External Validation of Artificial Intelligence-enabled Electrocardiography (AI-ECG) for the Detection of Left Ventricular Dysfunction (LVD)

Data were collected from 13 hospitals, excluding the medical center that developed the artificial intelligence-enabled electrocardiography (AI-ECG) algorithm. The primary objective of the study was to evaluate the sensitivity and specificity of the AI-ECG model in detecting left ventricular dysfunction, defined as left ventricular ejection fraction (LVEF) ≤ 40%. To ensure clinical applicability, predefined thresholds required both sensitivity and specificity to exceed 0.80 in external validation cohorts. Sample size calculations were based on testing the null hypothesis that sensitivity equals 0.80. In the development hospital cohort, the model demonstrated a sensitivity of 0.869 and a specificity of 0.896. With a two-sided significance level (α) of 0.05 and a power of 90%, an estimated 310 cases of LVEF ≤ 40% were required.

Given that the prevalence of left ventricular dysfunction was 4% in the development hospital cohort but expected to be lower-between 2.5% and 3%-in external validation settings (i.e., regional and local hospitals), the total sample size needed to accrue the target number of cases was estimated to range between 10,333 and 12,400 patients. To achieve this, six regional hospitals and seven local hospitals were selected as external validation sites. Because both electrocardiography and echocardiography were required within a seven-day interval-leading to anticipated exclusions-approximately 1,500 patients were targeted from each regional hospital and 500 from each local hospital, resulting in a final target sample size of approximately 12,500 patients.