Using a Deep Learning Approach for Model-based Control of Deep Brain Stimulation

Deep brain stimulation (DBS) has been developed as a treatment method for various neurological disorders, including Parkinson’s disease, essential tremor and depression. Although DBS is effective, it often loses efficacy over sustained periods because a constant stimulation is applied without adapting to the patient’s current clinical state. In contrast, an adaptive closed-loop DBS system can offer more tailored stimulation in real-time based on a feedback biomarker. In early 2024, we developed a model-based DBS control framework that consists of three main functions: (1) a biophysically reasonable encoding model, (2) a simple decoding model, and (3) a controller. We used a polynomial fit function in the decoding model to approximate the neural-motor relationship, from DBS-induced Vim neural activity to muscle fiber electromyography (EMG). Despite promising results, the polynomial method is inaccurate in capturing the full representation of the neural-motor (EMG) function across different DBS frequencies. In this work, to capture the nonlinear intricate relationship between the neural and EMG patterns, we developed a one-dimensional convolutional neural network (1-D CNN) as a decoding model to predict the EMG signal directly from the DBS-induced Vim neural activity. The 1-D CNN network outputted a high R ² value of 0.997 which significantly outperformed the polynomial method (R ² = 0.277) and a deep learning approach based on long short-term memory (R ² = 0.296). We anticipate that our work highlights the need for a data-driven approach that can reliably map neural activities to symptomatic signals like EMG for better adjusting DBS parameters.