Objectives: To assess the prevalence of common arrhythmias and to compare the automated analysis of an artificial intelligence (AI) model with physician interpretations on 24-hour ambulatory electrocardiography in Con Dao special zone.
Subjects and methods: A cross-sectional descriptive study was conducted on 187 outpatients at the military-civilian Medical Center of Con Dao from October 2025 to April 2026. Twenty-four-hour ambulatory electrocardiography was recorded using a 3-lead OctoBeat wearable device integrated with AI. The AI model was based on a ResNet architecture and validated on the MIT-BIH, AFDB, and ESC datasets. 5 binary outputs were evaluated: sinus rhythm, atrial fibrillation, premature atrial complexes, premature ventricular complexes, and sinus pause/asystole. Physician interpretation was used as the reference comparator. Performance metrics included sensitivity, specificity, positive predictive value, negative predictive value, accuracy, and McNemar test.
Results: The mean age was 51.4 ± 18.7 years, with a predominance of males (64.2%). Premature atrial complexes and premature ventricular complexes were the most common arrhythmias, accounting for 79.7% and 58.3%, respectively. Atrial fibrillation was observed in 4.8% and sinus pause/asystole in 5.9%. The AI model demonstrated high sensitivity across most outputs but low specificity, particularly for premature atrial and ventricular complexes. AI yielded higher positivity rates than physicians for all outputs (p < 0.001), except for sinus rhythm (p = 0.062).
Conclusions: AI integrated into wearable devices shows potential for arrhythmia screening in resource-limited settings. However, due to low specificity and a substantial false-positive rate, AI findings should be confirmed by physicians, and further calibration and validation are required before clinical implementation.