Adaptive EMG Pattern Classification via Probabilistic Knowledge Transfer With Scale Mixture-Based Bayesian Sequential Learning

Electromyogram (EMG) signals, measured non-invasively from the skin surface, reflect human motion intentions and enable device control through pattern classification, particularly in applications such as myoelectric prostheses. However, continuous use of EMG-based interfaces remains challenging due to temporal variations in signal characteristics caused by muscle fatigue and electrode shift, leading to gradual degradation in classification accuracy. To address this limitation, we propose an adaptive method that integrates a scale mixture classification model (SMCM) with Bayesian sequential self-training (BSST), enabling probabilistic knowledge transfer across trials. In this framework, the posterior distribution of model parameters is sequentially updated through Bayesian updates using pseudo-labels assigned based on prediction confidence. This approach enables adaptation to evolving signal characteristics without storing historical data. Furthermore, SMCM represents signal variability through variance uncertainty modeling, thereby improving both the representation of EMG signal distributions and the reliability of prediction confidence estimation. We evaluated the proposed method using both short-term (within-day) and long-term (30 days) EMG datasets. The results showed that our method outperformed conventional methods in both classification accuracy and confidence estimation, while effectively mitigating accuracy degradation over time. These results demonstrate that the combination of SMCM and BSST provides effective adaptation to EMG signal variations, offering a practical solution for reliable EMG-based interfaces. Code is available at https://github.com/Yseitaro/smcm-bsst.