Background: Neuronal oscillations are linked to symptoms of Parkinson’s disease. This relation can be exploited for optimizing deep brain stimulation (DBS), e.g. by informing a device or human about the optimal location, time and intensity of stimulation. Whether oscillations predict individual DBS outcome is not clear so far.
Objective: To predict motor symptom improvement from subthalamic power and subthalamo-cortical coherence.
Methods: We applied machine learning techniques to simultaneously recorded magnetoencephalography and local field potential data from 36 patients with Parkinson’s disease. Gradient-boosted tree learning was applied in combination with feature importance analysis to generate and understand out-of-sample predictions.
Results: A few features sufficed for making accurate predictions. A model operating on five coherence features, for example, achieved correlations of r > 0.8 between actual and predicted outcomes. Coherence comprised more information in less features than subthalamic power, although in general their information content was comparable. Both signals predicted akinesia/rigidity reduction best. The most important local feature was subthalamic high-beta power (20-35 Hz). The most important connectivity features were subthalamo-parietal coherence in the very high frequency band (>200 Hz) and subthalamo-parietal coherence in low-gamma band (36-60 Hz). Successful prediction was not due to the model inferring distance to target or symptom severity from neuronal oscillations.
Conclusion: This study demonstrates for the first time that neuronal oscillations are predictive of DBS outcome. Coherence between subthalamic and parietal oscillations are particularly informative. These results highlight the clinical relevance of inter-areal synchrony in basal ganglia-cortex loops and might facilitate further improvements of DBS in the future.
Keywords: Deep brain stimulation; Machine learning; Neuronal oscillations; Parkinson’s disease; Subthalamic nucleus.
Analysis pipeline. (A) Feature extraction. Following contact selection, STN power and STN-cortex coherence were computed from the Fourier spectrum. STN power underwent 1/f-correction and was averaged within frequency bands. STN-cortex coherence was source-localized using beamforming. Each source was assigned to one of 30 cortical parcels and source coherence was averaged within parcels and frequency bands. Band-limited STN power and STN-cortex coherence formed the hemisphere feature vector. (B) Leave-one-out regression. Left and right hemisphere feature vectors were stacked vertically to form the subject feature vector. The subject feature vectors were stacked horizontally to form the feature matrix. In each iteration through the leave-one-out cycle, one subject was set aside (test set). The remaining train set was divided into 3 folds for cross-validated hyper-parameter tuning and feature selection. The test features served as input to the regression model, which predicted UPDRS III sum score reduction.