Feature extraction and classification algorithms for assessing muscle condition using multi sensors

Abstract

Hamstring muscle strain injury is the most common injuries in sport environment especially in football. Beside sport, dancing is another physical activity that has high risk for hamstring muscle strain injuries. Muscle fatigue is one of the risk factors that cause muscle injury. Muscle fatigue occurs when muscle fail to maintain desired muscle strength. The aim of this work is to assess and distinguish muscle fatigue using 3 different myograms (Electromyogram, Mechanomyogram and Acoustic myogram). To achieve the aim of this project, 3 different myograms were recorded simultaneously from hamstring muscle (Biceps Femoris and Semimembranosus) during isometric contraction with different loads (5, 7.5 and 10kg). 20 healthy male subjects were recruited in this work with aged 22.4 ±2.6 years. The recorded myograms were denoised and segmented. Pearson correlation and linear regression were applied on 9 time and frequency domain features (were extracted from each frame of each myograms) to find the behavior of the myograms and relationship between the myograms. Based on the pearson correlation results there are a strong relationship between AMG and EMG signals and a weak relationship between MMG and EMG signals. However, the relationship of MMG and EMG signals becomes significant when the load is increased. Two types feature sets were used to classify muscle condition and there are conventional features set (CFS) and M-band wavelet transform based feature set (MFS). In CFS, 16 features were extracted from each myograms, which were used in previous studies to assess muscle condition. In MFS, the recorded myograms were decomposed using 4-band wavelet transform and 9 features were extracted from each node of each myograms. Therefore, 108 features were generated for each myograms. The feature sets of EMG, MMG and AMG of CFS and MFS were augmented to form a new feature set. Two stage feature selection was used to reduce feature space in the new feature set. In stage I of features selection, linear locally embedding (LLE) was used. New reductive feature space was used in stage II for feature selection. In that stage, binary particle swarm optimization (bPSO) was used. k-nearest neighbor (k-NN) classifier was used to classify muscle condition (non-fatigue or fatigue) by using individual myogram feature sets, concatenated / augmented feature sets and after 2 stage feature selection feature sets. In the individual assessment, EMG feature sets have obtained high accuracy. However, compare with the concatenated feature sets, this feature sets have given high accuracy. Two stage feature selection was applied to the concatenated feature sets and the features were reduced by 69% and 89% for CFS and MFS in stage 1 respectively. In stage 2 the features were reduced by 7% and 40% for CFS and MFS respectively. The classification result was improved to above 75% and 92% for CFS and MSF respectively. It can be conclude that, the combination of three myogram provides more useful information than a single myogram.