From residual nerve/muscle signal to real-time prosthetic motor command — targeted muscle reinnervation, multi-channel EMG, and pattern-recognition myoelectric control
After limb amputation, the peripheral nerves that once controlled the hand and forearm are still alive and still carry motor intent signals from the brain — they simply have nowhere to go. Targeted Muscle Reinnervation (TMR) surgically redirects these nerves into nearby, otherwise nonfunctional residual muscle, turning that muscle into a biological amplifier that produces a strong, localized EMG signal every time the patient thinks about moving their missing hand.
A transhumeral or transradial amputation severs major peripheral nerves along with the muscles they once controlled. The proximal nerve stump is still fully functional — axons still fire in the pattern the brain intends — but with no muscle fibers left nearby to innervate, that signal cannot be picked up cleanly by surface electrodes. What little EMG remains from residual stump muscle is often weak, cross-talking between adjacent muscle groups, and insufficient to reliably separate more than 2–3 movement classes.
TMR solves this by microsurgically transferring the residual nerve endings onto "target" muscle segments nearby — typically transected sections of pectoral, chest wall, or remaining upper-arm muscle that no longer serve their original biomechanical function. Over 3–6 months the nerve reinnervates the new muscle, and axon firing patterns tied to specific original hand/wrist movements now produce a strong, spatially distinct EMG burst directly over that muscle segment.
Surgeons deliberately map each transferred nerve to a separate muscle segment so that each channel of a subsequent electrode array corresponds, as closely as possible, to a single original motor intent: median nerve fascicles associated with finger flexion go to one segment, radial nerve fascicles associated with finger/wrist extension go to another, and so on. This creates several independent, high-amplitude EMG sources instead of one blended, low-amplitude signal — the foundation that makes multi-degree-of-freedom myoelectric control possible.
A newer extension, Regenerative Peripheral Nerve Interfaces (RPNI), wraps free muscle grafts around individual nerve fascicle ends instead of using existing muscle bulk, allowing even more independent control sites in patients who lack sufficient residual muscle for classic TMR.
Once reinnervated muscle segments exist, an array of electrodes must sit over them and continuously sample the electrical activity each generates. The choice between surface (skin-mounted) and implanted (intramuscular) electrodes shapes signal quality, channel count, robustness, and how much surgery the patient undergoes — a central engineering tradeoff in every modern myoelectric system.
Surface EMG electrodes sit inside the prosthetic socket, pressed against the skin over each reinnervated segment. They are non-invasive, cheap to replace, and easy to reposition — but signal quality depends heavily on skin contact, sweat, socket fit, and daily donning consistency. Cross-talk between nearby muscle segments is common, and small shifts in electrode position between mornings can meaningfully change the recorded pattern, forcing classifiers to be retrained or recalibrated periodically.
Commercial pattern-recognition systems such as Coapt Complete Control and Ottobock MyoPlus use 6–8 surface electrode pairs arranged around the residual limb, streaming raw EMG into an onboard processor embedded in the socket itself.
Intramuscular or epimysial implanted electrodes are surgically placed directly on or within the target muscle. Because they sit closer to the source and are unaffected by skin impedance or socket slip, they typically deliver a substantially cleaner, more stable signal with less cross-talk — enabling more independent channels and finer motor-class separation. Systems like the Implantable Myoelectric Sensor (IMES) place multiple wireless sensor units directly in muscle, telemetering data out through the skin.
The tradeoff is invasiveness: implantation requires surgery and carries infection and lead-migration risk, and any hardware failure means another operation to fix. Most clinically deployed systems today still favor surface electrodes for their simplicity and reversibility, reserving implanted arrays for research systems and patients needing the highest fidelity control.
A well-placed surface electrode array over TMR muscle can recover most of the functional benefit of implanted sensing — the surgery (TMR) does more of the signal-quality work than the electrode technology itself.
Raw EMG is a noisy, oscillating voltage signal with no obvious "meaning" on its own. Before any classifier can interpret it, the signal must be filtered to remove noise and motion artifact, then summarized over a short time window into a small set of numeric features that compactly describe how active and how "shaped" each channel's activity is — the feature vector that becomes the classifier's input.
Raw EMG is first band-pass filtered, typically 20–450 Hz, to remove low-frequency motion artifact (electrode cable sway, socket shift) and high-frequency electrical noise, then often notch-filtered at 50/60 Hz to remove powerline interference. The cleaned signal is broken into short overlapping analysis windows, commonly 150–250 ms with 50% overlap, each window producing one updated classification decision.
Window length is a direct tradeoff against decode latency: longer windows average out noise and improve classification accuracy, but every millisecond spent collecting data is a millisecond added to the delay between intention and movement.
For each channel and each window, a handful of statistics are computed that compress hundreds of raw samples into a few descriptive numbers:
• Mean Absolute Value (MAV): overall signal amplitude / muscle activation level • Root Mean Square (RMS): power of the signal, correlates with contraction force • Zero Crossings (ZC): how often the signal crosses zero, reflecting frequency content • Waveform Length (WL): cumulative signal complexity/length over the window • Slope Sign Changes (SSC): another proxy for frequency content and signal texture
Stacking these features across every channel (e.g., 6 features × 8 channels = 48-dimensional vector) produces the input the classifier will actually see — a drastic, information-preserving compression of the raw waveform stream.
