Confidence-gated temporal voting for EMG-based prosthetic arm control: a low-cost end-to-end pipeline with quantified...

Surface electromyography (sEMG) is the dominant non-invasive route to volitional prosthetic control, but a closed-world classifier issues a command for every input, including inputs corresponding to no trained gesture. This study evaluated a low-cost, three-gesture sEMG control pipeline combining a Random Forest classifier with rejection on the maximum Random-Forest class probability and temporal majority voting, and measured the resulting rejection behaviour rather than assuming it.

Three data sources were analysed: modelled synthetic sEMG; Ninapro DB5 (10 subjects, 16 channels at 200 Hz, documented three-channel subset); and live three-channel sEMG from 13 healthy limb-intact adults recorded over three sessions with MyoWare sensors spaced equally around the forearm flexor group. Each of the 117 recordings lasted 120 s, comprising a 60 s continuous phase and four 10 s sustained contractions separated by 5 s of rest, yielding 21,294 windows of which 4,212 were rest. Channel interleaving gave an effective 333.3 Hz per channel. All datasets were segmented in physical time (0.6 s window, 0.2 s overlap) without resampling, and four time-domain features (MAV, RMS, WL, ZC) were classified by a 100-tree Random Forest. Evaluation used repetition-level, cross-session and leave-one-participant-out (LOPO) splits, with calibration and threshold selection confined to a participant-disjoint partition. Models were trained and tested within each domain, so the three accuracies index benchmark difficulty rather than transfer.

LOPO accuracy was 80.2% (synthetic), 74.5% (DB5) and 80.2% (live). The pooled window-level split used originally overestimated live accuracy by 7.7 percentage points. Expected calibration error was 0.147 uncalibrated and 0.038 after Platt scaling. Open-set AUROC was 0.842 against rest, 0.784 under electrode displacement and 0.741 for held-out movements, and 7.4% of windows were confidently incorrect. The complete pipeline reduced false actuation from 19.80% to 5.00% and command switching from 0.16 to 0.02  $s^{-1}$ , while rejecting 18.40% of windows and raising median latency from 0.60 to 1.95 s. The prototype completed 12 of 15 benchtop trials.

The results support a benchtop engineering proof of concept with quantified rejection and latency behaviour, not clinical safety or deployment readiness.