|
|
 |
| |
|
|
|
FTCL: Federated Transfer Continual Learning with Outlier-Aware Adaptive EWC v2 and Episodic Replay for Personalized sEMG Gesture Recognition |
|
|
|
PP: 1183-1195 |
|
|
doi:10.18576/amis/200504
|
|
|
|
Author(s) |
|
|
|
Atef Ahmed Obeidat,
Naqaa Atef Obeidat,
Tamara Amjad Al-Qablan,
|
|
|
|
Abstract |
|
|
| Surface electromyography (sEMG)-based gesture recognition for myoelectric prosthetic control is hindered by inter-subject variability, privacy constraints on centralizing biomedical data, and catastrophic forgetting during continual multi-subject adaptation. This paper aims to address these challenges jointly through a novel federated continual learning framework. We present FTCL, a federated Transfer continual learning framework. It pre-trains a hybrid BiLSTM-CNN model, distributes it via FedProx aggregation, and performs subject-specific calibration through classifier-head-only fine-tuning regularized with two novel methods: (i) OA-EWC v2 (Outlier-Aware Adaptive Elastic Weight Consolidation v2) combining rolling Mahalanobis-distance recalibration with gradient- interference lambda boosting; and (ii) a lightweight Episodic Replay Buffer (LERB) that maintains direct exposure to old subject data during sequential fine-tuning. The framework was evaluated on the FORS-EMG dataset (19 subjects, 12 gestures, 54,720 windows) under Leave-One-Subject-Out Cross-Validation (LOSO-CV). FTCL achieves 80.69% accuracy (Macro-F1 = 0.8091), outperforming LDA (44.49%) and SVM (49.49%), with all 19 folds exceeding the 70% clinical usability threshold. Ablation study results show that the lightweight replay buffer mitigates forgetting significantly, and the combination of OA-EWC v2 and replay provides a mean sequential accuracy of 76.93% and True BWT of −9.07, signifying a 52% decrease in forgetting compared to the baseline. These results prove that FTCL is a privacy-aware clinically applicable method for personalized myoelectric prosthesis control, ensuring compatibility between data privacy and continual adaptability. |
|
|
|
|
|
|
|
