At-home movement state classification using totally implantable cortical-basal ganglia neural interface

Abstract

Decoding human movement from invasive neural signals has traditionally relied on complex machine learning algorithms using data collected from short-term laboratory tasks, limiting understanding of brain function during natural behavior and hindering development of clinically viable closed-loop neuromodulation. Here, we demonstrate the first in-human, at-home classification of a specific movement state—walking—using a fully implantable, bidirectional neurostimulator. In four individuals with Parkinson’s disease, we recorded chronic motor cortex and globus pallidus activity synchronized with wearable kinematic data across over 80 hours of unsupervised daily activity. We identified highly predictive personalized spectral biomarkers of gait and validated their performance. Critically, we showed that these biomarkers could drive real-time movement state classification using the neurostimulator’s embedded linear discriminant classifier, satisfying device-level constraints for closed-loop stimulation. Our results establish a previously unidentified pipeline for real-world neural decoding and scalable framework for personalized adaptive neuromodulation, expanding the translational reach of implantable brain-computer interfaces.

Chronic at-home neural recordings in Parkinson’s disease reveal personalized biomarkers that accurately classify walking state.

INTRODUCTION

Movement decoding from invasive human intracranial recordings has typically relied on advanced machine learning algorithms programmed into external computers with corresponding outputs for behavioral state classification. Despite the notable strides made by brain-computer interfaces (BCIs) in motor and speech restoration after stroke and spinal cord injuries (1–4), these systems are trained on data largely collected in laboratory environments, where subjects perform standardized tasks while neural activity is recorded. As a result, the implementation of most BCI technology is resource-intensive, time-limited, and detached from natural behavior (5). The reliance on rigid, task-specific experimental designs contrasts sharply with the variety and spontaneity of daily activities performed by subjects outside observational encounters. In addition, subjects are usually physically tethered to recording equipment, limiting the ability to study freely mobile behavior such as gait (6). Together, these logistical and methodological considerations have limited our understanding of neural activity in the real world.

The recent development of implantable neurostimulators (INS) with sensing functions has substantially expanded our knowledge of human brain function and pathophysiology. The advanced capabilities of these bidirectional interfaces to simultaneously record neural activity and provide stimulation have enabled investigations of critical brain networks implicated in an array of neurological and psychiatric disorders (7–17). By facilitating the collection of larger datasets across longer timescales, these devices have made it possible to identify intracranial biomarkers of language processing (18), seizure risk (7, 8), pain (13), as well as immobile and dyskinetic states in patients with Parkinson’s disease (PD) (14, 19).

Although bidirectional neurostimulators have proven useful for identifying biomarkers of pathological brain networks, the utility of these devices for classifying movement state in the naturalistic setting is conspicuously unknown. Accurate classification of movement state is central to improving the treatment of many circuit pathologies since motor activity affects connectivity within brain networks and modulates the oscillatory patterns of various cortical and subcortical regions (20, 21). This paradigm is particularly evident in PD, where both medication and movement states substantially alter cortical-basal ganglia network oscillations, although standard treatment with continuous deep brain stimulation (DBS) does not account for these dynamic fluctuations. Current strategies in adaptive DBS (aDBS) have focused on changing stimulation amplitude based on neural biomarkers of low medication state [increased beta frequency (13 to 30 Hz) synchronization] (22, 23) or dyskinetic episodes [increased narrowband gamma (~65 Hz) activity] (24), but not on specific movement states. The ability to identify distinct motor states in individuals with PD would greatly enhance aDBS algorithms to target and improve patient mobility.

Specifically, gait disturbances represent one of the most prevalent and debilitating symptoms of PD, making the identification of neural signatures of gait a critical first step in designing aDBS protocols aimed at improving gait functions (25). A growing body of evidence suggests that parkinsonian gait disorders may be better addressed by stimulation frequencies lower than those used for conventional clinical stimulation (e.g., ~60 Hz) (26–30). However, gait-optimized frequencies are often less effective for appendicular symptom control (e.g., tremor, bradykinesia, and rigidity) (31, 32). This divergence in therapeutic stimulation settings for appendicular and axial symptoms necessitates reliable and ecological biomarkers of a person’s movement state, which has not yet been explored in the real world (23, 24, 33, 34).

To address these areas of need, our study aimed to develop a pipeline for identifying naturalistic neural biomarkers of gait state in patients with PD. By leveraging long-term at-home intracranial recordings paired with kinematic data collection through external wearable sensors, we sought to decode neural signatures of gait using a totally implanted system (Fig. 1).

Fig. 1. Pipeline for identification of cortical-pallidal neural biomarkers of at-home movement state in Parkinson’s disease.

(A) Schematic of Summit RC+S bidirectional neurostimulator cortical electrodes overlying M1 and PM (left) and subcortical depth electrode implanted in the GP (right). (B) Representation of Rover accelerometer WDs, worn around the ankles bilaterally. Sample acceleration signals are shown from the left and right foot. (C) Neural data from M1, PM, and GP and acceleration data from WDs and INS were streamed from patients’ homes. (D) Acceleration signals were aligned, partitioned into 10-s epochs, and labeled as periods of continuous walking (W) or nonwalking (NW) using WD data. (E) From each epoch, average power was calculated within all possible frequency bands from 1 to 50 Hz. (F) These power value features were used to train and test LDA models. (G) Last, testing was performed to simulate on-board classification of movement state using system-constrained biomarkers derived from at-home data. Illustrations of coronal brain cut in (A), legs in (B), and man walking, sitting, and standing in (C), (D), and (G) were BioRender.com. Created in BioRender. Ramesh, R. (2026) https://BioRender.com/cpzbqp2. FFT, fast Fourier transform; LD, linear discriminant.

