Accuracy of Automated Sleep Classification with Intracranial Spectral Data from the Temporal Lobe. (16th November 2020)

• Record Type: Journal Article
• Authors: Daly, Samuel; Herman, Alexander; Saiote, Catarina; Darrow, David P; Park, Michael C
• Is Part Of: Neurosurgery. Volume 67(2010)Supplement 1
• Publication Date: 2020-11-16
• Subjects: Nervous system -- Surgery -- Periodicals; 617.48005
• DOI: 10.1093/neuros/nyaa447_641
• Languages: English
• ISSNs: 0148-396X

Abstract

Abstract: INTRODUCTION: Automated methods for differentiating sleep from wakefulness with spectral data from intracranial electrodes will be important for the study of conscious states, sleep neural processing, and the development of state-dependent neuromodulation therapies. Existing methods have shown promising results but do not assess single-channel classifier accuracy across patients. METHODS: Five epileptic patients with depth electrodes surgically implanted through the temporal cortex into the anterior hippocampus were included. Separate classifiers were built using spectral data from each contact along the depth electrode from four patients (N = 8–12 contacts/patient). Five-fold cross-validated support vector machine learning with a quadratic kernel was performed to build classifiers over only single channel Berger bands (Delta, Theta, Alpha, Beta, Gamma). These models were then tested on an out-of-sample patient with validated sleep/wake data. RESULTS: Overall, models created from estimated sleep and wake periods (>1000 20-second windows) performed significantly better than chance (p < 1.53e-16). When divided into medial, white matter, and lateral contact, we found the highest average accuracy to be medial models predicting out-of-sample lateral activity (0.89). Within region models also predicted high accuracy (0.82) for medial and lateral. The use of bipolar montage pre-processing yielded higher accuracy results than summed spectral power across all depth electrode contacts. CONCLUSION: This work suggests that it is possible to create an accurate sleep classifier from a single temporal lobe contact using only band power on estimated sleep wake cycles. Out-of-sample validation on sleep/wake data revealed high accuracy rates within electrodes embedded in or near gray matter. This work has direct application to inform existing neuromodulation platforms to automatically integrate and potentially modulate sleep dynamics.