Evaluation of artificial intelligence-based electrocardiogram analysis tools in patients with hypertrophic cardiomyopathy

Background Artificial intelligence (AI)-based electrocardiogram (ECG) analysis tools have shown promise in detecting various cardiac conditions. However, their performance in specific patient populations, such as those with hypertrophic cardiomyopathy (HCM), remains incompletely characterized. Purpose To evaluate the performance of three AI-based ECG analysis tools in patients with confirmed HCM: (1) a tool calculating HCM probability, (2) a tool calculating structural heart disease (SHD) probability, and (3) a tool providing ECG-based diagnoses across multiple categories. Methods We analyzed digitized 12-lead ECGs from patients with confirmed HCM using three AI tools. We applied three validated AI tools—ECGVision-HCM for HCM probability, PRESENT-SHD for structural heart disease screening, and ECGDx for automated ECG diagnoses—using 12-lead ECGs. We assessed the distribution of AI-calculated probabilities and their associations with clinical parameters and evaluated agreement between AI-based and manually assigned ECG diagnoses using Cohen’s kappa. Results Despite all patients having confirmed HCM, the AI-calculated HCM probabilities showed a relatively uniform distribution (median 38.8% [IQR: 12.8–63.4%]), with only 41.2% and 12.5% of patients receiving a probability score >50% and >75%, respectively. HCM probabilities were significantly higher in patients with abnormal versus normal ECGs (p<0.001), and they were positively associated with markers of disease severity. SHD probabilities were generally higher (median 51.4% [IQR: 28.7–74.5%]), with 51.2% and 25% of patients receiving scores >50% and >75%, respectively. As illustrated in Figure 1, the distribution of predicted probabilities was skewed toward lower values for HCM probability, particularly among patients with visually normal ECGs. In contrast, SHD probabilities were more broadly distributed across the cohort. The AI tool provided specific ECG diagnoses in only 17.7% of cases, with most patients classified as having no diagnostic label. Agreement between AI-based and manually assigned ECG diagnoses varied across diagnostic categories (kappa range: 0.08–0.52), with the highest agreement observed for LBBB, RBBB, and AF (κ=0.52, 0.45, and 0.40) and the lowest for AV block (κ=0.08). Figure 2 summarizes the baseline clinical and ECG characteristics of the study population, along with AI-derived probability scores and their correlations with clinical markers. Conclusions AI-based ECG analysis tools demonstrated modest performance in our HCM cohort. These findings highlight the challenges of applying AI tools developed in general populations to specific disease cohorts and underscore the need for disease-specific validation before clinical implementation.