Prosthetic Hand Control: Reading Intent From Muscle

Reading intent from the muscle

Surface electromyography (sEMG) picks up the summed electrical activity of motor units in a muscle as it contracts, through electrodes placed on the skin rather than needles in the tissue. After an amputation, the muscles that once moved the missing hand or fingers are often still there in the residual limb, and they still fire with the same motor commands the brain sends — the intention to close a hand does not vanish with the hand itself. A myoelectric prosthesis reads that residual activity and decodes it back into a grasp command.

From raw signal to usable features

Raw sEMG is a noisy signal, typically under a millivolt in amplitude and spread across roughly 20 to 500 Hz. Rather than classify the raw waveform, controllers slide a short window — commonly 100 to 250 milliseconds — over the signal and reduce it to a handful of time-domain features that are cheap to compute and robust to noise:

RMS = sqrt( (1/N) * Σ x_i^2 )         // root mean square — overall effort
MAV = (1/N) * Σ |x_i|                 // mean absolute value — cheap proxy for RMS
ZC  = count of sign changes in x      // zero crossings — frequency-content proxy
WL  = Σ |x_i - x_(i-1)|               // waveform length — signal complexity

Classifying the grasp: LDA and beyond

Features from several electrode channels are stacked into one feature vector and fed to a classifier, trained on a handful of the user's own labelled contractions — open hand, close hand, pinch, wrist rotation and so on. Linear Discriminant Analysis remains the clinical workhorse: it trains in seconds on a small amount of data, runs comfortably in real time on embedded low-power hardware, and gets useful accuracy from as few as two to eight electrode channels. Convolutional networks on raw or spectrogram EMG can push accuracy higher in the lab, but the felt threshold for a prosthesis to seem responsive is a round-trip latency under roughly 300 milliseconds, and that tight, low-power budget keeps simple linear classifiers genuinely competitive in clinical devices.

Pattern recognition versus direct control

Older myoelectric hands use direct control: one muscle pair opens and closes the hand, a second pair or a deliberate co-contraction gesture switches to another degree of freedom such as wrist rotation. It is simple and reliable but limited to one motion at a time and requires the user to consciously mode-switch. Pattern recognition control instead classifies the whole multi-channel activation pattern at once, so a single set of electrodes can distinguish many distinct grasp shapes directly, without a dedicated mode-switch gesture — closer to how an intact hand is actually controlled, at the cost of needing a trained classifier rather than a fixed wiring rule.

What makes it reliable outside the lab

Two effects explain most of the gap between clean lab benchmarks and real-world reliability. Electrode shift: taking the socket off and putting it back on repositions the electrodes by a few millimetres, changing which motor units dominate the recorded signal. The limb-position effect: gravity and mechanical loading change a muscle's activation pattern for what the user still intends as the identical gesture, once the arm moves from a seated resting position to, say, reaching overhead. Robust systems train on data collected across multiple sessions and limb positions, use classifiers that adapt online, or rely on targeted muscle reinnervation — a surgery that redirects residual nerves to new muscle sites, creating additional, more separable control signals for people with higher-level amputations.

Frequently asked questions

Try it live

Everything above runs in your browser — open Prosthetic Hand Control — EMG and change the parameters while it is running. Nothing is installed, nothing is uploaded, the whole model lives in one tab.