To our knowledge, our study represents the first attempt to collect and decode at-home movement state using chronic, multisite neural activity in patients with PD. We used wearable sensors (Rover, Sensoplex, Inc.) to label at-home movement state in four patients with PD, while simultaneously recording subcortical local field potential (LFP) and motor cortex electrocorticography (ECoG) signals via an investigational sensing-enabled neurostimulator (Summit RC+S, Medtronic, Inc.). By comparing cortical-pallidal oscillatory activity during walking and nonwalking epochs, we identified both shared and subject-specific neural oscillatory bands that distinguish these movement states. Using machine learning techniques, we characterized the relative importance of different frequency ranges and brain regions to accurately classify gait state and derived highly predictive individualized biomarkers of gait. Last, we tested the accuracy of the on-board classifier to distinguish gait versus nongait states with these biomarkers using offline replication of embedded decoding. These findings expand our understanding of gait neurophysiology and advance the viability of at-home data collection as a modality for the identification of patient-specific movement state biomarkers for closed-loop aDBS.

RESULTS

Patient characteristics and contact localization

Four individuals with PD (2 male and 2 female) undergoing evaluation for DBS implantation were recruited and implanted with a bidirectional investigational device with dual stimulation-sensing capabilities (Summit RC+S, Medtronic, Inc.). Subject demographics are presented in Fig. 2A. All participants enrolled in this study exhibited gait dysfunction, which was assessed with the Movement Disorder Society Unified Parkinson’s Disease Rating Scale Part III (MDS-UPDRS III) (score range: 21 to 57). Specifically, this scale’s posture instability and gait disorder (PIGD) subscore was used to characterize challenges with balance, freezing, and postural instability (score range: 2 to 10).

Fig. 2. Subject demographics and electrode localization.

(A) Demographic and clinical characteristics are shown for the four subjects enrolled in this study. (B) Subject-specific reconstructions are shown of ECoG contacts targeting M1 and PM and depth electrodes targeting GP. Stimulation contacts are indicated in red and recording contacts are indicated in blue. C, case; F, female; L, left; M, male; R, right.

Two subjects received unilateral implants (left hemisphere), whereas the remaining two received bilateral implants. Each DBS device consisted of quadripolar depth leads implanted in the globus pallidus (GP) and subdural quadripolar paddles overlying the sensorimotor cortices. Cortical electrodes recorded neural data from the primary motor (M1) and premotor (PM) cortices in all subjects. Localization of depth and surface electrode contacts is shown in Fig. 2B, and further details are provided in fig. S1. Electrodes were connected to INS, which generated electrical impulses and streamed kinematic and neural data (Materials and Methods). Throughout this study, all subjects received continuous conventional stimulation using their clinically optimized stimulation parameters.

Movement state labeling using wearable accelerometer devices

To collect chronic recordings of at-home movement, subjects were provided wearable ankle accelerometer devices (WDs) (Rover, Sensoplex, Inc.), each containing a sensor composed of a triaxial gyroscope, accelerometer, and magnetometer, which recorded kinematic data locally. To validate the ability for these devices to accurately identify walking in our subjects, we collected data from trials of observed walking and nonwalking periods in controlled environments. During these in-laboratory trials, data were recorded from participants’ INS, WDs, and force-sensitive resistors (FSRs) placed under the feet bilaterally. Signals from these three devices were aligned, and WD-based labeling of movement state was compared with ground-truth labels (Materials and Methods). An example data collection session is illustrated in Fig. 3A.

Fig. 3. Validation of WD accuracy in labeling walking and nonwalking epochs.

(A) Acceleration signals from left INS and bilateral WDs are shown aligned with force signals from FSRs on the subject’s feet during a 300-s session of overground walking with interspersed periods of standing or seated rest. The bar at the top indicates subject behavior throughout the session (blue, walking; red, nonwalking). A 5-s period of continuous walking indicated by the gray box is enlarged and displayed within the subpanel on the right. (B) Subject-specific mean accuracies of WD-based movement state labeling are shown (fig. S2B). Error bars represent standard deviations. a.u., arbitrary units; FS, force signal; HST, heel strike threshold; LH, left heel; LT, left toe; RH, right heel; RT, right toe.

For all subjects, WDs enabled excellent discrimination of walking and nonwalking states (Fig. 3B). Mean accuracy for WD-based labeling was above 95.0% for all subjects, with a range from 95.8% (Subject 1) to 99.0% (Subject 2). Sensitivity varied from 94.4% (Subject 2) to 98.9% (Subject 1), and specificity ranged from 94.7% (Subject 1) to 100.0% (Subjects 2 and 4). See fig. S2 for further details. These results demonstrate robust WD-based movement state identification across the variety of gait metrics displayed by our subjects (e.g., stride length, velocity, and cadence; table S1) and validate the use of WDs to accurately identify periods of walking and nonwalking in our cohort.

Chronic at-home streaming of kinematic and neural data

Subjects streamed chronic at-home neural and kinematic data from their Summit RC+S INS and WDs as continuously and for as many days as possible. Subjects were fully autonomous in wearing the sensors and initiating recordings. All subjects were receiving clinically optimized continuous GP stimulation and standard pharmacological treatment during these recordings (table S2). Acceleration signals from the INS and WDs were aligned to synchronize neural time-domain data with kinematic data. A sample aligned at-home recording is presented in Fig. 4A. To facilitate subsequent analyses, we partitioned neural signals into 10-s epochs and used WDs to label each epoch’s gait state. Across all six hemispheres, a total of 84.5 hours of neural and kinematic data were aligned and analyzed (range: 8.3 to 25.5 hours) from an average of 13 days per hemisphere (range: 5 to 23 days) (fig. S3).

Fig. 4. Spectral analysis of at-home walking and nonwalking epochs.

(A) Sample at-home recording is shown from Subject 4, with aligned acceleration signals from left WD (top) and INS (bottom). Periods of walking (blue) and nonwalking (red) identified with WD-based labeling are indicated with shading. (B) Sample mean PSDs are shown from 0 to 50 Hz for all walking (blue) and nonwalking (red) 10-s epochs analyzed from GP, M1, and PM for Subject 4. (C) Violin plots are shown comparing mean power within each canonical frequency band between all 10-s nonwalking (red) and walking (blue) epochs. Two-sided Wilcoxon rank sum tests were used, with Benjamini-Hochberg correction for multiple comparisons. Across all hemispheres, M1 α and β power was significantly lower during walking epochs, highlighted with yellow boxes. *P < 0.05 (table S3). (D) Normalized mean feature coefficients from logistic regression classifiers of movement state are visualized for all hemispheres (table S4). δ, 1 to 4 Hz; θ, 4 to 8 Hz; α, 8 to 13 Hz; β, 13 to 30 Hz; γ_low, 30 to 50 Hz.

