Derived feature dataset and benchmark results for subject-independent sEMG gesture classification

This dataset contains derived results, benchmark tables, and trained model artifacts produced by a subject-disjoint machine-learning pipeline for surface electromyography (sEMG) gesture classification, applied to NinaPro Database 2 (40 subjects, 12 channels, 49 movements plus rest, 692,276 labelled 200 ms windows). NinaPro DB2 itself is a pre-existing, publicly available third-party dataset and is not redistributed here; every file in this deposit is an original output of the authors' own analysis. The dataset includes: held-out and 40-fold Leave-One-Subject-Out (LOSO) cross-validation results for three tuned gradient-boosted classifiers (CatBoost, XGBoost, LightGBM); per-subject LOSO ranking and calibration statistics; global feature-importance rankings from two independent methods (SHAP and permutation importance) plus a feature-level stability analysis across subjects; feature- and channel-level ablation results quantifying the cost of removing top- vs. bottom-ranked features; embedded-deployment benchmarking results (stage-by-stage inference latency, model export size across four formats, and a direct test of post-training quantization applicability); and the tuned CatBoost model exported in both native binary (.cbm) and ONNX formats. These derived data support independent verification and reuse of the reported classification performance, explainability cross-checks, and deployment feasibility claims without requiring access to the original NinaPro DB2 recordings or re-execution of the full training pipeline.