Machine learning prediction of motor response after deep brain stimulation in Parkinson’s disease

Introduction Despite careful patient selection for subthalamic nucleus deep brain stimulation (STN DBS), some Parkinson’s disease patients show limited improvement of motor disability. Non-conclusive results from previous prediction studies maintain the need for a simple tool for neurologists that reliably predicts postoperative motor response for individual patients. Establishing such a prediction tool facilitates the clinician to improve patient counselling, expectation management, and postoperative patient satisfaction. Predictive machine learning models can be used to generate individual outcome predictions instead of correlating pre- and postoperative variables on a group level. Methods We developed a machine learning logistic regression prediction model which generates probabilities for experiencing weak motor response one year after surgery. The model analyses preoperative variables and is trained on 90 patients using a ten-fold cross-validation. We intentionally chose to leave out pre-, intra- and postoperative imaging and neurophysiology data, to ensure the usability in clinical practice. Weak responders (n = 27) were defined as patients who fail to show clinically relevant improvement on Unified Parkinson Disease Rating Scale (UPDRS) II, III or IV. Results The model predicts weak responders with an average area under the curve of the receiver operating characteristic of 0.88 (standard deviation: 0.14), a true positive rate of 0.85 and a false positive rate of 0.25, and a diagnostic accuracy of 78%. The reported influences of the individual preoperative variables are useful for clinical interpretation of the model, but cannot been interpreted separately regardless of the other variables in the model. Conclusion The very good diagnostic accuracy of the presented prediction model confirms the utility of machine-learning based motor response prediction one year after STN DBS implantation, based on clinical preoperative variables. After reproduction and validation in a prospective cohort, this prediction model holds a tremendous potential to be a supportive tool for clinicians during the preoperative counseling.