Decoding and Restoring Movement in Parkinson’s Disease

## Predict movement, then test the interpretation Gait involves both legs, balance and changing context. A single feature such as beta power may miss useful information. **Neural decoding** uses recorded signals to estimate a measured aspect of behavior; it does not mean reading thoughts. In Choi et al. (2025), **N2GNet**, a neural-to-gait model, used bilateral STN LFPs from eighteen people performing harnessed stepping in place with stimulation off. Five seconds of neural input were used to estimate a force-plate weight-shift measure over the last two seconds. These estimates concern a stepping measure, rather than a person’s intentions or an unrestricted prediction of daily walking. Models were trained for each person on an early visit, validated later, and tested on the latest visit. This separation matters: performance on training data would not show that a model can handle a later recording. Estimates tracked stepping more closely than beta power from either lead alone. Frequency analyses highlighted beta alongside contributions from other bands. Prediction is not a mechanistic explanation. Different models or signal features may support similar estimates. N2GNet was tested offline with stimulation off; stimulation-related changes would need to be handled before such a decoder could guide therapy. ## Brain and movement can both provide biomarkers **Neural and kinematic biomarkers for gait** offer complementary information. IMUs measure step timing and rotation. Freezing detection models ask whether sensor patterns identify an episode; they do not demonstrate that an intervention will prevent it. A 2023 report used shin-sensor measures of irregular stepping to drive adaptive DBS in one participant. It demonstrated feasibility during that test, with no freezing episodes observed in that condition, but also delivered slightly more total stimulation. A single-person result does not establish benefit for everyone. Cui et al. (2026) reported four weeks each of adaptive and continuous stimulation in one person at home, using a device unlocked for research. Beta power varied across the day, and stimulation tracked it. Daily QDG mobility measures stayed stable during adaptive stimulation and declined slightly during continuous stimulation. This is a case report, not a universal comparison. The source package also discusses the larger, nonrandomized ADAPT-PD trial in people previously stable on continuous DBS. ## Personalization needs real-life testing Multimodal control could combine brain signals, wearable movement measurements and context. It remains a research direction requiring validation. Cognitive-motor research asks how attention and walking interact. The source package describes an investigational nucleus basalis of Meynert pilot alongside STN stimulation; it reports no established cognitive benefit. The course chain is **measure the person → measure the brain → find the signal → build the algorithm → adapt the therapy → test it in real life**. Each arrow requires evidence, not simply a promising prediction. ## Source trail See `sources/bs3-1.md` and the claim audit: Choi (2025), Melbourne (2023), Cui (2026), ADAPT-PD (2025), and NCT05968703.