AI foundation model for pathogenic gene variant prediction in dilated cardiomyopathy
F El Kayouhi, L P A Arts, F D Raijmakers, M Kolk, S Ruiperez-Campillo, N Kalf, M Venner, L Alvarez-Florez, E Poel, J Verdonschot, E J Bekkers, F W Asselbergs, C R Bezzina, A S Amin, F V Y TjongAbstract
Background
Dilated cardiomyopathy (DCM) is a common heritable cardiomyopathy with substantial clinical and genetic heterogeneity (≈1 in 200). Determining whether a patient carries a likely pathogenic/pathogenic (LP/P) variant is clinically important since genotype status is associated with differences in disease course and arrhythmic risk, informing prognosis, follow-up intensity, device-therapy thresholds (including earlier ICD consideration in higher-risk genotypes), and facilitating familial screening. Current guidelines support genetic testing and, in resource-limited settings, the use of prioritisation tools such as the Madrid DCM Genotype Score. However, existing scores depend on detailed clinical variables (e.g., family history and specific phenotypic features) that may not be readily available or consistently captured in routine care, while artificial intelligence (AI) may capture electrocardiographic features that reflect genotype status. Therefore, a lower-barrier ECG-AI guided screening approach may facilitate earlier risk stratification and timely referral for genetic evaluation in DCM.
Objective
To develop and evaluate a foundational ECG-AI model for prediction of LP/P variant status in DCM, compared with the Madrid DCM Score.
Methods
Retrospective DCM cohort (probands and familial cases) with genetic testing. Outcome: any likely pathogenic or pathogenic (LP/P) variant. Input: earliest 10-second 12-lead ECG as sole input for the AI model. A pre-trained ECG foundation model (ECGFounder; trained on ~10M ECGs) was fine-tuned in a development cohort from hospital A (n=576) and evaluated in a balanced held-out test set (n=144). External validation used an independent cohort at hospital B (n=513; 19% LP/P) with a different LP/P gene distribution. Discrimination was assessed by AUROC/AUPRC and compared with the Madrid DCM score; a hybrid model combined the Madrid score with the AI-predicted LP/P probability.
Results
The ECG-only foundational model outperformed the Madrid DCM Score (AUROC 0.84 vs 0.76; AUPRC 0.84 vs 0.72). The hybrid model improved further (AUROC 0.87; AUPRC 0.89), an absolute gain of +0.11 AUROC and +0.17 AUPRC versus Madrid. These results suggest potential complementary value between clinical score–based risk stratification and ECG-derived AI predictions. In external validation (Hospital B), overall performance decreased but ranking was preserved; the hybrid model (AUROC 0.73) remained best over the ECG-Only (AUROC 0.70) and Madrid score (AUROC 0.67), despite clinically relevant differences in patient characteristics and gene distribution relative to the development cohort.
Conclusion
An ECG-only foundational AI model demonstrated improved discrimination for LP/P variant status compared to a clinical score and may enable lower-barrier genetic risk stratification. Combining ECG-AI with the Madrid DCM score further improved performance, suggesting incremental value where clinical variables are available.AUROC curve and other metrics for all methodsBaseline characteristics & top LP/P genes