Abstract Body (Do not enter title and authors here): Background
Artificial intelligence-enabled electrocardiography (AI-ECG) has emerged as a promising tool for detecting hypertrophic cardiomyopathy (HCM). While demonstrating high specificity (>99%), its moderate sensitivity (≈67%) raises concerns about its ability to identify individuals at increased risk for sudden cardiac death (SCD).

Hypothesis
We hypothesized that the algorithm’s diagnosis is most influenced by structural changes and scar burden, thereby increasing the likelihood of positively identifying patients with high SCD risk.

Methods
We retrospectively analyzed 12-lead ECGs of 230 patients with confirmed HCM, diagnosed per standard clinical criteria, using the Viz-HCM AI algorithm. Patients were categorized into AI-ECG positive (n=155) and AI-ECG negative (n=75) cohorts. Data were collected including septal phenotype, genotype variant, and clinical markers for high SCD risk as defined in AHA SCD prevention guidelines. A sensitivity analysis for algorithm positivity was conducted for high SCD risk features. Statistical comparisons using multivariable logistic regression and Wilcoxon rank-sum testing were performed to evaluate associations between patient characteristics and AI-ECG positivity.

Results
In AI-ECG positive patients, baseline algorithm sensitivity increased with presence of left ventricular aneurysm (90%), LGE ≥15% (84%), BNP >155 pg/mL (83%), and HCM Risk-SCD score ≥6% (74%). Genotype analysis conducted using a multivariable logistic regression model with covariate adjustment demonstrated that those with no pathogenic mutation had the strongest association with AI-ECG positivity (OR=27.9), followed by unknown genotype (OR=15.5) and pathogenic mutation (OR=7.7). Phenotypically, an apical septal morphology conferred the highest odds of AI-ECG positivity (OR=15.7) when compared to sigmoid. Incorporation of BNP into the model increased discrimination, raising AUC from 0.78 to 0.83. AI-ECG positive patients exhibited greater septal thickness compared to AI-negative patients (18.7mm vs. 16.9mm, p<0.001), and had higher BNP (303 pg/mL vs. 127 pg/mL, p=0.002).

Conclusions
AI-ECG positivity is most strongly predicted by features indicating advanced disease. The algorithm demonstrates preferential sensitivity for high-risk phenotypes, notably in those with structural changes known to be associated with adverse clinical outcomes. This likely contributes to the algorithm’s high likelihood of accurately diagnosing HCM in patients at high risk for SCD.