Towards automated sleep-stage classification for adaptive deep brain stimulation targeting sleep in patients with Parkinson’s disease

Abstract Sleep dysfunction affects over 90% of Parkinson’s disease patients. Recently, subthalamic nucleus deep brain stimulation has shown promise for alleviating sleep dysfunction. We previously showed that a single-layer neural network could classify sleep stages from local field potential recordings in Parkinson’s disease patients. However, it was unable to categorise non-rapid eye movement into its different sub-stages. Here we employ a larger hidden layer network architecture to distinguish the substages of non-rapid eye movement with reasonable accuracy, up to 88% for the lightest substage and 92% for deeper substages. Using Shapley attribution analysis on local field potential frequency bands, we show that low gamma and high beta are more important to model decisions than other frequency bands. These results suggest that the proposed neural network-based classifier can be employed for deep brain stimulation treatment in commercially available devices with lower local field potential sampling frequencies.