UCSF

Motor control relies heavily on a functional network between the basal ganglia and the sensorimotor cortex. In Parkinson’s disease (PD), disruptions within this network can lead to hypokinetic motor symptoms, such as bradykinesia and rigidity. While deep brain stimulation (DBS) of the basal ganglia has been shown to mitigate these symptoms, the conventional constant-amplitude approach is limited in its therapeutic efficacy by its inability to adapt to fluctuating movement states and symptom severity. Sensing-enabled neurostimulators – which enable the detection of cortico-basal oscillations – have the potential to offer a dynamic alternative: closed-loop, movement-responsive DBS (mDBS) which delivers targeted stimulation during volitional motion. However, its translation to real-world environments requires machine learning (ML) models that can reliably decode movement across changing stimulation amplitudes and extended timeframes. To address this, we first evaluated the predictive power of cortico-basal signals during normal, daily activities under cDBS. Analyzing over 530 hours of continuous cortical and subcortical recordings from 15 PD patients (27 hemispheres), synchronized with wrist-worn accelerometer data, revealed distinct neural biomarkers: movement-related beta (13-30 Hz) desynchronization and gamma (40-80 Hz) synchronization in both regions consistently distinguished mobile from stationary states. Using these spectral features, we developed movement-predictive ML models. Cortical models outperformed subcortical ones, though combining both signals yielded the highest overall accuracy. Notably, while increasing stimulation amplitudes resulted in a decrease subcortical model performance, cortical and combined models remained resilient. These results support the feasibility of developing robust cortico-basal ML models for real-time movement decoding in naturalistic settings. Building on these findings, we conducted a blinded, randomized, crossover feasibility study in four PD patients evaluating chronic home use of subthalamic mDBS using embedded cortical ML models. The mDBS algorithm maintained precision and stability across different timescales (milliseconds to months) and settings (clinical and naturalistic). During structured motor tasks, mDBS increased forearm speed and prevented progressive bradykinetic slowing compared to cDBS, accompanied by cumulative changes in sensorimotor cortical beta activity and coherence. Additionally, during unsupervised, daily activities, mDBS significantly reduced bradykinesia severity compared to cDBS. The progressive symptom relief observed over hours of mDBS - and its subsequent reversal upon returning to cDBS - suggests that movement-responsive stimulation strengthens functional motor circuits and engages activity-dependent plasticity mechanisms rather than merely masking symptoms. Collectively, these findings support the feasibility and therapeutic efficacy of mDBS as a restorative neuromodulation strategy for movement disorders.