Mudrakola, Harsha V.

Abstract Introduction Obstructive sleep apnea (OSA) is associated with atrial fibrillation (AF). Several features of OSA are thought to play role in the development of AF. A recently developed artificial intelligence (AI) enabled electrocardiogram (ECG) algorithm predicts the probability of future development of AF from a single ECG. We sought to examine the relationship between polysomnogram (PSG) and clinical features of patients with OSA and the probability of future AF (P-AF). Methods Consecutive adults (age ≥ 18) with OSA and with an ECG obtained within 2 years of their attended polysomnograms (PSG) were examined. We excluded patients with AF, left ventricular ejection fraction < 50%, implanted pacemaker or cardioverter defibrillator. We recorded demographics, apnea-hypopnea index (AHI), respiratory arousal threshold (RAT), percent sleep time with SpO2 < 89%, arousal index, and a comorbidity score composed of presence of coronary artery disease (CAD), hypertension (HTN), diabetes mellitus (DM), stroke, and chronic lung disease. One and two-way ANOVA was used to examine the relationship of OSA severity (mild, moderate, or severe) with P-AF. Multiple linear regression further characterized associations with P-AF. Results Mean AI determined P- AF was 0.052±0.012 (mean±SE) in mild, 0.064±0.014 in moderate, and 0.105 + 0.014 in severe OSA, demonstrating a significant association of OSA severity with P-AF (p = 0.01). Post-hoc pairwise tests showed a significant difference in P-AF between mild and severe (p = 0.008) but not between moderate and severe OSA (p = 0.083). Multiple linear regression with all the variables listed above showed the association between OSA severity and P-AF remains significant (P = 0.044). Age (p<0.001) and comorbidity score (p<0.001) were the only other variables significantly associated with P-AF. Conclusion OSA severity determined by AHI was significantly associated with P-AF as determined by a novel ECG-based AI algorithm independent of other variables. Given the strength of association with age and the comorbidity score, OSA severity may have been associated as it is known to increase with age and comorbid conditions. Future directions include incorporating AI enabled algorithms to identify individuals with the highest risk of AF in whom the role of OSA treatment in reducing risk of AF may be examined. Support (If Any)

Abstract Introduction The rules for classifying apneas as either obstructive or central using usual polysomnography (PSG) channels are well established, but classification of hypopneas is less straightforward without special sensors. Visual scoring methods have been proposed by the American Academy of Sleep Medicine (AASM) and by Randerath, et al. These two scoring methods have never been compared. We evaluated these two scoring methods for clinical use. Methods We selected 50 hypopnea segments from patient’s PSGs with very clear obstructive physiology (average total AHI 48.6, central AI 0), assumed to be obstructive hypopneas, and from patient’s PSGs with very clear central physiologies (average total AHI 34.3, obstructive AI 0.3), assumed to be central hypopneas. These 100 hypopnea-containing PSG segments (HCPS) were deidentified, placed in randomized order, and sent to two groups of 6 PSG scorers (2 RPSGTs, 2 sleep medicine fellows, 2 sleep medicine specialists). One group scored using the AASM criteria and the other used the Randerath algorithm. After a washout period, re-randomized HCPS were sent to be scored using the alternative method. We used Fleiss’ kappa to determine inter-rater reliability—i.e., how consistently multiple scorers came to the same conclusion about a given hypopnea segment using each method, and accuracy—how often the scorers rated the HCPS in a manner consistent with the assumed physiology. We also recorded the time it took to score. Results Overall accuracy of both methods was 68%. Among 12 scorers, Fleiss’ Kappa coefficient was 0.32 and 0.27 for the AASM and Randerath scoring methods, respectively. Average scoring time (24.3 minutes for AASM and 26.2 minutes for Randerath) was similar (p=0.79). Conclusion Inter-rater agreement was only fair using these methods, and accuracy was only 68%. More work is needed to discover a convenient, non-invasive way to reproducibly and accurately characterize hypopneas. Support This project does not have any funding support.