Self-Organizing Maps as a Tool to Analyze Movement Variability

Self-Organizing Maps possess unique properties that remove redundancies in a high-dimensional input space and map that input space to a low dimensional output space, thereby showing non-linear relationships in the input data. This ability makes Self-Organizing Maps attractive for measuring inter-limb coordination patterns. The current study takes previously published data (Bartlett, Bussey, & Flyger, 2006) used to compare the reliability of different operators digitizing gait patterns with and without anatomical markers. The trained network was simulated and analyzed qualitatively using the trajectory of activated nodes for each input vector, similar to Barton, Lees, Lisboa, & Attfield (2006). Qualitative differences in map trajectories were seen between Marker and No-Marker conditions and supported the results of the original publication that, when using 2-D videography, manually digitized markers allowed accurate estimation of movement variability, whereas the No-Marker condition did not. This finding is shown by the No-Marker trajectory travelling further from the center of the network cluster than the Marker condition. Additionally, the map trajectories revealed that changes in coordination at different phases of the movement can be identified. For most trials the Marker trajectory travels closer to the centre of the map than the No-Marker condition which, as is explained by the neighbourhood function, indicates less variability. The consistency between conventional biomechanical analysis techniques and the qualitative assessment of Self-Organizing Map outputs adds to the validity of the Self-Organizing Map as an accurate measurement tool for coordination. The ability to identify changes in coordination using the map trajectories illustrates the potential for the SelfOrganizing Map to show novel information about the coordination pattern. KEYWORDS, ARTIFICIAL NEURAL NETWORKS, COORDINATION, RELIABILITY, SELF-ORGANIZING MAPS, VARIABILITY