Spectral analysis of walking and nonwalking epochs

To examine differences in cortical-pallidal neural activity between movement states, we performed within-subject spectral analysis comparing frequency representations between walking and nonwalking periods. Power spectral densities (PSDs) for each subject and region are presented in Fig. 4B and fig. S4. For each 10-s epoch, we computed the average power within canonical frequency bands [i.e., delta (1 to 4 Hz), theta (4 to 8 Hz), alpha (8 to 13 Hz), beta (13 to 30 Hz), and low gamma (30 to 50 Hz)] for the GP, M1, and PM neural signals.

Overall, when comparing canonical frequency band power across movement states, we observed significantly lower M1 alpha and beta power during walking compared to rest in all subjects (P ≤ 10⁻⁴, two-sided Wilcoxon rank-sum test) (Fig. 4C). In addition, M1 low gamma power was lower during walking in all but Subject 3’s right hemisphere (P ≤ 10⁻⁹). Trends in power within all other canonical bands and regions differed by individual; summary power values and P values are detailed for all comparisons in table S3.

In Subject 1, left GP power was higher across all canonical frequency bands except beta during walking (P ≤ 10⁻¹⁹), whereas right GP power was lower in all bands except delta (P ≤ 10⁻²¹). With the exception of delta power in the left M1, bilateral M1 power was decreased across all bands during walking (P ≤ 0.033). PM power exhibited relative variability bilaterally.

Subject 2 demonstrated significantly increased GP and PM power in all five canonical frequency bands during walking epochs compared to nonwalking epochs (P ≤ 10⁻¹⁴). In M1, walking was associated with higher low-frequency (delta and theta) power and lower high-frequency (alpha, beta, and low gamma) power (P ≤ 10⁻⁵).

In Subject 3, left GP power was lower in all canonical bands except delta during walking epochs (P ≤ 10⁻¹⁵), whereas right GP power was higher in all bands except beta (P ≤ 0.001). Left M1 power was decreased in all bands during walking epochs (P ≤ 10⁻⁸), whereas right M1 power was more variable. Similar to Subject 1, PM power showed bilateral variability.

In Subject 4, GP power in delta, theta, alpha, and low gamma bands was higher during walking, whereas beta power was lower (P ≤ 10⁻²⁷). In M1, walking epochs were associated with higher delta and theta power, and lower alpha, beta, and low gamma power (P ≤ 10⁻⁵). In PM, power within all frequency bands except theta was lower during walking (P ≤ 0.003).

To assess the relative contribution of each frequency band to movement state classification, we constructed subject- and hemisphere-specific logistic regression models to classify walking versus nonwalking states. Significant decoding performance was achieved in all four subjects (one-sided empiric permutation-based P values of <0.001) (table S4). We visualized mean regression coefficients rescaled from −1 to 1 for each subject to identify the most important frequency bands across individuals (Fig. 4D). Normalized coefficients and further model details are presented in table S4. This analysis revealed that both the most influential canonical power band and the most important region varied widely between participants and even between hemispheres in subjects with bilateral implants. These findings suggest that the oscillatory markers of movement state differ between individuals.

Identification of personalized cortical-pallidal movement state biomarkers

While canonical frequency bands offer the advantage of standardized comparison across individuals, they also carry several limitations including the potential masking of endogenous narrowband peaks (35). It is possible that there exist subject-specific frequency bands that enable more accurate differentiation of walking and nonwalking than these generalized canonical ranges. To account for this important consideration, we shifted our focus from canonical ranges to all possible frequency bands of sizes 1 to 49 Hz in the range of study (1 to 50 Hz), which was selected to minimize contributions from clinical stimulation frequencies and subharmonics.

First, using random forest (RF) models, which are well suited to handle collinearity and high dimensionality, we calculated the importance of each frequency band for accurate discrimination of gait state. By visualizing these feature importances rescaled from 0 to 1 for each subject, we found notable variety in both the regions and bands that supported classification for each subject and hemisphere (Fig. 5A). The most important bands for each subject and region are presented in table S5. Pallidal features were the most important for most models (4/6), tending to focus on bands falling within delta, theta, and beta range ranges. Notably, in the hemispheres where delta and theta GP activity demonstrated a peak in feature importance, these ranges showed higher activity during periods of walking (Subject 1 left, Subject 2, and Subject 3 right). Contrarily, in hemispheres where beta range GP activity demonstrated a peak in feature importance, pallidal beta activity was lower during periods of walking (Subject 1 right, Subject 3 left and right, and Subject 4) or, in the remaining case, not significantly different between movement states (Subject 1 left). Pallidal features were followed in importance by those from M1, which were similarly distributed in low delta and theta frequencies or captured the mid to high beta range. PM features were consistently the least important; on average, the most important features within this region reflected higher frequency bands than those in GP and M1.

Fig. 5. Identification of subject-specific movement biomarkers.

(A) Normalized feature importance is shown for all cortical-pallidal frequency bands using mean decrease from impurity (MDI) from 1000 RF iterations. (B) LDA movement state classifier performance is shown for each hemisphere and type of model (single-region, multiregion, and complete). Black dots indicate permuted chance-level performance (n = 1000), which were used to calculate one-sided empirical P values. *P < 0.001 (table S6).

We next trained and tested cross-validated linear discriminant analysis (LDA) models to classify movement state using all frequency band features. LDA models were selected as they generate feature weights that can be entered into the Summit RC+S on-board classifier when embedding adaptive architecture. Separate models were constructed for each hemisphere using features from all combinations of brain regions: single-region (GP, M1, and PM), multiregion (GP+M1, M1+PM, and GP+PM), and complete (GP+M1+PM). In all hemispheres, successful movement state classification was possible with activity from a single region [area under the curve (AUC) range: 0.67 to 0.98, P < 0.001] (Fig. 5B and table S6). Models using pallidal data were the best-performing single-region models in all individuals except for Subject 4, in whom the M1 model was superior. In three of the six independent hemispheres analyzed, the GP single-region model led to the highest binary classification AUC of any type of model tested (AUC range: 0.86 to 0.98, P < 0.001). In the remaining cases, the best-performing models were either multiregion or complete models incorporating data from the GP (AUC range: 0.77 to 0.96, P < 0.001). The best-performing models across all subjects ranged in sensitivity from 74.8 to 95.8% and specificity from 66.9 to 94.0%. Positive predictive value (PPV) was similarly high, varying from 69.7 to 94.1%. These results support the hypothesis that cortical-pallidal oscillatory activity is modulated by movement state.

