Automated detection of atrial fibrillation from the electrocardiogram channel of polysomnograms

Abstract

Purpose Accurate identification of atrial fibrillation episodes from polysomnograms is important for research purposes, but requires manual review of a large number of long electrocardiographic tracings. As automated assessment of these tracings for atrial fibrillation may improve efficiency, this study aimed to evaluate this approach in polysomnogram-derived electrocardiographic data. Methods A previously described algorithm to detect atrial fibrillation from single-lead electrocardiograms was applied to polysomnograms from a large epidemiologic study of obstructive sleep apnea in older men (Osteoporotic Fractures in Men [MrOS] Sleep Study). Atrial fibrillation status during each participant's PSG was determined by independent manual review. Models to predict atrial fibrillation status from a combination of algorithm output and clinical/polysomnographic characteristics were developed and their accuracy was evaluated using standard statistical techniques. Results Derivation and validation cohorts each consisted of 1395 individuals; 5% of each group had atrial fibrillation. Model parameters were optimized for the derivation cohort using the Akaike Information Criterion. Application to the validation cohort of these optimized models revealed high sensitivity (85-90%) and specificity (90-95%) as well as good predictive ability, as assessed by the C statistic (> 0.9) and generalized R2 values (∼ 0.6). Addition of cardiovascular or polysomnogram data to the models did not improve their performance. Conclusions In a research setting, automated detection of atrial fibrillation from polysomnogram-derived electrocardiographic signals appears feasible and agrees well with manual identification. Future studies can evaluate the utility of this technique as applied to clinical polysomnograms and ambulatory electrocardiographic monitoring.

Cardiac Tachyarrhythmia Detection by Poincaré Plot-Based Image Analysis

“…Poincaré plot analysis, describing each RR interval versus the previous one, has been widely used to classify different rhythms. Previous studies were based on the definition of parameters and thresholds describing particular patterns in the plot [4][5][6][7][8][9]. However, some patterns either cannot be easily described by simple parameters or cannot be intuitively appreciated.…”