Parkinson’s disease (PD) is a progressive neurodegenerative disorder caused by moto impairments which include tremor, rigidity, bradykinesia and significant gait disturbances like freezing of gait and postural instability. While deep brain stimulation (DBS) has become an established treatment to manage motor symptoms, conventional DBS delivers continuous stimulation without adapting to fluctuations in behavioral state.
Emerging bidirectional neurostimulators enable simultaneous stimulation and neural sensing, creating the possibility of adaptive DBS (aDBS) systems which dynamically adjust therapy based on real-time biomarkers. To detect reliable neural signatures of naturalistic behaviors like walking in real-world environments remains a critical challenge.
The published study was aimed to determine whether neural oscillatory activity recorded from globus pallidus (GP) and sensorimotor cortex could reliably distinguish walking from non-walking during both laboratory and at home conditions.
A secondary objective was to assess whether such decoding could be implemented in technical constraints of an implantable bidirectional device which enables real-world closed loop stimulation.
4 individuals with clinically significant PD (2 male, 2 female) undergoing DBS evaluation were implanted with a bidirectional investigation neurostimulator capable of sensing and stimulation. 2 participants received unilateral implants, and 2 received bilateral implants which result in 6 hemispheres for analysis. Electrodes targeted the GP and cortical motor regions (primary motor cortex and premotor cortex).
Participants remained on clinically optimized continuous stimulation throughout the data collection. Wearable ankle accelerometers were used to detect gait to label behavioral states. Validation experiments conducted in controlled laboratory sessions as compared wearable-derived gait labels with synchronized inertial navigation and force-sensitive resistor measurements, which showed high accuracy.
Gait detection accuracy exceeded 95% (range 95.8% to 99.0%) with sensitivity between 94.4% and 98.9% and specificity between 94.7% and 100.0% which confirms reliable movement-state labelling.
Chronic at-home streaming yielded 84.5 total hours of synchronized neural and kinematic recordings collected over about 13 days per participant. Neural data were segmented in 10 second epochs and categorised as walking or non-walking. Spectral analyses revealed consistent movement related modulation of cortical oscillations.
Primary motor cortex alpha (8 to 13 Hz) and beta (13 to 30 Hz) power were significantly reduced during walking compared with non-walking (P ≤ 10⁻⁴) in all hemispheres. It was consistent with movement-related desynchronization. Low gamma (30 to 50 Hz) power was significantly reduced during walking in most hemispheres (P ≤ 10⁻⁹). Pallidal activity showed greater interindividual variability.
Delta and theta power increased significantly during walking, while beta power often decreased in many hemispheres, For example, in one subject, GP delta, theta, alpha and low gamma power increased during walking (P ≤ 10⁻²⁷), whereas beta decreased which highlight subject-specific oscillatory patterns.
Machine learning classifiers were applied to assess movement-state decoding performance. Logistic regression models by using canonical frequency bands achieved significant classification in all hemispheres (empirical permutation P < 0.001). To improve personalization, a data-driven random forest approach identified the most informative frequency bands between 1 to 50 Hz.
Pallidal signals emerged as most influential features in 4 of 6 hemispheres specifically in delta, theta and beta bands. Linear discriminant analysis models built by using individualized frequency bands achieved strong discrimination performance with area under the curve (AUC) values ranging from 0.67 to 0.98 (P < 0.001). Sensitivity ranged from 74.8% to 95.8%, specificity from 66.9% to 94.0% and positive predictive value from 69.7% to 94.1%.
In several hemispheres, GP-only models achieved highest AUC (0.86 to 0.98) which indicate that pallidal signals alone were often sufficient for accurate decoding. At-home classification achieved AUC values between 0.64 and 0.81 (P < 0.001) with sensitivity 66.0% to 83.3% and specificity 51.6% to 69.1% under these constraints.
When validated in independent in-laboratory walking sessions, AUC ranged from 0.63 to 0.87 (P < 0.001) with sensitivity up to 90.2% and specificity up to 81.3%. These findings demonstrate that personalized neural biomarkers derived from chronic recordings can be implemented in real-world adaptive DBS systems.
This study provides the first in human demonstration of reliable at home movement state decoding by using chronic multisite intracranial recordings in PD. Walking is consistently linked to cortical alpha and beta desynchronization and subject-specific pallidal oscillatory changes.
Personalized and data driven biomarker selection significantly improves classification performance, and clinically feasible device compatible models retain meaningful accuracy. These results support the development of embedded, closed loop adaptive DBS systems capable of switching stimulation parameters based on real-time behavioral state.
The study establishes a critical step toward personalized, ecologically valid neuromodulation strategies for Parkinson’s disease and other movement disorders.
Reference
Ramesh R, et al. At-home movement state classification using totally implantable cortical-basal ganglia neural interface. Sci Adv. 2026;12:eadz4733. doi:10.1126/sciadv.adz4733
