Artificial intelligence-enhanced electrocardiogram for arrhythmogenic right ventricular cardiomyopathy detection

Comprehensive Summary

Haq et al. trained and internally validated a convolutional neural network to detect arrhythmogenic right ventricular cardiomyopathy (ARVC) from standard 12-lead ECGs in 77 genetically confirmed patients (mean age: 47 years; 51% male), predominantly with PKP2 variants, compared with a large age- and sex-matched control cohort. ECGs were separated at the patient level into training, validation, and testing sets. In the held-out test cohort, the model achieved an AUC of approximately 0.75-0.76 with 77% sensitivity and 63% specificity. The positive predictive value was low (~3%), while the negative predictive value was high (~99%); however, this high NPV was strongly influenced by the low disease prevalence and 1:100 case-control structure. Performance declined when restricted to algebraically independent leads, suggesting the network leveraged redundancy across the full 12-lead signal. Overall, the model demonstrated modest discrimination with potential rule-out use in this cohort.

Outcomes and Implications

Given the diagnostic complexity of ARVC, where early ECG changes are subtle and imaging can be overinterpreted, an AI-enhanced ECG with high negative predictive value may serve as a first-pass screening tool within a stepwise evaluation strategy. However, its modest specificity, very low positive predictive value, and reliance on internal validation limit immediate clinical application. The findings support further development in larger, multicenter cohorts with external validation before integration alongside established Task Force criteria rather than as a replacement for them.