Long-term wireless streaming of neural recordings for circuit discovery and adaptive stimulation in individuals with Parkinson's disease

Nature Biotechnology 39:1078–1085, doi 10.1038/s41587-021-00897-5; NIH author manuscript, PMC8434942 (NIHMS1684406; the manuscript title says "patients with"). Read for research direction R1, 4 October 2026. Provenance: papers/amadeus/gilron2021_chronic_streaming.provenance.json.

What was read

• Read in full: the PMC author manuscript (NCBI efetch XML converted to text): abstract, introduction, results, discussion, Methods, Extended Data Figures 1–7 captions, all main figure captions, Table 1, ethics, data and code statements, and references.
• Not read: the Supplemental Information (surgical details, in-clinic statistics, overfitting controls, data-continuity procedures, adaptive-DBS risk management), the supplementary video and the Reporting Summary. None is in the XML. Figures were not inspected as images.

Design

• Who: five people with Parkinson's disease and motor fluctuations (four men; aged 28–63), referred for deep brain stimulation (DBS). Cognitive impairment and untreated mood disorder were excluded. UCSF IRB, an FDA investigational device exemption (G180097) and ClinicalTrials.gov NCT03582891; protocol on OSF (ya5jf).
• Implant: on each side,
  • a quadripolar DBS lead in the subthalamic nucleus (STN);
  • a quadripolar subdural paddle over primary motor cortex (4 mm contacts, 10 mm apart);
  • an investigational Medtronic Summit RC+S pulse generator in the chest.
• Device capability:
  • Channels: 16 channels; streams four bipolar time-series channels at 250 or 500 Hz, or two at 1,000 Hz. Also up to eight on-device band-power channels and an accelerometer.
  • Transmission: to a pocket relay carried by the patient (191 g), then by Bluetooth to a Windows tablet within about 12 m.
  • Battery: "up to 30 hours" of sensing per generator before recharge (Figure 1 caption).
  • Losses: 1–5% of data packets are dropped even in range.
  • Software: investigators must write their own device-control software to FDA design-control standards; an open consortium shares it (openmind-consortium.github.io).
• Recording:
  • Before therapy: 2–4 weeks after surgery and before therapeutic stimulation, patients streamed at home in 1–2-week "sprints", told to carry the tablet and "stream continuously if possible". Most data were at 250 Hz; four bipolar channels per side, eight in total.
  • Total: 2,655 hours of 8-channel data over up to 15 months. Of these, 2,142 hours were at home before stimulation and 1,502 hours during sleep.
• Labels:
  • Wrist monitor: a clinically validated actigraphy watch (Parkinson's KinetiGraph) worn on both wrists scored bradykinesia and dyskinesia every 2 minutes. Ten-minute windows were labelled "on" (mobile), "off" (immobile), sleep, or "other" when fewer than 60% of epochs agreed.
  • Self-report: a patient app recorded medications taken and self-rated motor signs. Motor diaries were used during the adaptive-stimulation trial.
  • Synchronisation: neural and watch data were aligned through the implant's accelerometer.
• Analysis:
  • Features: 30 s Welch spectra and STN–cortex coherence, averaged into 10-minute windows.
  • Tests: per-person rank-sum tests of on against off in alpha, beta and gamma bands.
  • Decoding: per-person linear discriminant decoders, 5-fold stratified cross-validation, label shuffling 10,000 times, Bonferroni.
  • Unsupervised: density-peak clustering, and template matching to in-clinic on and off spectra.

Main results (verified)

• Person-specific "fingerprints": which bands and sites separated mobile from immobile states differed by patient. STN beta disappeared and cortical gamma near 75 Hz appeared when "on" with dyskinesia (one patient, 42.7 h of labelled data). STN–cortex coherence was often the best discriminator.
• Decoding motor state at home during daily life: within-person decoders reached AUC 0.81–1.0 when both sites and coherence were combined; single-site features worked in most but not all hemispheres. Accuracy tended to rise with the severity of a patient's fluctuations (p = 0.065). The patient with the largest fluctuations was decoded best.
• Unsupervised clustering recovered states that agreed with the watch labels in example patients (Extended Data Figure 5, caption only).
• Sleep changes everything: it suppressed STN beta and cortical gamma and raised cortical low-frequency activity. The authors say this is "impractical to obtain from brief in-clinic recordings".
• Adaptive stimulation at home: fully embedded adaptive DBS, driven by STN beta in one patient and cortical gamma in another, ran for 1 and 4 days. It increased "on" time without dyskinesia compared with standard stimulation, on watch and diary measures. Not a formal trial.
• Data: "available from the corresponding author upon reasonable request". Code is public on GitHub.

Limits

• Five patients with a movement disorder; the decoded state is a motor symptom with large, medication-driven swings and an objective wearable label. Nothing cognitive or affective was decoded.
• Coverage is two bipolar channels per site per hemisphere: a few square centimetres of motor cortex and one deep nucleus.
• Streaming needed the tablet within about 12 m, the relay carried on the body, recharging, and patient effort. Data came in sprints, not as an unbroken record.
• The devices were provided free by the manufacturer (Medtronic). Some authors hold a patent on cortical detection of movement-disorder biomarkers.

What it means for Amadeus

• Chronic implants can record a person at home for months (verified), but only where a therapy justifies the implant. The coverage is tiny (eight bipolar channels here), and continuity depends on the person carrying hardware and recharging. That is the honest form of "electrodes for a month": patients, not volunteers, and very sparse.
• The recipe that worked is the one brief 6.7 proposes (verified for motor state): a long unlabelled stream, plus a frequent, independent label stream (a watch scoring every 2 minutes), plus within-person models. With weeks of data per person, person-specific decoders were accurate. Two conditions mattered: the state fluctuated strongly and often, and the label was objective and dense.
• Transfer to thought is the open question (inference). Spontaneous thought has no watch. Its label comes from sparse experience-sampling probes, and its neural signature is not concentrated in two contacts. The decodability shown here should not be read as evidence that a month of recording would decode thinking.
• Sleep dominates a continuous record (verified numbers; the share is our arithmetic).
  • The 1,502 "asleep" hours, defined as 10 pm to 8 am, are all from the pre-stimulation period: Extended Data Figure 2 counts sleep only before stimulation. That makes them about 70% of the 2,142 pre-stimulation home hours.
  • Streaming was evidently easiest when the person stayed near the tablet (inference).
  • A month-long continuous stream will therefore be weighted towards sleep and quiet rest, and decoders must condition on sleep and arousal state before anything else.
• E3 (verified): data on request only, motor state, no recall. Not an E3 candidate.