Movement-responsive deep brain stimulation for Parkinson’s Disease using a remotely optimized neural decoder

Deep brain stimulation (DBS) has garnered widespread use as an effective treatment for advanced Parkinson's Disease (PD). Conventional DBS (cDBS) provides electrical stimulation to the basal ganglia at fixed amplitude and frequency, yet patients’ therapeutic needs are often dynamic with residual symptom fluctuations or side-effects. Adaptive DBS (aDBS) is an emerging technology that modulates stimulation with respect to real-time clinical, physiological, or behavioral states, enabling therapy to dynamically align with patient-specific symptoms. Here, we report an aDBS algorithm intended to mitigate movement slowness by delivering targeted stimulation increases during movement using decoded motor signals from the brain. Our approach demonstrated improvements in dominant hand movement speeds and patient-reported therapeutic efficacy compared to an inverted control, as well as increased typing speed and reduced dyskinesia compared to cDBS. Furthermore, we demonstrate proof-of-principle of a machine learning pipeline capable of remotely optimizing aDBS parameters in a home setting. This work illustrates the potential of movement-responsive aDBS as a promising therapeutic approach and highlights how machine learning assisted programming can simplify complex optimization to facilitate translational scalability.

Methods
Neural data sensing configuration and signal processing

All neural recordings were performed during monopolar stimulation. Bipolar recordings in the STN were collected using a “sandwich” configuration (electrodes on either side of the stimulating contact) for common-mode rejection of the stimulation artifact. Two channels of neural data were recorded along the ECoG array, using the first and second electrodes for one channel (postcentral gyrus) and the third and fourth electrodes for the other (precentral gyrus). Neural time-domain data were recorded at a sampling rate of 500Hz. Device-embedded signal processing consisted of short-time Fourier transforms (STFT) with an FFT size of 256 pt and an interval of 50 ms. 

Behavioral data streaming, signal processing, and synchronization

Detailed descriptions of the behavioral data collection platform can be found in Strandquist et al., 2023. Briefly, arm movement data were collected using Apple Watches (Apple, Inc) worn on each wrist and uploaded to a remotely accessible third-party data repository (Rune Labs, Inc.). These data were sampled at 50 Hz, and power was calculated using STFT with an FFT size of 64 pt and interval of 100 ms. Power was calculated in the 0-5 Hz range for movement decoding, 1-4 Hz for quantifying dyskinesia (40), and 4-7 Hz for quantifying tremor (41). Increased power in the 4-7 Hz band of the accelerometry signals was previously found to indicate tremor for this participant when receiving low stimulation (data not shown). Video data for participant observation and kinematic pose estimation were recorded on three cameras arranged at varied angles in the participant’s home office. A custom keylogging application was built to provide statistics regarding typing performance without compromising privacy by removing the letter identity of each keypress. This application was deployed on the participant’s home office computer with patient knowledge and consent. All data was securely transferred to remotely accessible databases over a virtual private network (VPN). To overcome drift and time discrepancies across system clocks and align the data streams, a synchronization event was performed at the start of all recordings. The participant tapped each side of their chest over the implanted neurostimulator while standing in view of the cameras. This movement created identifiable spikes in the neurostimulator and Apple Watch accelerometry signals and could be discerned on the video, allowing realignment of all data to a common time.

Data collection protocols

Three different datasets were collected remotely for this study. For identifying personalized power band candidates, a single two hour session of free behavior was collected during monopolar stimulation at 1.6 mA. The participant was instructed to go about their regular daily routine while occasionally taking breaks to perform hand open-close movements and rest their hands on their lap. Since this dataset was only used to identify the most prominent components of the neural signals through PCA, there were not firm constraints on the participant’s actions during data collection other than ensuring that there were periods of both movement and non-movement. Separate datasets were used for identifying the personalized power bands and evaluating their use in decoding movement to ensure that we avoided any risk of overfitting and represented generalizable performance.

A second dataset was collected for optimizing the movement classifier and performing feature selection from the identified power band candidates. This dataset included six days of data where the participant performed a prescribed set of motor tasks at three different levels of monopolar stimulation (1.2 mA, 2.4 mA, 2.6 mA). Each day consisted of approximately 25 minutes of total recording time. The participant was given an instruction sheet with the list of motor tasks, which included standardized movement tasks from the Unified Parkinson’s Disease Rating Scale Part III (UPDRS; sections 3.17-3.18 rest tremor, 3.4 finger tapping, 3.5 hand movements - open-close, 3.6 hand movements - pronation-supination) as well as blocks of text for typing that were each approximately 100 words long (Figure 2A,B). The participant was a proficient typer prior to beginning the study.

For evaluating the movement responsive aDBS algorithm, 12 days of data were collected while the participant performed a similar set of tasks in three different stimulation conditions (Movement Responsive, Inverted, Constant). Constant stimulation was delivered at 1.9mA, while Movement Responsive and Inverted stim toggled between 1.6mA and 2.2mA. The order in which the conditions were performed in each session was pseudorandomized in balanced fashion and was included as a covariate in all across-condition statistical testing. Each day again consisted of approximately 25 minutes of total recording time. The open-close hand movements (UPDRS section 3.5) from the initial six-day dataset were replaced with finger-to-nose movements (repeatedly extending arm out in front, then bending at elbow to touch nose), and no movements of the lower body were performed. All movements were instructed to last 20 seconds, and the rest phase 10 seconds.

Human subject informed consent

The study participant acknowledged in the associated manuscript provided written informed consent prior to participating in the study, and provided additional written informed consent consenting to acknowledgement in the manuscript prior to publication.