1079 Robust Automated Sleep Staging Using Only EEG Signals

Abstract Introduction Automated algorithms for assisting sleep technologists and clinicians in the staging of sleep have the potential to significantly speed the scoring of PSGs, reduce inter-rater variability, and improve the reliability of diagnosis of sleep disorders. Here, we describe an automated sleep staging machine learning algorithm that scores PSG using only EEG signals. Methods We describe a convolutional neural network-based sleep staging algorithm, SleepStageML™, that was trained on a database of 19,000 polysomnography recordings from a heterogeneous patient population within the Beacon Clinico-PSG Database. The algorithm was evaluated on 100 held-out recordings from 5 clinical institutions across 11 sites, comprising a highly diverse population representative of patients evaluated in sleep clinics. Each recording was independently staged by 3 registered polysomnographic technologists to generate consensus ground-truth sleep staging for each recording. Results The automated algorithm achieved human-level sleep staging performance with per-stage positive-percent-agreements (PPA) of 89% for W, 65% for N1, 81% for N2, 90% for N3, and 92% for R. Negative-percent-agreements (NPA) were 98% for W, 95% for N1, 95% for N2, 94% for N3, and 98% for R. The algorithm achieved a macro-average F1-score of 0.77. In addition, the algorithm’s median absolute error in estimating total sleep time (TST) was 12 minutes, wake after sleep onset (WASO) was 7 minutes, latency to persistent sleep (LPS) was 2 minutes, and REM latency was 1 minute. The algorithm was also able to stage recordings with the minimum number of AASM recommended EEG electrodes without a statistically significant reduction in performance. The sleep staging model was able to generate sleep stages for all 100 recordings within 26 minutes on a computer with an available GPU. Conclusion We demonstrate a model trained on large amounts of highly diverse PSG data that is capable of automatically staging EEG channel data from the PSG to achieve performance matching human experts. An automated staging algorithm that operates on EEG alone has the potential to rapidly provide accurate diagnostic information in a variety of neuropsychiatric conditions with a reduced testing burden, and greatly accelerate the development of novel therapies. Support (if any)