Machine Learning Specialist / Ph.D. Student Mayo Clinic / Brno University of Technology Rochester, Minnesota
Rationale: Advanced implantable electrical brain stimulation devices (EBS) enable continuous intracranial electroencephalographic (iEEG) recording. Analysis of long-term iEEG data acquired by such systems reveals new opportunities for objective monitoring of EBS outcome and patient health and well-being. Long-term sleep analysis using the continuous iEEG has potential for evaluating the effects of EBS on sleep quality. With new technology new challenges also arise; synchronous EBS and iEEG streaming make accurate classification of iEEG data into sleep stages challenging. Methods: A human subject underwent long-term monitoring using the Medtronic investigational Summit RC+S (TM) implantable neural stimulator (INS) with electrodes implanted in bilateral hippocampus & anterior nucleus of the thalamus. Three consecutive nights of standard sleep clinical polysomnography (PSG) were recorded simultaneously with continuous iEEG data streaming from the INS. The first night was stimulation-free, and in the second and third nights we trialed 15-minutes of no stim, 2, 7, and 100 Hz stimulation with 15-minute wash-out periods between EBS parameter changes. PSG data were scored according to gold standard sleep categories using AASM2012 rules. A classification algorithm was designed and trained using the first-night iEEG data. The second-night data was used for a validation, and the third-night data was used for testing. Results: A behavioral state classifier (wake, REM sleep, and non-REM sleep) using long-term iEEG recordings with EBS stimulation artifacts was designed, and prospectively tested with overall F1-score 0.86, Cohen’s Kappa score 0.78 and accuracy 0.87 for three categories: wake, non-REM, and REM. To create a sleep/wake profile for a given patient, the model was then deployed on long-term data (over six months of continuous iEEG), and evaluated as a proof-of-concept for online sleep scoring in an ambulatory patient implanted with an INS capable of continuous iEEG sensing. Conclusions: The trained classifier enables the assessment of behavioral states of human subjects implanted with an INS for epilepsy treatment. The classifier uses the data from iEEG recordings in the presence of EBS stimulation artifacts. Such a system will enable a sleep quantification of long-term data and objective evaluation of the effect of EBS on sleep quality of patients with epilepsy. Funding: Please list any funding that was received in support of this abstract.: This research was supported by National Institutes of Health (UH2&3-NS95495, R01-NS092882), DARPA Morepheus, LQ1605 from the National Program of Sustainability II (MEYS CR, Czech Republic), and institutional resources from Mayo Clinic, Rochester MN USA, Medtronic Plc, Minneapolis, MN, USA, and Czech Technical University in Prague, Czech Republic.