Offline testing of on-board movement state classification using at-home data

The use of neural biomarkers to drive closed-loop aDBS in practice is limited by the technical requirements of the classifiers on-board adaptive-enabled neurostimulators. Therefore, to evaluate the ability to use at-home neural-kinematic data to derive biomarkers compatible with actual aDBS, we performed offline testing of movement state classification on-board the Summit RC+S device, which allows only four total features with no more than two from a single region.

First, we tested on-board classification of at-home movement state using biomarkers derived from at-home data. Classification using biomarkers consisting of one to four features meeting device constraints was tested (see Materials and Methods for further details). Above-chance performance was achieved for at least one biomarker in all hemispheres (P value range: <0.001 to 0.006) (Fig. 6A and table S7). AUC for the best-performing simulations ranged from 0.64 (Subject 1 left) to 0.81 (Subject 3 right) (P < 0.001). In all hemispheres, the best-performing models were those using three or all four of the permissible features. These models varied in sensitivity (66.0 to 83.3%), specificity (51.6 to 69.1%), and PPV (59.5 to 71.3%).

Fig. 6. Testing of on-board movement state classification.

(A) Performance of simulated on-board LDA classifiers is shown for models tested on at-home data (blue) and validated with testing on in-laboratory observed trials of overground walking with interspersed periods of standing or seated rest (orange). Empirical one-sided P values were calculated by comparing model performance with a permuted chance-level distribution (n = 1000) (table S7). Models deemed statistically insignificant (P ≥ 0.05) are shown with grayed bars. (B) Sample visualization of continuous on-board classification using 5-s epochs is shown for Subject 1 using biomarker derived from at-home data. An observed trial is shown with true movement state indicated by the bar at the top. Simulated LD output is shown with the threshold indicated by red line, followed by the corresponding aDBS state and true INS acceleration signal. Illustrations of man walking, sitting, and standing in (B) were created using BioRender.com. Created in BioRender. Ramesh, R. (2026) https://BioRender.com/cpzbqp2.

We next tested device-compatible biomarkers generated with at-home data on recordings of observed in-laboratory trials of overground walking with interspersed periods of standing or seated rest. Again, above-chance performance was demonstrated in all hemispheres, with best-performing models’ AUC varying between 0.63 (Subject 1 left) and 0.87 (Subject 1 right) (P < 0.001) (Fig. 6A and table S7). Sensitivity (range: 73.0 to 90.2%), specificity (range: 59.1 to 81.3%), and PPV (range: 61.5 to 82.8%) similarly differed across models. Unlike on-board classification of at-home movement state, the best-performing biomarkers for classification of in-laboratory trials were not consistently those that used three or four features. A sample on-board simulation of on-board movement state classification is visualized for Subject 1’s right hemisphere in Fig. 6B. These findings support the feasibility of using at-home data to derive effective device-compatible movement state biomarkers for closed-loop adaptive stimulation.

DISCUSSION

Sensing-enabled interfaces now offer the opportunity to dynamically modulate stimulation in response to changes in brain activity, although a major challenge has been the study of complex behavior and identification of naturalistic neural biomarkers. In particular, the ability to accurately discern movement states in the real-world represents a critical gap in designing neuromodulation therapies targeting specific symptoms. Here, using WDs and a totally implanted bidirectional neurostimulator, we provide the first in-human evidence for classification of a specific movement state at home from over 80 hours of synchronized kinematic and neural data. Comparing frequency representations within cortical-pallidal signals, we found shared and subject-specific spectral features enabling accurate within-subject discrimination of walking and nonwalking periods. With machine learning techniques, we characterized the variable importance of canonical and noncanonical frequency bands and developed highly predictive individualized cortical-pallidal signatures of gait. Last, we constrained biomarkers to the practical limitations of the classifier embedded on-board the Summit RC+S neurostimulator and tested their performance, showing above-chance decoding in all hemispheres. These results have important implications for our understanding of the supraspinal oscillatory dynamics of locomotion and demonstrate the feasibility of successfully collecting and decoding movement state using chronic at-home multisite neural recordings in patients with PD.

Our pipeline provides an effective approach for investigating brain circuit control of natural behaviors in freely moving humans, which have proved difficult to study in the laboratory setting. Specifically, our understanding of human gait has been limited by methodological constraints. Studies of supraspinal control of gait are typically performed with electroencephalography systems or externalized DBS leads in operating rooms and laboratories. As gait is a highly dynamic state that requires constant adapting and updating based on environmental changes and behavioral goals, studying this movement in an artificial environment presents a substantial challenge. However, understanding how gait-related circuits are modulated in the real world with changes in the surrounding environment is critical for treating dynamic symptoms such as freezing of gait, which are often triggered by directional changes, crowded spaces, or emotional states that cannot be easily replicated in experimental settings (36).

Our approach of identifying neural signatures of gait across different medication cycles over long-term recordings at home yielded important insights about circuit dynamics during walking. Notably, we found lower M1 alpha and beta power and higher GP low-frequency delta and theta activity during periods of walking compared to rest. Oscillatory patterns reflect synchronized activity across neural ensembles and may represent modulations in neuronal excitability that enable the encoding and processing of information within brain networks (37–39). Studies of human bipedal walking have demonstrated gait cycle-related changes in brain rhythms, although few have used invasive methods of recording neural activity and even fewer have recorded data outside laboratory conditions (40–44). In our study of freely ambulating subjects, lower alpha and beta power in M1 during gait align with prior demonstrations of alternating suppression of these bands during periods of walking (45). One potential explanation is that desynchronization within these frequency ranges may act to dampen irrelevant sensorimotor signals and direct the spatiotemporal precision of motor activity (21, 46–48).

