Abstracts

Rationale: Studies investigating the relationship between epilepsy and sleep usually involve the assignment of sleep stages by an expert sleep technician, which is time-consuming and labor-intensive. Most automated methods depend on handcrafted features trained on relatively small datasets that are fewer than 100 individuals, limiting their generalizability. Recent advances in machine learning have enabled the training of algorithms on large polysomnography (PSG) datasets. However, most of these detectors use referenced channels that may preclude their use on scalp EEG recordings that do not use these references. In this work, we use deep learning methods on a large PSG dataset to develop a reference-free sleep scoring algorithm that could be applicable to EEG recorded in many clinical settings. Methods: EEG data recorded from 6,431 patients who underwent routine diagnostic PSGs were scored in nonoverlapping 30-second epochs according to the American Academy of Sleep Medicine (AASM) standards. For each recording, the EEG signals were filtered, downsampled, and re-referenced to a longitudinal bipolar montage with 4 channels (F3-C3, C3-O1, F4-C4, and C4-O2). The recordings from 5,041 patients were used for training a deep neural network consisting of convolutional layers (CNN) and bi-directional long short-term memory layers (LSTM). The rest of the recordings were equally split for validation (650 patients) and testing (650 patients). Results: Application of the hybrid CNN/LSTM neural network on the testing set results in an agreement of 81.1% (Cohen’s kappa = 0.74), which is close to the inter-rater agreement between human experts of 82.0% (Cohen’s kappa = 0.76). Conclusions: We have developed an end-to-end, reference-free, deep learning model that can achieve near human level sleep staging performance on a large dataset of PSG recordings. Future work will aim at determining the applicability of this detector to long-term scalp EEG recordings for use in both epilepsy research and clinical interpretation. Funding: ADL was funded by NIH NINDS K23 NS101037 and R25 NS065743, and the American Academy of Neurology Institute. MBW was funded by NIH NINDS 1K23NS090900, 1R01NS102190, 1R01NS102574, and 1R01NS107291. No funding sources had any involvement in the study design, collection, analysis, interpretation of data, or the decision to submit this abstract.