P 34 StimFit – A data-driven algorithm for automated deep brain stimulation programming

Abstract

Question: Finding the optimal deep brain stimulation (DBS) parameters out of a multitude of possible combinations by trial-and-error is time-consuming and requires highly trained medical personnel. We developed an automated algorithm to identify optimal stimulation settings in Parkinson’s disease (PD) patients treated with subthalamic nucleus (STN) DBS based on imaging-derived metrics. Methods: Electrode locations and monopolar review data of 612 stimulation settings acquired from 31 PD patients were used to train a predictive model for therapeutic and adverse stimulation effects. Model performance was then evaluated within the training cohort using cross-validation and on an independent cohort of 19 patients. We inverted the model by applying a brute-force approach to determine the optimal stimulation sites in the target region. Finally, an optimization algorithm was established to identify optimal stimulation parameters. Suggested stimulation parameters were compared to the ones applied in clinical practice. Results: Predicted motor outcome correlated with observed outcome (R = 0.57; p < 10 -10 ) across patients within the training cohort. In the test cohort, the model explained 28% of the variance in motor outcome differences between settings. The stimulation site for maximum motor improvement was located at the dorso-lateral border of the STN. When compared to two empirical settings, model-based suggestions more closely matched the setting with superior motor improvement. Conclusion: We developed and validated a data-driven model which can suggest stimulation parameters leading to optimal motor improvement while minimizing the risk of stimulation-induced side-effects. This approach might provide guidance for DBS-programming in the future.