Our findings also showed that a single region was sufficient to classify movement state in all hemispheres; the preferential importance of GP features to gait decoding seen in most cases supports the hypothesis that pallidal oscillatory activity is central to movement. The most important features from GP signals tended to be narrowband activity from the delta, theta, and beta ranges. Low frequency (delta and theta) GP activity has previously been associated with dystonia, suggesting that synchronization in these ranges promotes a prokinetic state, perhaps signaling supraspinal activation of large muscle groups (49, 50). In subjects who showed high importance of decreased low-frequency activity during gait, GP beta power was up-regulated during walking epochs, consistent with prior findings seen even in patients without marked gait abnormalities, which may indicate that this oscillatory change is not unique to the parkinsonian state (51). In contrast, for subjects where models showed high beta band importance, GP beta activity during walking was lower compared to the nonwalking state, which may suggest a more pathological gait pattern that requires greater beta desynchronization to promote an antikinetic state (50–55). The specificity of this finding, however, may be affected by the relationship between medication state and beta power, requiring further investigation. Overall, it is possible that the balance between beta desynchronization and low-frequency synchronization are necessary for gait, with the former releasing the “brake” and the latter activating motion.

These results also underscore differences between the GP and the subthalamic nucleus (STN) that may influence their respective utility for gait decoding and aDBS design. Both nuclei are tightly coupled within the basal ganglia–thalamocortical network, and pathophysiological beta activity is present in both structures, with modulation during movement and dopaminergic treatment (56, 57). Recording from the STN, however, has been more common in existing literature. STN LFPs display modulation across both low- and high-beta bands during gait and freezing episodes, with spatially distinct encoding in posterior sensorimotor versus associative subregions (58). STN recordings capture a broad range of gait-related processes, including initiation, termination, and muscle activation dynamics (59). GP recordings have similarly demonstrated robust modulation of low-beta activity during stepping and walking, and our group has observed correlations between pallidal beta power during specific gait phases and improved gait performance (60, 61). One potential advantage of GP over STN is that its signals may be less contaminated by nonmotor processes. While the STN is functionally heterogeneous, with associative and limbic territories in close proximity to motor regions, the GP is more exclusively motor. Furthermore, GP signals may more directly reflect motor output pathways, given its position as the final common inhibitory projection of the basal ganglia. Clinically, GP is also typically preferred as a DBS target over STN in patients with neurocognitive symptoms, with some studies suggesting GP DBS may additionally better preserve gait function (62–65). Comparative studies, however, are still limited, and future work explicitly differentiating STN and GP decoding strength for gait states ideally in the same subjects will be critical for clarifying nucleus-specific biomarker utility and optimizing aDBS therapy.

The ability to decode movement state with acceptable accuracy using subcortical signals alone also raises important questions regarding the added value of cortical recordings in implantable DBS systems. In our study, while M1 ECoG signals improved decoding performance in a subset of patients and provided the most informative features in one subject (Subject 4), we found that GP-derived signals alone were often sufficient for reliable classification of walking versus nonwalking states. This suggests that for many patients, the benefits of including ECoG for binary movement state classification may not outweigh the additional surgical complexity, clinical management burden, and regulatory challenges of cortical arrays. The added risks of subdural array placement and long-term stability issues such as lead migration and cortical irritation may further limit the practicality of chronic cortical interfaces outside investigational settings. However, cortical features have been implicated in more complex movement phenomena, such as elevated M1 phase-amplitude coupling during freezing of gait (FoG), indicating that cortical input may become increasingly valuable for higher-dimensional or symptom-specific decoding frameworks (66). Future work should evaluate whether less invasive cortical sensing options such as epidural electrodes or minimally invasive cortical interfaces could offer a balance between decoding performance improvement and clinical feasibility.

The ability to distinguish walking and nonwalking states is critical to the treatment of advanced gait disorders in PD, which remain some of the most debilitating and treatment-resistant aspects of PD. Even modest improvements in gait function can have profound impact on patient outcomes, given that gait disorders are the primary drivers of falls, hospitalization, and loss of independence in advanced PD (67–71). Critically, pathological features such as reduced gait speed (25, 72), shorter step length (72, 73), and increased interlimb asymmetry (74, 75) are present across everyday walking in PD, not only during complex episodes such as FoG. We have previously demonstrated that optimized parameters can significantly enhance overall walking performance, with measurable improvements in stride velocity, arm swing amplitude, and step time and length (61). Therefore, the binary classification of walking versus nonwalking addresses a fundamental therapeutic branch point: By distinguishing periods of ambulation, we can more precisely target gait-specific deficits with optimized stimulation parameters. Despite this clinical need, the treatment of gait disturbances with conventional clinical stimulation has been a challenge. Early trials of aDBS for PD have used cortical or subcortical control signals to selectively regulate stimulation amplitude during periods of symptom worsening, demonstrating improved symptom control (23, 24, 33, 34). While pilot studies have produced nearly uniformly encouraging findings, adaptive approaches to gait dysfunction remain unexplored due to our limited understanding of neural changes in the real world. With the pipeline discussed here, we propose one method of identifying cortical-pallidal biomarkers that can be used on-board embedded systems to automatically switch stimulation from conventional clinical settings when patients are at rest to gait-optimized parameters when patients are walking. While this initial binary classification scheme does not yet detect critical gait states such as FoG, it provides a strong foundational framework for identifying periods when patients are most vulnerable to gait-related impairments. A key consideration for extending this pipeline to more complex movement states is the challenge of aligning multimodal data streams and reliably generating high-quality ground truth labels. For example, robust detection of FoG in the home setting would require reliable discrimination of freezes from voluntary stops, potentially necessitating both validated kinematic wearables and multiangle video verification. Until such infrastructure is more readily available, the scalability of at-home movement classification frameworks will remain constrained. Nevertheless, the binary walking versus nonwalking paradigm already offers a practical and clinically actionable foundation for adaptive neuromodulation.

