Abstract
BACKGROUND: EMG-based hand-gesture recognition can enable home-based post-stroke rehabilitation, yet one-size-fits-all feature sets overlook differences across recovery stage METHODS: Thirteen post-stroke participants performed seven gestures while EMG was recorded from six forearm sensors. From 38 time- and frequency-domain features, we derived stage-specific subsets for Low (Brunnstrom 1-2, minimal movement), Medium (3-4, partial movement), and High (5-6, near-normal movement) using a wrapper approach Sequential Forward Selection (SFS). For reference, we included a filter comparison using minimum Redundancy-Maximum Relevance (mRMR). To provide fair baselines, we reproduced two literature feature sets within an identical Light Gradient Boosting Machine (LightGBM) pipeline: (i) a healthy-cohort feature set and (ii) a patient-cohort feature set that was not stage-stratified and did not focus on feature selection (we adopted the features as reported). Multiple classifiers-Linear Discriminant Analysis, Support Vector Machines, Random Forest, LightGBM, Logistic Regression, and K-Nearest Neighbors-were evaluated via group-wise cross-validation. Within-stage variability was quantified using pairwise Jaccard overlap of selected features. RESULTS: Stage-tailored subsets achieved compact yet accurate models: High = 81.5% (14 features, LightGBM), Medium = 80.2% (9 features, LightGBM), Low = 65.0% (7 features, Random Forest). SFS exceeded the mRMR filter comparison and outperformed both literature baselines under the same LightGBM pipeline (paired tests across CV folds, [Formula: see text]). Relative to the healthy-cohort baseline, gains were +6.5% (High), +6.2% (Medium), and +12.0% (Low); relative to the non-stage-stratified patient baseline, gains were +9.5%, +10.2%, and +21.0%, respectively. Time-domain metrics-particularly Difference Absolute Standard Deviation Value and Sample Entropy were most frequently selected. Jaccard analyses indicated within-stage heterogeneity alongside convergence on a small set of core discriminative features. CONCLUSIONS: Brunnstrom stage-specific feature engineering substantially improves EMG gesture-classification accuracy over both healthy-derived and non-stage-stratified patient baselines while reducing computational load. These findings support adaptive, stage-aware designs for wearable rehabilitation systems and motivate larger Low-stage cohorts and models robust to sparse or low-SNR signals.