Neural & myoelectric interface decoding (BCIs, intelligent prostheses)

People with limb loss, spinal-cord injury, stroke or ALS cannot express intent through their normal motor channels. ML decoders read the residual signals (muscle EMG or neural recordings), infer the intended movement or speech in real time and drive a prosthesis, a communication system or a stimulator. Myoelectric pattern recognition is already commercial; implanted brain-computer interfaces (BCIs) are in clinical trials with striking but small-cohort results.

Problem & value

• Today: conventional myoelectric prostheses use two-site threshold control (one muscle opens, another closes) and switch modes by co-contraction, which is slow and tiring. Many users abandon the device. People with locked-in syndrome rely on eye trackers or switches, which are slow and fatigue-prone.
• AI value: classifiers and regressors on multi-channel EMG let users select among several grips or control several degrees of freedom without mode switching. Intracortical decoders translate attempted speech or handwriting to text; BCI-triggered spinal stimulation has restored walking-related control in a person with chronic tetraplegia.
• Evidence (peer-reviewed, small cohorts): a speech neuroprosthesis for a participant with ALS reported accurate, rapidly calibrating decoding of attempted speech [1]; a second study decoded attempted speech at high rates from intracortical arrays [2]; a wireless "brain-spine interface" linked cortical decoding to epidural spinal stimulation in a person with tetraplegia (CH/FR) [3]. These are feasibility studies, not proof of broad clinical benefit. Do not extrapolate accuracy figures to the general population.

Variants

• Pattern-recognition myoelectric prostheses (commercial): upper-limb prostheses with multi-electrode EMG and a classifier (Coapt Complete Control, Ottobock Myo Plus) [4].
• Targeted muscle reinnervation (TMR) + decoding: surgical rerouting of nerves gives richer EMG sources for decoders.
• Implanted speech and communication BCIs: intracortical arrays (BrainGate-type, Blackrock arrays), ECoG, endovascular stentrodes (Synchron).
• BCI-coordinated stimulation: decoded movement intent triggers spinal-cord or functional electrical stimulation (ONWARD ARC-BCI, NL/CH, EU-based).
• Exoskeleton and rehabilitation control: intent-based triggering of powered orthoses; EEG or EMG-gated neurorehabilitation after stroke.
• Non-invasive EEG BCIs: lower bandwidth, used for rehabilitation and research; consumer EMG wristbands for HCI are a related but non-medical use.
• Not covered: monitoring-only biosignals (HLTH-007, HLTH-013, HLTH-035).

How it works

1. electrodes (EMG / ECoG / intracortical)
2. amplification, filtering, artefact rejection
3. feature extraction (RMS, spectral power, spike-band power, threshold crossings)
4. decoder (LDA/SVM classifier | Kalman filter | RNN / transformer)
5. intent estimate
6. language prior for speech (n-gram + LLM rescoring)
7. device command (grip, cursor, stimulation, voice output)
8. closed-loop feedback to the user
9. online recalibration (human-in-the-loop)

• Decoders: classic pattern recognition (LDA, SVM) is robust and cheap on embedded hardware; speech and handwriting BCIs use recurrent or transformer networks trained on days to weeks of cued sessions, followed by LLM or n-gram rescoring of phoneme sequences [1][2].
• Non-stationarity: electrode shifts, sweat, socket changes and neural drift degrade decoders. Recalibration, unsupervised adaptation and transfer learning across days are the central research topic.
• Latency budget: the closed loop must stay within roughly a few hundred milliseconds end to end, so inference runs on the device or a bedside unit, not in the cloud.
• Human-in-the-loop: clinician and prosthetist set up and validate the decoder; the user can always override; safety limits on actuator force and stimulation amplitude are hard-coded outside the ML component.

Data preparation

• Data and location: EMG, EEG or intracortical recordings from lab acquisition systems and device logs, calibration sessions with cued movements or speech, IMU data, clinical outcome scores, and pre-injury voice recordings. In practice these sit on research workstations, device manufacturer clouds or hospital servers, rarely in ERP or DMS.
• Typical gaps: very small cohorts, per-user and per-session variation, electrode drift, inconsistent file formats and sampling rates, missing metadata on electrode placement and session conditions, artefacts, and unclear consent scope for secondary use.
• Preparation steps: define a session and metadata schema (PREP-019); convert vendor formats to open standards such as BIDS or LSL streams (PREP-007); label calibration cues and rejected artefact segments (PREP-020, PREP-018); keep held-out sessions from later days as the test set; store features rather than raw signals where possible and pseudonymise participants (PREP-022); document consent and purpose (PREP-025).
• Tools: BrainFlow, MNE-Python, Lab Streaming Layer, MOABB datasets.
• Effort share: about 50 percent, mostly consistent, well-annotated calibration data and data-lineage documentation required for MDR. See GUIDE-003.