Because features must be recomputed every window, feature extraction is usually the cheapest stage of the pipeline computationally — the real latency cost comes from the window length itself, not the math.
The feature vector produced every 150–250 ms must be translated into one of a small set of discrete motor intents: hand open, hand close, wrist pronate, wrist supinate, a specific grip pattern, or rest. Pattern recognition myoelectric control — trained on the patient's own labeled muscle-contraction data — has largely superseded older direct control schemes, enabling many more independent movements from the same electrode set.
Early myoelectric prostheses used direct control: one electrode pair's amplitude directly drove one motor function, and a second pair (or co-contraction) switched between functions — e.g., flex hard to close the hand, extend hard to open it, co-contract to switch to wrist rotation mode. This is simple and robust but limits patients to sequential, one-function-at-a-time control with awkward mode-switching.
Pattern recognition control instead treats the full multi-channel feature vector as a single pattern and trains a classifier to recognize which trained motion it most resembles, drawing on all channels simultaneously. This lets many more movements be distinguished from the same electrode set, and multiple channels of correlated but distinct information (e.g., different TMR segments) can be combined instead of used one at a time.
Linear Discriminant Analysis (LDA) remains the most widely deployed classifier in commercial systems (Coapt, Ottobock MyoPlus) because it is fast, requires little training data, and is robust to the moderate noise typical of surface EMG — often reaching 90-97% offline classification accuracy with good electrode placement. Support Vector Machines (SVM) offer somewhat better separation for harder, higher-dimensional class boundaries at higher computational cost. Shallow and, more recently, deep neural networks (including convolutional and recurrent architectures processing raw or lightly-processed EMG) are an active research direction, promising better robustness to electrode shift and more natural simultaneous/proportional control, at the cost of needing more training data and more onboard compute.
A confidence threshold is typically applied to the classifier's output: if no class exceeds the threshold, the system holds the previous state or defaults to "no movement" rather than guessing — trading a small amount of responsiveness for a large reduction in false triggers.
Total decode latency — from muscle contraction to a labeled decision — combines the analysis window length, feature computation, and classifier inference time. Studies on myoelectric control consistently find that latencies above roughly 300 ms feel noticeably sluggish and break the illusion of direct control, while latencies under ~150 ms feel responsive and are barely perceptible as a delay. This is why so much engineering effort goes into keeping window length short without sacrificing the averaging that gives features their statistical reliability — the two goals are in direct tension.
Clinically deployed pattern-recognition systems target end-to-end decode latency under roughly 150 ms — fast enough that the prosthesis feels like it responds to intention rather than to a command.
A decoded motor class is only useful once it drives physical movement in the prosthetic hand — and a truly embodied bionic limb also needs to send information back the other way, so the brain can feel what the hand is touching. Targeted Sensory Reinnervation and osseointegration are the two technologies extending myoelectric control into a genuinely bidirectional human-machine interface.
Once a motor class clears the confidence threshold, a command is sent to the prosthetic's embedded microcontroller, which drives independent motors in each finger (or a shared underactuated linkage in simpler hands) to execute the target grip or wrist motion. Proportional control maps EMG amplitude within the winning class to speed or force — a harder contraction closes the hand faster or grips tighter, not just triggering an on/off motion. Force and position sensors in the motors provide closed-loop control so the hand can, for example, stop closing once a stable grasp is detected, preventing crushing of fragile objects.
Targeted Sensory Reinnervation (TSR) is the sensory counterpart to TMR: residual sensory nerve fibers that once innervated the hand are rerouted to reinnervate a patch of skin on the chest or residual limb. After healing, touching that skin patch produces a sensation the patient perceives as coming from the missing hand itself — a phenomenon exploited by embedding pressure sensors in the prosthetic fingertips and driving small vibrotactile or electrotactile actuators against the reinnervated skin patch in proportion to sensed grip force. This closes the loop: the hand acts on the decoded intent, and the resulting contact force is relayed back as a felt sensation, improving grip reliability and reducing the visual attention needed to use the prosthesis.
Osseointegration anchors the prosthetic limb directly to the residual skeleton through a titanium implant fused into the bone (via an abutment protruding through the skin), replacing the traditional socket entirely. It is not a neural decoding technology, but it substantially improves the myoelectric pipeline's real-world performance: without a socket, electrodes can be permanently and precisely fixed relative to the reinnervated muscle rather than shifting daily with donning, and osseoperception (bone-conducted mechanical feedback) adds another passive sensory channel. Combined with TMR/TSR and implanted electrode arrays, osseointegration represents the current frontier toward a fully integrated, always-attached bionic limb.
The most advanced bionic limb systems today combine all three surgical/technological threads — TMR for motor signal, TSR for sensory feedback, and osseointegration for stable, socket-free mechanical attachment — into a single continuously worn, bidirectional neural-mechanical interface.