Unsupervised Learning for Biomechanical Data Using Self-organising Maps, an Approach for Temporomandibular Joint Analysis

We proposed to apply a specific machine learning technique called Self-Organising Maps (SOM) to identify similarities in the performance of muscles around human temporomandibular joint (TMJ). The performance was assessed by measuring muscle activation with the use of surface electromyography (sEMG). SOM algorithm used in the study was able to find clusters of data in sEMG test results. The SOM analysis was based on processed sEMG data collected when testing subjects performing four mandibular motions: opening, closing, protrusion and retrusion. Muscle activation of four TMJ muscles (masseter right, masseter left, temporalis right and temporalis left) were used as input variables for SOM algorithm. The results of the network are presented on U-matrix maps. These maps consist of formed groupings that correspond to similarities in data points that clustered together. The clustering implies similarity in muscle activation of different subjects. The results show that it is possible to cluster medical datasets with SOM algorithm in the analysis of full jaw motions, which may support the diagnostic process.