Implementation paths

Which path for whom

• Enterprise: Path C. Device makers and research groups build decoders on open acquisition and ML tooling with an embedded runtime; custom training (Path D) is added for each user and clinical use needs full MDR conformity.
• Developer: Path C. BrainFlow, MNE-Python, MOABB and PyTorch cover prototyping with healthy volunteers.
• Public: Path A. Public clinics, rehabilitation centres and insurers should buy certified, fitted prostheses; implanted BCIs are accessible only through trials.

Path A — Buy (right for patients and clinics)

• Commercial prostheses: Coapt Complete Control (US) and Ottobock Myo Plus (DE, EU), plus other EU/UK bionic-hand makers (e.g. Össur, Touch Bionics/PSYONIC-class products). Fitted by certified prosthetists; reimbursement varies by country.
• Implanted BCIs: only via clinical trials (BrainGate, Synchron, Neuralink, Precision Neuroscience, ONWARD ARC-BCI). No general market product for speech BCIs in the EU as of 2026.
• Choose A when: the goal is patient care. Research-grade decoding is not a purchasable service.

Path B — Build on cloud APIs

Limited role. Cloud is unsuitable for the control loop (latency, safety). Use cloud/EU-region services for offline model training, fleet analytics, and an LLM rescoring or voice-synthesis stage (TTS, personalised voice from pre-injury recordings) where a short delay is acceptable. Patient data must stay in EU regions under a DPA.

Path C — Self-host (open source)

• Acquisition and tooling: BrainFlow, MNE-Python, Lab Streaming Layer, OpenBCI hardware for non-invasive work; MOABB for EEG benchmarking [5]; Neural Latents Benchmark datasets for intracortical decoding research.
• Models: PyTorch decoders (GRU/transformer), open-weight small LLMs or n-gram toolkits (KenLM) for language priors, ONNX/TensorRT/TFLite for the embedded runtime.
• Hardware: HW1 (embedded CPU/NPU or Jetson-class) for the real-time loop; a workstation GPU for training.
• Use: research, prototyping, device development; clinical use requires full MDR conformity.

Path D — Custom training

This is the core of the field. Each user needs personal training data (cued attempted movements or speech, hours to weeks). Manufacturers add pooled pre-training across users to shorten calibration. Requires ethics approval, IRB/EC-reviewed protocols, GCP-compliant trials (ISO 14155 for devices) and long-term follow-up.

State of the art (2026)

• Speech neuroprostheses decode attempted speech from motor-cortex arrays into text and synthesised voice with large vocabularies and short calibration; results come from a handful of participants [1][2].
• Brain-spine interfaces couple cortical decoding with spinal stimulation, with wireless implants [3].
• Myoelectric control is moving from classification toward continuous regression of multiple degrees of freedom, with sensory feedback (vibrotactile, neural) in research.
• Foundation-model ideas: pre-training neural decoders across sessions and subjects to reduce calibration; evidence is still mostly academic.
• Limits and failure modes: signal drift, electrode degradation and infection risk for implants, small cohorts, lab-bound set-ups, caregiver burden for calibration, decoders that fail silently when signals change, and the gap between lab accuracy and daily-life reliability. Abandonment of myoelectric prostheses remains a real problem.

Shared resources & combinations

• Same stack (time-series + streaming + edge-runtime): HLTH-007 (wearable analytics), HLTH-013 (ECG signals), HLTH-035 (EEG and sleep analysis) share filtering, artefact rejection and streaming code.
• Downstream and adjacent: HLTH-034 (closed-loop insulin delivery) shares the regulatory and safety-case pattern for AI in the control loop; HLTH-037 (robotic surgery) for intent-driven actuation; VIS-024 (pose and gesture) for camera-based rehab assessment.
• Remote follow-up: HLTH-004-style clinician dashboards for usage logs and drift alerts.

Effort, cost & KPIs

• Path A: prosthesis fitting costs are in the range of a medical device purchase plus therapy and training; reimbursement by health insurers is country-specific.
• Path C: months for a research prototype with off-the-shelf hardware.
• Path D: years and multi-million-euro budgets for an implanted system (device, GMP manufacturing, trials, notified-body review). Do not plan below this order of magnitude.
• KPIs: online decoding accuracy and word error rate (speech), information transfer rate (bits/min), task completion time (e.g. Box and Block, SHAP test for hands), latency, daily usage hours, recalibration frequency, device abandonment rate, adverse events, user-reported quality of life.

