Stochastic Generative Model and Signal Analysis of Surface EMG

Surface electromyography (EMG) signals generated during muscle contraction contain various information related to human motion intention and muscle contraction, and are widely used for robotic prosthetic hand control and diagnosis of neuromuscular diseases. EMG signals have traditionally been assumed to follow a zero-mean Gaussian distribution, and various signal processing and analysis methods have been developed based on this assumption. However, recent studies have experimentally reported that EMG signals can exhibit non-Gaussian properties due to changes in muscle activity characteristics such as changes in muscle contraction level and the presence of muscle fatigue. Therefore, conventional approaches based on Gaussian distribution models cannot adequately handle the features contained in EMG signals. In this research, we assume that the recently reported non-Gaussian nature of EMG signals is caused by uncertainty in the amplitude (=variance) of EMG signals, and we are working on creating new signal analysis methods by constructing stochastic generative models that can express this. Beyond signal analysis such as muscle fatigue evaluation and quantification of muscle activity, the model also serves as a foundation for learning and inference in motion classification. Recently, we have extended it to hybrid generative-discriminative learning, which combines the generative nature of scale mixture models with the classification performance of discriminative models, aiming at a framework that offers both the interpretability of probabilistic models and the accuracy of deep learning.