Surface EMG signal segmentation based on HMM modelling: Application on Parkinson’s disease

The study of burst electromyographic (EMG) activity periods during muscles contraction and relaxation is an important and challenging problem. It can find several applications like movement patterns analysis, human locomotion analysis and neuromuscular pathologies diagnosis such as Parkinson disease. This paper proposes a new frame work for detecting the onset (start) / offset (end) of burst EMG activity by segmenting the EMG signal in regions of muscle activity (AC) and non activity (NAC) using Discrete Wavelet Transform (DWT) for feature extraction and the Hidden Markov Models (HMM) for regions classification in AC and NAC classes. The objective of this work is to design an efficient segmentation system of EMG signals recorded from Parkinsonian group and control group (healthy). The results evaluated on ECOTECH project database using principally the Accuracy (Acc) and the error rate (Re) criterion show highest performance by using HMM models of 2 states associated with GMM of 3 Gaussians, combined with LWE (Log Wavelet decomposition based Energy) descriptor based on Coiflet wavelet mother with decomposition level of 4. A comparative study with state of the art methods shows the efficiency of our approach that reduces the mean error rate by a factor close to 2 for healthy subjects and close to 1.3 for Parkinsonian subjects.