Automated Sleep Detection in Movement Disorders Using Deep Brain Stimulation and Machine Learning

Background: Automated sleep detection in movement disorders may allow monitoring sleep, potentially guiding adaptive deep brain stimulation (DBS). Objectives: The aims were to compare wake-versus-sleep status (WSS) local field potentials (LFP) in a home environment and develop biomarkers of WSS in Parkinson's disease (PD), essential tremor (ET), and Tourette's syndrome (TS) patients. Methods: Five PD, 2 ET, and 1 TS patient were implanted with Medtronic Percept (3 STN [subthalamic nucleus], 3 GPi [globus pallidus interna], and 2 ventral intermediate nucleus). Over five to seven nights, β-band (12.5–30 Hz) and/or α-band (7–12 Hz) LFP power spectral densities were recorded. Wearable actigraphs tracked sleep. Results: From sleep to wake, PD LFP β-power increased in STN and decreased in GPi, and α-power increased in both. Machine learning classifiers were trained. For PD, the highest WSS accuracy was 93% (F1 = 0.93), 86% across all patients (F1 = 0.86). The maximum accuracy was 86% for ET and 89% for TS. Conclusion: Chronic intracranial narrowband recordings can accurately identify sleep in various movement disorders and targets in this proof-of-concept study.