structural Heart Disease Findings Among False Positive Electrocardiographic Foundation Model Predictions
By Eswar K, Ahmad S, Vavalle J et al.
Background: False-positive (FP) predictions of structural heart disease (SHD) from AI-ECG models remain a major challenge for screening deployment. We hypothesized that FP predictions reflect alternative SHD phenotypes sharing electrical remodeling signatures rather than random model error. Methods: We applied a transformer-based foundation model (ECG-FM) pretrained on 1.5 million ECGs to the EchoNext-Mini dataset (100,000 ECG-echocardiogram pairs; 36,286 patients). Eight endpoints were evaluated: aortic stenosis (AS), mitral regurgitation (MR), tricuspid regurgitation (TR), ≥ moderate right ventricular systolic dysfunction (RVSD), LVEF ≤35%, ≤40%, ≤45%, and left ventricular wall thickness ≥13 mm (LVWT). For each endpoint, frozen-embedding single-head logistic-regression probes (linear probes) were compared with partial fine-tuning of ECG-FM's top 3 of 12 transformer layers. We tested whether FP were enriched for alternative SHD by comparing FP with true negatives (TN) across 50 endpoint pairs after excluding nested LVEF-threshold comparisons, using age- and sex-adjusted logistic regression with patient-clustered robust standard errors. False discovery rate was controlled using Benjamini-Hochberg (BH) correction. Results: Fine-tuned model discrimination was moderate to high across endpoints (AUROC 0.755 [MR] to 0.873 [LVEF ≤ 40]). FP were enriched for alternative SHD in 48 of 50 endpoint pairs (BH q<0.05; adjusted ORs 1.41-6.41). The strongest enrichment was for LVEF ≤ 35% among FP for RVSD (unadjusted OR 6.77, adjusted OR 6.41, 95% CI 5.08-8.09). Stronger FP enrichment was seen for structural abnormalities that more commonly co-occurred clinically (Spearman ρ=0.67, p<0.001), supporting shared structural remodeling rather than random error. Across fine-tuned models, 50% - 70% of FP versus 19% - 39% of TN had ≥1 alternative structural finding (Figure 1). Linear probes achieved near equivalent partial fine-tuning discrimination (<2.5% AUROC differences) and reproduced the enrichment pattern, indicating that enrichment was present in pretrained ECG-FM representations rather than introduced by partial fine-tuning. Conclusions: False-positive ECG-FM predictions for structural heart disease were systematically enriched for alternative cardiac structural abnormalities, suggesting these models encode shared electrical signatures of structural remodeling. False-positive rates calculated against single-endpoint labels, while accurately reflecting endpoint-specific misclassification, may underestimate broader structural screening value.