Uncertainty-Aware Single-Channel Surface EMG Classification for Low-Resource Hand Gesture Recognition: A Subject-Independent Evaluation

Single-channel surface electromyography (sEMG) could reduce the cost and complexity of hand-gesture interfaces, but performance commonly degrades when models are tested on unseen subjects. We evaluated five-class hand gesture classification using only the Flexor Carpi Radialis (FCR) channel from 83 quality-screened subjects drawn from 113 recordings. Time-domain (TD) features were classified with multilayer perceptrons (MLPs) and three uncertainty quantification (UQ) methods: Monte Carlo (MC) Dropout, Deep Ensembles, and Evidential Deep Learning (EDL). A raw-window one-dimensional convolutional neural network (1D-CNN) and a Temporal Convolutional Network (TCN) were included as end-to-end comparators. The best within-subject result was obtained by Random Forest at 73.3%± 13.3% , while full-coverage leave-one-subject-out cross-validation (LOSOCV) accuracy was 51.1%± 10.3% for the MLP baseline; all UQ methods were statistically indistinguishable from this baseline after Holm correction. Confidence-based selective prediction improved per-subject accuracy to $$68.3%$$ at $$50%$$ coverage and $$83.9%$$ at $$30%$$ coverage, but this gain was achieved equally by plain MLP softmax confidence and by MC Dropout, indicating that the improvement arises from the confidence-ordering property of softmax outputs rather than from stochastic inference. The deterministic MLP achieved the best calibration (Expected Calibration Error, ECE $$=0.024$$ ), while EDL showed vacuity collapse and was the only method to degrade under selective prediction (ECE $$=0.341$$ ). Both 1D-CNN and TCN showed numerically lower LOSOCV accuracy than the TD+MLP pipeline under the evaluated preprocessing configuration, though the differences were not statistically significant after Holm correction; loss of amplitude information through per-sample normalisation is a plausible contributing factor. Per-class analysis revealed that selective prediction gains are structurally concentrated in the rest class, which is retained at $$97.5%$$ coverage with near-perfect accuracy ( $$99.7%$$ ); excluding rest, active-gesture accuracy at $$50%$$ coverage was only $$46.5%$$ . This rest-class dominance effect has direct implications for how selective prediction should be evaluated in sEMG datasets that include a rest class. External validation on NinaPro DB2 (40 subjects, one circumferential forearm channel) confirmed above-chance classification ( $$p<0.0001$$ ) but showed no meaningful correlation between uncertainty and per-subject accuracy ( $$r=+0.047$$ ), indicating that the UQ mechanism does not generalise to this dataset. These results indicate that standard softmax confidence, requiring no additional training or inference overhead, is sufficient for rejection-based low-resource sEMG deployment under the evaluated conditions; computationally expensive stochastic inference methods provide no measurable additional benefit on low-dimensional tabular features.