Predicting Parkinson’s disease behavioral state from neural and kinematic data

We aimed to identify the biomechanical behavioral state of patients with Parkinson’s disease during a clinical research task from local field potential voltage recordings from surgically implanted electrodes in the brain. A logistic regression, LSTM and 1D CNN model were explored initially. The 1D CNN proved most promising and thus extensive experiments were performed to tune hyperparameters. The best 1D CNN model performed with an average area under the receiver operating characteristic of 0.70 during holdout cross-validation. This is a promising initial step toward our ultimate goal of predicting freezing behavior in Parkinson’s disease. To continue to improve classification performance, objective labels of freezing of gait will be explored.