0932 Using Novel EEG Phenotypes and Artificial Intelligence to Estimate OSA Severity. (12th April 2019)

• Record Type: Journal Article
• Authors: Fernandez, Chris; Rusk, Sam; Glattard, Nick; Piper, David; Solis, Jonathan; Hensen, Brock; Orr, Nick; Tekchandani, Jatin; Shokoueinejad, Mehdi; Hungerford, James
• Is Part Of: Sleep. Volume 42(2019)Supplement 1
• Page Start: A375
• Page End: A375
• Publication Date: 2019-04-12
• Subjects: Sleep -- Physiological aspects -- Periodicals; Sleep disorders -- Periodicals; Sommeil -- Aspect physiologique -- Périodiques; Sommeil, Troubles du -- Périodiques; Sleep disorders; Sleep -- Physiological aspects; Sleep -- physiological aspects; Sleep Wake Disorders; Psychophysiology; Electronic journals; Periodicals; 616.8498
• DOI: 10.1093/sleep/zsz067.930
• Languages: English
• ISSNs: 0161-8105

Abstract

Abstract: Introduction: EEG studies are widely used for monitoring and diagnosis of neurological conditions including epilepsy, seizure disorders, among others. Ambulatory EEG, EEG-video monitoring, and long-term EEG monitoring typically result in several full nights of sleep EEG data. In this work, we leverage artificial intelligence methods that achieved breakthrough performance in related domains with large clinical EEG datasets, to explore our hypothesis that neurological phenotypes that highly correlate with sleep disordered breathing can be extracted from overnight EEG recordings. Furthermore we hypothesize that these EEG phenotypes can be used to accurately predict a patients OSA severity, without accompanying cardiopulmonary data. Methods: We used cross-sectional analyses of adult patients (N = 4650) who completed an overnight PSG study. All signals were excluded from analysis except for the standard 10-20 EEG sensor array, to simulate an ambulatory or video-EEG acquisition for the present study. Global phenotypic features were derived from the patients full-night sleep architecture and fragmentation profiles. Local phenotypic features were derived by analyzing biomarker patterns and respiratory cycle-related EEG changes exhibited in the EEG signals directly. Artificial Intelligence methods including Bidirectional-LSTM and Deep-CNN were trained, optimized, and evaluated to model the relationship between global and local EEG phenotypes and OSA severity. Performance for predicting moderate and severe OSA (AHI ≥ 15) was evaluated using randomized 10-fold cross-validation. Results: The best performance was obtained by a combination of the Bidirectional-LSTM and Deep-CNN architectures, with an average accuracy, sensitivity, and specificity of 91.1%, 86.9%, and 99.5% respectively for predicting moderate and severe OSA. Conclusion: This and prior work have demonstrated a promising opportunity to estimate OSA severity with a host of EEG study types using applied artificial intelligence. Future research involving a cohort of ambulatory EEG subjects, controlled for OSA severity, can validate the efficacy of this approach in the clinical setting. Following further validation, AI based risk estimates could be incorporated into diagnostic EEG reports, to provide clinicians with additional means for identifying patients with moderate and severe OSA that may benefit from follow-up diagnosis and treatment. Support (If Any)