MACHINE LEARNING DETECTION OF AROUSALS FROM AT-HOME SLEEP EEG DEVICE DEMONSTRATES COMPARABLE PERFORMANCE TO EXPERT RATERS IN SLEEP DISORDERS WITH AND WITHOUT ANTIDEPRESSANT USE

Introduction

Cortical arousals provide insights into sleep pathologies and sleep quality through relationships to daytime impairment, respiratory event related arousals (RERAs), and hyperarousal patterns in patients with various sleep and psychiatric conditions. Scoring arousals from in-lab polysomnography (PSG) by expert readers is the gold standard, but is burdensome, time-consuming, and subject to significant inter/intra-rater variability. Here, we compare the detection of arousals from PSG to an FDA-cleared, at-home, dry-electrode EEG device that captures cortical arousals and other sleep metrics. We evaluate this system on individuals with sleep disorders with and without concurrent antidepressant use — a meaningful subgroup comparison, as antidepressants are associated with elevated cortical arousability.

Methods

The at-home EEG device and PSG were simultaneously recorded from 50 individuals (ages 25-65, 54% female) overnight, including 12 with sleep disorders taking SSRI antidepressants (SDAD) and 38 not taking antidepressants (SDCTRL). Arousals were scored by five Registered PSG Technologists following AASM guidelines using PSG signals to serve as ground truth for algorithm evaluation. Arousals on the at-home EEG device were detected by a machine-learning algorithm trained on over 15,000 nights. Performance was evaluated by computing the intraclass correlation coefficient (ICC) of the model’s Arousal index (ArI; arousals per hour of sleep) to the scorers’ ArI, as well as by comparing 30-second epochs with or without arousals to the scorers’ majority vote to compute Cohen’s Kappa.

Results

The mean ArI ICC between pairs of human scorers was 0.74 (SD=0.092, range: 0.610.85). The mean ICC between the at-home EEG device’s ArI and individual scorers was 0.78 (SD=0.043, range: 0.72-0.84) overall, 0.83 (SD=0.078, range: 0.69-0.91) for SDAD and 0.77 (SD=0.044, range: 0.70-0.81) for SDCTRL. Comparing the at-home EEG device ArI to the average ArI of all scorers resulted in an ICC of 0.87 (95% CI: 0.80-0.91) overall, 0.91 (95% CI: 0.81-0.97) for SDAD and 0.86 (0.78 - 0.91) for SDCTRL. The at-home EEG device’s epochlevel kappa of 0.57 (95% CI: 0.54-0.60) exceeded the pairwise agreement of most individual scorers (0.45 to 0.58, evaluated relative to other scorers’ majority vote).

Conclusion

An at-home sleep EEG device with an associated machine learning algorithm can detect arousals with a high epoch-level agreement with expert scorers and provides a reliable estimate of arousal index in multiple populations with disrupted sleep. These findings demonstrate the potential for scalable, objective quantification of sleep fragmentation in realworld settings, enabling more efficient evaluation of therapies and future diagnostic applications.