Compared to aDBS architecture that relies on external wearable sensors to detect gait and trigger stimulation changes, a totally embedded approach such as the one advanced in our study offers several benefits. First, leveraging neural signals allows for direct interfacing with the circuitry underlying gait control, enabling identification of intrinsic biomarkers of dysfunction and improvement that may be more closely linked to pathophysiology than external kinematics alone. For example, we have shown that GP beta power during double limb support correlates with improved walking performance and may inform aDBS titration (61). Neural signals also permit more complex multistate decoding (e.g., medication status and sleep) and can uniquely enable timely recognition of fast-evolving states such as FoG, where pre-state neural signatures may be critical. Beyond these advantages, a fully implantable, autonomous neural-signal based system avoids reliance on external hardware, reducing patient burden, compliance issues, and potential points of failure. Last, and most practically, the lack of integration between external sensors and current commercial aDBS devices limits the near-term feasibility of wearable-based approaches, although next-generation devices may benefit from incorporation of multimodal inputs.

Despite the ability to decode movement state in all hemispheres with high within-subject sensitivity and specificity, we observed considerable variation in the cortical-pallidal biomarkers enabling this performance. Differences in movement state biomarkers may be explained by variance in gait dysfunction severity, PD phenotype, or electrode placement (21). In addition, these biomarkers may be influenced by the amount of time spent in the various gait cycle phases by each subject as our approach averages across many individual cycles. Regardless, the variability observed within and between subjects reflects the real-world heterogeneity of patients with PD, underscoring the likely need for personalized biomarker identification and individualized thresholds. Biomarkers also showed a range of sensitivity and specificity across subjects. In practical implementation, these biomarkers will likely be fine-tuned with threshold modification due to the tradeoffs in optimization. In other words, inappropriately switching to lower-frequency settings while a patient is at rest sacrifices appendicular symptom control, and the opposite error may lead to substandard improvement in gait functions. To what extent either of these faults is tolerable may very well be patient- and impairment-dependent, which further emphasizes the value of a customizable pipeline for biomarker identification.

In this study, we also demonstrated the viability of an ambulatory rather than in-laboratory data collection pipeline. Advances in sensor technology have increasingly improved our ability to capture and study neural activity in naturalistic environments. Here, we illustrated the successful use of WDs to accurately identify walking states in subjects with a wide range of pathologic gait patterns. This permitted the collection of large datasets across a range of days for all subjects at home more efficiently than would have been possible under laboratory conditions (5). The ability to synchronize high-quality behavioral data with neural signals represents a major advancement in our ability to derive ecologically valid biomarkers for BCIs (76). Looking ahead, the proposed machine learning workflow can be implemented in new patients using currently available sensing-enabled DBS systems and wearable sensors, with further optimization as next-generation adaptive platforms mature. One critical bottleneck for widespread deployment remains the precise temporal alignment of multimodal data streams, which is essential for reliable model training. Continued progress in device interoperability and synchronized streaming architectures will therefore be key to clinical translation. Notably, our results demonstrate that accurate movement-state decoding can be achieved using neurophysiology alone, permitting use on-board existing aDBS-enabled neurostimulators. One limitation, however, is that commercially available systems currently stream and support a more constrained set of features than those used in our offline analyses, which may limit direct widespread translation in the near future. As next-generation platforms expand computational flexibility and bandwidth for real-time signal processing, the reach of this pipeline will also expand.

While our results are specific to individuals with PD, particularly in the context of gait dysfunction and DBS therapy, the overall pipeline including naturalistic neural and behavioral signal acquisition, feature extraction, and state classification is designed to be generalizable. The methodological framework for understanding neuroscience in the naturalistic extra-laboratory environment can, in principle, be applied to other movement disorders or neurological conditions where behavioral state–dependent stimulation may be beneficial, such as sleep disorders, pain syndromes, and even neuropsychiatric conditions. Beyond research, this development may also hold value for clinical applications, including for remote outpatient therapy optimization and expansion of care to medically underserved areas (77–79). While we focused on distinguishing binary walking versus nonwalking epochs, future research may expand to more complex or ambiguous states and latent microstates influenced by medication state, walking characteristics (e.g., duration and exertion), and circadian period using clustering approaches.

A primary limitation of our study is the small sample size, although it remains in line with other invasive surgical studies using investigational devices as we have used here. Future studies with larger cohorts and healthy controls are warranted to further elucidate supraspinal locomotor control mechanisms in neurotypical and pathological states. Another limitation of our study is the manual approach to neural and kinematic signal alignment, which is time-intensive and intractable for processing larger datasets. Considering the translational value of this work, the scalability of this alignment pipeline will be better addressed in the future by automatic/Bluetooth synchronization of external WDs with INS.

Overall, the identification of effective gait state biomarkers is critical to a variety of diseases. While the pipeline discussed here identifies that oscillatory signatures will enable DBS devices to address both appendicular and axial dysfunction in PD, our approach is generalizable to the many neurological diseases impairing movement. The pipeline discussed here offers a way to collect high-quality, longitudinal neural data in the real world. The insights gained from naturalistic data collection will advance therapy for PD and has the potential to accelerate BCIs across a multitude of debilitating conditions.

MATERIALS AND METHODS

Patient selection

Subjects with idiopathic PD were recruited from those being evaluated for DBS surgery at the University of California, San Francisco (UCSF) (2 male and 2 female, age range: 62 to 68 years, disease duration range: 4 to 21 years) (Fig. 2A). Inclusion criteria included the absence of major cognitive impairment, ability to comply with study follow-up visits, Unified Parkinson’s Disease Rating Score between 20 and 80 and an improvement of at least 30% in the baseline on-medication score compared to the off-medication score, and gait impairments (slowed gait, shuffling steps, postural instability, or freezing of gait). Subjects were assessed and diagnosed with PD by neurologists specialized in movement disorders. All participants provided written informed consent. Institutional Review Board (IRB) at UCSF provided formal ethical approval for this study (IRB number 20-32847). Throughout this study, all subjects received continuous conventional stimulation using their clinically optimized stimulation parameters; no other stimulation settings were tested or applied.