People & change

• Involve early: management, the clinical and engineering leads, prosthetists and therapists who fit and tune the system, security and legal/compliance (MDR, AI Act, cybersecurity), the DPO, ethics committee and notified body. The most important stakeholders are the users, patients and their caregivers; involve them in design and calibration from the start.
• Works council: no, in general. The system is used by patients, not to monitor employees (§87(1) no. 6 BetrVG does not apply); inform the works council only if staff-level usage data of clinicians or engineers is evaluated, and note §90(1) no. 3 for AI-supported work processes (GUIDE-004).
• Data protection: neural and EMG recordings are health and potentially biometric data (Art. 9); the DPO, explicit consent, a DPIA (Art. 35) and an AVV with every processor are required (GUIDE-005). Informed consent for people with communication impairments needs special care.
• AI literacy: take measures to support AI literacy (Art. 4) of clinicians and prosthetists: drift, recalibration and limits of lab accuracy (GUIDE-006).
• Pilot and communication: start with healthy volunteers using EMG, then supervised clinical studies (GUIDE-009). Typical concerns are reliability, abandonment, cyber attacks on implants and long-term vendor support; address them with hard-coded safety limits, a user-controlled stop, honest communication of cohort-size limits and long-term support commitments (GUIDE-008, GUIDE-013).

Risks, regulation & compliance

• EU AI Act: high-risk. A decoder that is a safety component of, or is itself, a medical device under MDR (Class IIb/III for active implants and many stimulators) falls under Art. 6(1) with Annex I, so Chapter III obligations apply (risk management, data governance, logging, human oversight, accuracy and robustness), integrated with the MDR conformity assessment by a notified body. Consider the transition timetable and any amendments (Digital Omnibus) when planning. Art. 5 bears watching: do not use neural or biosignal data to infer emotions of workers or students (prohibited, except for medical or safety reasons), and avoid neuro-marketing style uses.
• MDR (EU) 2017/745: implantable and active devices are among the strictest classes; clinical investigation under Art. 62 ff., ISO 14155, ISO 14971, IEC 62304, IEC 60601, cybersecurity (MDCG 2019-16) and post-market surveillance [6].
• GDPR: neural and EMG recordings are health data and potentially biometric data (Art. 9); explicit consent, DPIA, strict purpose limitation; neural data can reveal more than intended (mental privacy debate, e.g. neurorights legislation outside the EU). Data minimisation: store features rather than raw signals where possible.
• Other risks: cybersecurity of implants (remote attack could harm the user), vendor lock-in or company failure leaving implanted users without support, informed consent for people with communication impairments, equity of access.
• Mitigations: hard-coded safety limits independent of ML, fail-safe default behaviour, logged decoder confidence, user-controlled stop, long-term support commitments.

Getting started

1. Define the target population and the control task (grip selection vs. speech vs. stimulation trigger); this decides the regulatory class.
2. For research: acquire an open dataset (MOABB EEG sets, Neural Latents Benchmark) and reproduce a baseline decoder offline.
3. Build a real-time loop on BrainFlow or LSL with latency measurement, then test with healthy volunteers using EMG.
4. Involve a notified-body consultation, clinical partners and an ethics committee early; draft the MDR intended purpose and risk file before building.
5. Plan the recalibration workflow and the user-support model as part of the product.

Sources

1. Card N. et al., "An Accurate and Rapidly Calibrating Speech Neuroprosthesis", NEJM 2024 — https://www.nejm.org/doi/full/10.1056/NEJMoa2314132
2. Willett F. et al., "A high-performance speech neuroprosthesis", Nature 2023 — https://www.nature.com/articles/s41586-023-06377-x
3. Lorach H. et al., "Walking naturally after spinal cord injury using a brain-spine interface", Nature 2023 — https://www.nature.com/articles/s41586-023-06094-5
4. Coapt Engineering — Complete Control pattern recognition — https://coaptengineering.com
5. MOABB (Mother of All BCI Benchmarks) — https://github.com/NeuroTechX/moabb
6. Regulation (EU) 2017/745 (MDR) — https://eur-lex.europa.eu/eli/reg/2017/745/oj
7. Regulation (EU) 2024/1689 (AI Act) — https://eur-lex.europa.eu/eli/reg/2024/1689/oj
8. BrainFlow — https://github.com/brainflow-dev/brainflow