Personalized Sensor Selection Under Channel Compression in Surface EMG Gesture Recognition: An Identity-Preserving, Leakage-Controlled Study on NinaPro DB2

This independent research study investigates whether subject-specific sensor selection becomes more valuable as the number of available surface electromyography (sEMG) channels is reduced. Using NinaPro DB2 Exercise 1 data from 40 intact subjects, the study applies a frozen, leakage-controlled protocol separating training repetitions (1, 3, 4), calibration (6), and sealed-test repetitions (2, 5). The research was conducted through a staged OrganOS EMG evidence lineage: 1. Channel Compression Baseline A frozen baseline showed progressive recognition-performance loss as the available sEMG channel budget decreased. 2. Sensor Selection Calibration-based channel selection recovered a substantial portion of the performance lost under channel compression. 3. Standard Selection Controls Sequential Forward Selection (SFS), mutual-information selection, prefix baselines, and exhaustive subset search were compared. Exhaustive search did not demonstrate a meaningful advantage over SFS, providing an important negative result and indicating that expensive exhaustive enumeration was unnecessary under this protocol. 4. Global vs Personalized Sensor Selection Subject-specific SFS channel subsets were compared with leave-one-subject-out (LOSO) global subsets constructed without using the target subject. Personalized selection improved mean sealed-test accuracy by: - 8 channels: +0.93 percentage points - 6 channels: +1.35 percentage points - 4 channels: +3.30 percentage points At the four-channel budget, personalized selection outperformed the LOSO global subset in 35 of 40 subjects. The four-channel paired improvement was +3.30 percentage points, with a bootstrap 95% confidence interval of +2.24 to +4.42 percentage points and a two-sided Wilcoxon p-value of 2.54×10^-7 (Holm-adjusted p = 7.61×10^-7). The results support a deliberately narrow conclusion: Under this frozen NinaPro DB2 protocol, subject-specific sensor selection becomes increasingly valuable as the available sensor budget becomes more constrained. This work does not claim state-of-the-art sEMG gesture-recognition performance, clinical efficacy, or superiority of a proprietary OrganOS classifier. The accompanying reproducibility package contains the final manuscript, frozen derived evidence, execution scripts, figures, metadata, and SHA-256 manifests. Original NinaPro raw data are not redistributed. Author: Min-Gi Kim Independent Researcher Research Program: OrganOS Version: v1.1 Date: 22 September 2026