DBS implantation and electrode reconstruction

The subjects enrolled in this study underwent surgical implantation of quadripolar DBS leads into the GP (model 3387, Medtronic; contact length: 1.5 mm; intercontact distance: 1.5 mm) and subdural quadripolar paddles overlying the sensorimotor cortices (model 0913025, Medtronic; contact diameter: 4 mm; intercontact distance: 10 mm) inserted through the same skull opening. Two subjects were implanted unilaterally (left hemisphere), and two were implanted bilaterally. Electrodes were connected to investigational bidirectional implantable pulse generators placed in a superficial pocket over the ipsilateral pectoralis muscle (Summit RC+S model B35300R, Medtronic). Further details of the surgical implantation procedure have been reported previously (34, 80, 81).

Localization of depth and cortical electrodes was performed using advanced image processing pipelines (Lead-DBS, LeGUI) (Fig. 2B and fig. S1) (82–85). High-resolution postoperative computed tomography (CT) images were coregistered to preoperative T1-weighted 3T magnetic resonance imaging (MRI) scans using a rigid, linear affine transformation. Coregistration accuracy was visually verified and refined when necessary using an additional brain shift correction routine to align subcortical anatomy. Subcortical electrode artifacts were identified on postoperative CT images and matched to known electrode geometry. Cortical electrodes were projected onto MRI-rendered pial surfaces and manually adjusted if needed.

Summit RC+S device and neural recording preprocessing

The Summit RC+S device (model B35300R, Medtronic, Inc.) is an investigational, rechargeable, bidirectional INS that can stream four bipolar time domain electrode channels at a sampling rate of 500 Hz concurrent with delivery of therapeutic stimulation through a maximum of two quadripolar leads (Fig. 1A). The INS also contains an embedded accelerometer, which samples at 64 Hz. Further details of the Summit RC+S device can be found in previous publications (80).

Data from the Summit RC+S devices were extracted and analyzed using open-source code: https://github.com/openmind-consortium/Analysis-rcs-data. Neural signals from the GP, M1, and PM were bandpass-filtered from 1 to 150 Hz using a sixth-order Butterworth filter. Signals were examined for any stimulation or electrocardiogram artifacts, and template subtraction was used to identify and remove all instances of these artifacts: https://github.com/lhart1216/PerceptHammer. Epochs with Summit RC+S packet loss or idiosyncratic artifacts were excluded from analysis.

Rover WD

The Rover WD consists of a 38.1 mm by 51.2 mm by 13.7 mm, lightweight (53.86 g) rechargeable inertial sensor module paired with a fabric strap, allowing the device to be worn around the ankle (Fig. 1B). The nine-axis motion sensor module contains a triaxial gyroscope, triaxial accelerometer, and triaxial magnetometer, which sample at 100 Hz and record data locally on a secure digital card. WD data were uploaded by subjects to the secure Rover cloud. Raw WD data were extracted and formatted with custom code.

Once a WD recording was uploaded to the cloud, the Rover Gait Analysis System was used to generate a multiple-page gait analysis (MPGA) spreadsheet. Each MPGA summarized activity details (e.g., recording duration, total walking time, total distance walked, etc.) for each recording. In addition, it provided a list of time-stamped strides for both legs, with gait cycle time (in seconds), length (in centimeters), swing period (in seconds), and heading (in degrees) recorded. MPGA reports were used to calculate average gait metrics for all subjects (table S1).

Validation of WD movement state identification

We performed independent validation of the Rover WD’s ability to accurately identify movement states in our cohort using kinematic data from observed walking and nonwalking periods collected in a controlled environment. Patients were equipped with the Rover WD and Delsys sensors, which consist of two Avanti FSR adapters, two Avanti goniometer adapters, and two Trigno surface electromyography (EMG) sensors with built-in accelerometers. The goniometer adapters were placed on the shanks bilaterally. EMG sensors were placed on top of the Summit RC+S INS. FSR adapters were each attached to four FSRs (model DC:F01, Delsys) under the calcaneus, hallux, first metatarsal, and fifth metatarsal. Under supervision, subjects performed overground walking loops with interposing periods of standing, arm swing, or seated rest, during which data were recorded from the Summit RC+S, Rover WD, and Delsys devices. See fig. S2A for sample trials for all subjects.

Following data collection, signals from these three devices were aligned with peak-to-peak matching across acceleration signals from the ipsilateral Summit RC+S INS, Rover WD, and Delsys Trigno EMG accelerometer using a custom-built graphical user interface (GUI) in MATLAB (version 2021b, Mathworks, Inc., Natick, MA, USA). Gait cycles were identified using Delsys FSR data; individual cycles were defined as the time between two consecutive heel strikes (calcaneal FSR signal crossing 5% threshold in positive direction). Each recording of kinematic data was partitioned into 1-s epochs, and each epoch was labeled as a walking period if it contained any portion of a gait cycle and nonwalking if it contained no portions of any gait cycle. These labels were verified with review of video recordings taken of subjects’ activity during these sessions and manually adjusted if necessary.

To then compare Rover WD–based labeling of movement state with the ground-truth labels, we classified all 1-s epochs using the time-stamped stride lists from the Rover MPGA. Epochs composed of any left or right leg strides, determined by calculating the overlap of stride times and epoch times, were considered walking. Epochs with no overlapping left and right leg stride times were considered nonwalking. WD-derived labels were then compared to ground-truth labels; mean accuracy, sensitivity, and specificity were calculated for each subject (fig. S2B).

Neural and kinematic signal alignment

Subjects streamed neural data while wearing the Rover WD during normal activities of daily living several times weekly, as continuously as possible. Data from Summit RC+S and WD devices for each subject were uploaded to a secure cloud folder accessible only by the research team. Summit RC+S and WD recordings from the same days were identified.

A custom-built GUI in MATLAB was used to manually align x-axis acceleration signals from the Summit RC+S and bilateral Rover WD by matching acceleration peak to peak across the devices. All alignments were reviewed extensively to ensure synchronization between Summit RC+S and Rover acceleration signals.

Movement state labeling

Each INS-WD alignment was partitioned into continuous 10-s epochs. The associated time-stamped MPGA stride lists and WD accelerometry signals were used to label each epoch with complete data as walking or nonwalking. Epochs composed of greater than 50% (5 s) of left and right leg strides, determined by the overlap of stride times and epoch times, were considered walking. Epochs with no overlapping left and right leg stride times were labeled as nonwalking. Remaining epochs were excluded from labeling and analysis as these periods were considered transition periods.

