1095 Detecting Major Depressive Disorder from Sleep EEG Using Artificial Intelligence

Abstract Introduction Obstructive Sleep Apnea (OSA) and Depression have co-linear and self-reinforcing relationships in their incidence, severity, and impacts on mood and cognition. Comorbidity and underdiagnosis of each is common. OSA patients have a 17.6% prevalence of major depressive disorders (MDD), while MDD patients have a 18% prevalence of OSA. Sleep EEG demonstrates characteristic changes in MDD patients including: impaired sleep continuity (increased fragmentation, reduced sleep efficiency), REM sleep disinhibition manifested as elevated REM density and reduced REM latency (≤65 minutes), changes in NREM sleep including reduced slow-wave sleep (≤8%) and lower delta sleep ratio (DSR) (< 1.1), and spectral differences including higher gamma power and elevated alpha asymmetry. These characteristics were shown predictive of MDD diagnosis and treatment response. In this study we sought to detect MDD from sleep EEG using a novel deep learning approach. Methods An AI model training and testing sample of N=10,000 retrospective diagnostic PSGs (26.2% MDD-positive, 73.8% MDD-negative) were collected utilizing stratified sampling with proportionate allocation across MDD diagnostic prevalence, OSA severity, Gender, Age, and comorbid sleep, psychological, and neurological disorders. An EEG-only input deep learning ensemble was pre-trained for sleep staging prior to MDD detection. A subset of N=3,500 PSGs were used for training, N=1,500 for hyperparameter tuning, and N=5,000 studies held-out for final model validation. A second out-of-distribution hold-out dataset of N=135 was collected from patients referred for MSLT (48% MDD-positive). Sensitivity, Specificity, and ROC-AUC were reported and compared to the Delta-Sleep Ratio as a comparative baseline. Results The AI model demonstrated moderate performance in the two independent held-out datasets, with an ROC-AUC of 0.729 (N=5,000) and 0.703 (N=135) respectively. In the N=5,000 PSG sample, the model demonstrated 65.5% Sensitivity at 70% specificity, 49.3% at 80%, and 31.8% at 90% respectively. In the second N=135 MSLT-referred sample, the model showed 56.7% sensitivity at 70% specificity, 46.7% at 80%, and 45% at 90% respectively. The AI approach substantially outperformed the Delta-Sleep Ratio baseline in sensitivity and specificity in both samples with an ROC-AUC of 0.521 (N=5,000) and 0.537 (N=135) respectively. Conclusion The Sleep EEG-based deep learning model showed moderate performance detecting MDD, and outperformed the Delta Sleep Ratio. Support (if any)