Objective: To evaluate motor classification and somatotopy of local field potentials (LFPs) recorded from subthalamic nucleus (STN) directional deep brain stimulation (DBS) electrodes.

Background: STN DBS is an effective therapy for motor symptoms of Parkinson’s disease, but robust treatments for resistant gait and speech symptoms are lacking. LFPs recorded from directional electrodes may encode spatial features of motor control, such as distinguishing neural activity in upper versus lower limb movements. We hypothesize these signals represent discrete motor states, informing the development of open and closed loop stimulation paradigms that respond robustly to behavioral demands.

Method: We obtained intracranial field potential recordings from the STN in 24 patients undergoing DBS surgery. Motor tasks were rest and simple repetitive movements of the mouth and contralateral hand and foot. Directional DBS leads in anterior, posterolateral, and posteromedial orientations were bandpass filtered from 4-55Hz, and wavelet transforms yielded spectrograms.

We compared three classification approaches. First, we trained a random forest (RF) model with mean powers across canonical frequency bands (theta, alpha, low beta, high beta, low gamma). This was compared to SeqNMF, an algorithm that extracts global spectral factors using non-negative matrix factorization. A second RF model was trained on the SeqNMF factors. Third, a transfer learning approach fine-tuned an image classification model on the resulting spectrograms.

Results: SeqNMF identified seven spectral factors across the sample. With a chance level of 0.25, the validation accuracies for rest, hand, mouth, and foot movements for the SeqNMF RF model were 0.81, 0.59, 0.53, and 0.49, respectively. The frequency bands RF model accuracies were 0.88, 0.57, 0.45, 0.40. The spectrogram classification model accuracies were 0.76, 0.57, 0.46, and 0.43. Feature importance analyses revealed that the posterolateral STN contacts contributed most to decoding mouth and hand movements, and rest, whereas anterior contacts contributed most to foot movements.

Conclusion: STN LFPs contain information about movement somatotopy, but current decoding accuracy is insufficient for real-time clinical application. Future work should extend analyses to high frequency bands, higher dimensional directional electrodes, and compare basal ganglia with motor cortex signals to optimize open and closed loop stimulation.

« Back to 2026 International Congress

MDS Abstracts - https://www.mdsabstracts.org/abstract/movement-decoding-from-subthalamic-nucleus-during-deep-brain-stimulation-surgery/