Gaussian mixture model based clustering of Manual muscle testing grades using surface Electromyogram signals

Abstract

Muscle strength testing has long been an important assessment procedure in rehabilitation setups, though the subjectivity and standardization of this procedure has been widely debated. To address this issue, this study involves the use of Electromyogram (EMG) features that are intuitively related to muscle strength to classifyManual muscle testing (MMT)grades of '4-', '4', '4 + ' and '5' of the Medical Research Council scale. MMTwas performed on Tibialis anterior muscle of 50 healthy participants whose MMT grades and EMG were simultaneously acquired. Chi square goodness of fit and Spectrum Decomposition of Graph Laplacian (SPEC) feature selection algorithms areused in selecting five features, namely Integrated EMG, Root Mean Square EMG, Waveform Length, Wilsons' amplitude and Energy. Gaussian Mixture Model (GMM) approach is used for unsupervised clustering into one of the grades. Internal cluster evaluation resulted in Silhouette score of 0.76 and Davies Bouldin Index of 0.42 indicating good cluster separability. Agreement between the machine-based grade and manual grade has been quantified using Cohens' Kappa coefficient. A value of '0.44' has revealed a moderate agreement, with greater differences reported in grading '4' and '4 + ' strength levels. The comparative advantage of EMG based grading over the manual method has been proved. The suggested method can be extended for muscle strength testing of all muscles across different age groups to assist physicians in evaluating patient strength and plan appropriate strength conditioning exercises as a part of rehabilitative assessment.