Canonical frequency band analysis

Neural time-domain data from each 10-s epoch were converted to the frequency domain using Welch’s method (MATLAB “pwelch” function; 1-s windows with 50% overlap). Average power was calculated within canonical frequency bands: delta (δ; 1 to 4 Hz), theta (θ; 4 to 8 Hz), alpha (α; 8 to 13 Hz), beta (β; 13 to 30 Hz), and low gamma (low γ; 30 to 50 Hz). Two-sided Wilcoxon rank-sum tests were used to compare power within canonical frequency bands between walking and nonwalking epochs.

To assess the relative importance of each canonical frequency band to within-subject discrimination of walking and nonwalking epochs, logistic regression models were trained and tested within 10-fold cross-validation (70/30 train-test split) using average power within canonical frequency bands from GP, M1, and PM signals (totaling 15 features per model). All features were standardized using z-scoring within each frequency band. AUC was calculated for models using class probabilities. For subjects with bilateral implants, separate models were generated for each hemisphere. Model coefficients were linearly rescaled from −1 to 1 and plotted on colored heatmaps to visualize relative feature importance.

Personalized gait biomarker identification

While canonical frequency bands offer the advantage of standardized comparison across individuals, they also carry several limitations including the potential masking of narrowband peaks (35). We therefore used a data-driven approach to identify individualized frequency bands that may serve as potential endogenous biomarkers for movement state classification. To generate new candidate bands, we calculated average power within all bands of varying sizes from 1 Hz [e.g., (1 to 2 Hz), (2 to 3 Hz), etc.] to 49 Hz [i.e., (1 to 50 Hz)], thereby generating 1225 features per region.

RF models were used to identify the most important frequency bands, as RF has been shown to overcome limitations of other feature selection algorithms including sensitivity to collinearity and high dimensionality (86, 87). Feature importance was calculated for each frequency band as the mean decrease in impurity (MDI), a metric measuring the increase in subset homogeneity, or Gini impurity, when each variable is used for node splitting. For each node, the Gini impurity was calculated as follows, where p_i is the proportion of samples belonging to class c

$$I_{\text{Gini}}=1-{\displaystyle \sum _{i=1}^c}{p}_i^2$$ (1)

Mean feature importance was calculated from 1000 RF iterations. We visualized feature importance for each hemisphere by rescaling the mean importance of each variable from 0 to 1 and plotting these values as overlapping bars for each frequency band.

We next turned to classification of movement state using these new frequency bands. LDA models were constructed to classify movement state for each subject using cortical-pallidal neural activity. LDA models were chosen as they generate feature weights that can be entered into the Summit RC+S on-board classifier when embedding adaptive architecture. Several types of LDA models were tested: single-region models using signals from one region (GP, M1, and PM), multiregion models using signals from two regions (GP+M1, GP+PM, and M1+PM), and complete models using signals from all three regions (GP+M1+PM).

Each LDA model was trained and tested with 10 iterations of stratified Monte Carlo cross-validation; for each fold, the model was trained with 70% of the data and tested on the remaining 30%. Data were stratified to balance the representation of walking and nonwalking epochs (50/50) in the training and testing sets. All features were standardized using z-scoring. Mean accuracy, AUC, sensitivity, specificity, and PPV were calculated.

Offline testing of on-board classification using at-home biomarkers

The implementation of closed-loop aDBS architecture requires adherence to the practical constraints of embedded classifiers. Specifically, the Summit RC+S device transforms neural time-domain data to the frequency domain and internally calculates power features in a distinct proprietary manner, which has been previously discussed and shown to be reliably replicable offline. In addition, the LDA classifier on-board the Summit RC+S device is limited to a maximum of four features with no more than two features from a single channel. As a result, the unconstrained models tested in previous analyses, while useful for understanding locomotor networks, are less informative for on-board classifier programming. Consequently, to test the ability for at-home cortical-pallidal activity to be used for generating aDBS-compatible biomarkers for practical implementation, we performed offline replication of on-board gait-state classification with device-constrained biomarkers identified using at-home data.

To test on-device decoding performance, we used the rcssim package, which enables in silico simulation and visualization of on-board feature calculation and LDA classification, using raw neural time-domain data and decoder programming parameters (i.e., frequency bands, channels, feature weights, and LDA threshold) as input: https://github.com/Weill-Neurohub-OPTiMaL/rcs-simulation. First, epochs were shortened to 5-s intervals to match the more rapid changes necessary for real-world adaptive stimulation, and power features were recalculated using rcssim computations mimicking Summit RC+S on-device conversions (see Supplementary Text for details). Next, we assessed on-board classification of at-home movement state by training LDA models on subsets of at-home data to calculate the necessary decoder input parameters and testing simulated decoder performance on held-out at-home epochs (stratified Monte Carlo cross-validation; 10 iterations; 70/30 train-test split). To further validate classification performance, we repeated this procedure by training models on at-home data but testing simulated decoder performance on recordings of in-laboratory trials of overground walking loops with interposing periods of standing or seated rest (stratified Monte Carlo cross-validation; 10 iterations; 70/30 train-test split). In all cases, we tested biomarkers composed of one to four input features; feature selection was guided by the RF feature importance values discussed in an earlier section.

General statistical methods

Correction for multiple comparisons

In analyses with multiple comparisons (e.g., comparison of power within canonical frequency bands between movement states), we accounted for the false discovery rate using the Benjamini-Hochberg procedure (88).

Empirical P value calculation

For each classification model tested, we characterized chance-level performance by generating 1000 (n) permuted null models of randomly shuffled movement state labels and calculating AUC for each model. We then calculated the number of permuted AUCs greater than each LDA or logistic regression model AUC (k) and corrected this empirical one-sided P value with the following adjustment similarly used in previous studies to avoid underestimation (13, 89)

$$p=\frac{k+1}{n+1}$$ (2)

Ethics statement

This study was reviewed and approved by the IRB of record at the UCSF, under the approval number 20-32847. The UCSF IRB conducted a full board review and granted approval, confirming that the study met all ethical requirements.

Supplementary Materials

This PDF file includes:

Supplementary Text

Figs. S1 to S4

Tables S1 to S7

References