Deep Neural Networks for Context-Dependent Deep Brain Stimulation

Abstract

Closed-loop deep brain stimulation (DBS) is a novel class of neuromodulation therapy in which stimulation is delivered on demand based on disease or activity state. Some applications of closed-loop DBS for essential tremor (ET) aim to trigger stimulation via detection of overt hand movement from implanted electrocorticographic (ECoG) sensing of motor cortex activity. In this study we examine ECoG activity recorded from three chronically implanted patients while performing a number of activities, and we investigate overt hand movement classification performance for standard beta band (12-30Hz) based spectral feature classifiers against a novel deep neural network (DNN) architecture with automated feature extraction. We find that the DNN architecture significantly outperforms beta band classifiers in overt hand movement detection in this limited cohort of patients, and that this classification performance generalizes to ambulatory activities as well. Finally, we motivate a discussion of context-dependent DBS applications and discuss possibilities for future closed-loop DBS with computationally intensive requirements.