Class-Incremental Learning for Myoelectric Locomotion Modes Prediction: Evaluating Replay Strategies Against Catastrophic Forgetting

Surface electromyography (sEMG)-based locomotion intent recognition enhances lower-limb wearable device control by enabling earlier adjustments of control strategies, thereby improving dynamic responsiveness in complex environments. While existing sEMG-driven frameworks achieve high accuracy in fixed locomotion mode sets, prior research has not addressed incremental learning scenarios—a critical gap for real-world deployment where new locomotion modes must be dynamically integrated without retraining from scratch. To bridge this gap, this study presents the first systematic exploration of class-incremental learning (CIL) in sEMG-based intent recognition, benchmarking three replay-based methods: incremental Classifier and Representation Learning (iCaRL), Bias Correction (BiC), and Dynamically Expandable Representation (DER) under two incremental learning protocols: Slow adaptation (five-phase incremental learning with two classes per phase) and Rapid adaptation (two-phase incremental learning with five classes per phase). Key findings reveal that network expansion (DER) optimally mitigates catastrophic forgetting, limiting old-class accuracy drop to 14.07% while maintaining 75.80% accuracy in rapid adaptation. This work establishes foundational insights for deploying incremental learning in adaptive wearable robotics.