Source-Validated Personalized Fusion for Calibration-Efficient Static Wrist sEMG Gesture Recognition
Abstract
This study proposes SVP-Fusion, a source-validated personalized fusion framework for calibration-efficient static wrist-sEMG gesture recognition in unseen participants. Cross-participant surface electromyography (sEMG) gesture recognition remains challenging because population-level models often generalize poorly to new users. We evaluate calibrated static wrist-sEMG recognition on an expanded GRABMyo/GRABMyoFlow cohort comprising 63 participants, 17 gesture classes, three recording sessions, and 22,491 trials. SVP-Fusion combines a source-trained global TemporalCNN with participant-specific classical and deep RBF-kernel support vector machines and a neural fine-tuning seed ensemble. Held-out source-validation participants are treated as pseudo-target users to select adaptation depth, posterior temperatures, and fusion weights before the resulting policy is applied to unseen outer-target participants. Evaluation uses five participant-disjoint outer folds and a fixed common test set containing 85 trials per participant at every calibration budget. The task is calibrated rather than zero-shot: <inline-formula> <tex-math notation="LaTeX">$K=16$ </tex-math></inline-formula> corresponds to 272 labeled calibration trials per participant distributed across the three recording sessions. The primary preprocessing protocol uses label-free but transductive participant-wise normalization, and a calibration-only normalization control is included to quantify its deployment-relevant effect. At <inline-formula> <tex-math notation="LaTeX">$K=16$ </tex-math></inline-formula>, SVP-Fusion achieved 93.80% <inline-formula> <tex-math notation="LaTeX">$\pm ~0.71$ </tex-math></inline-formula>% accuracy and 93.80% <inline-formula> <tex-math notation="LaTeX">$\pm ~0.71$ </tex-math></inline-formula>% macro F1-score, compared with 69.97% accuracy for the zero-target-label global model and 91.05% for the strongest individual personalized predictor. At <inline-formula> <tex-math notation="LaTeX">$K=8$ </tex-math></inline-formula>, it achieved 90.84% accuracy, while the calibration-only normalization control changed accuracy by only −0.24 percentage points at <inline-formula> <tex-math notation="LaTeX">$K=8$ </tex-math></inline-formula> and + 0.09 percentage points at <inline-formula> <tex-math notation="LaTeX">$K=16$ </tex-math></inline-formula>. These results show that source-validated fusion provides a favorable accuracy–calibration trade-off for calibrated trial-level static wrist-sEMG recognition.