Kinematic classification of gait patterns using a neural network evaluation system

JORGE BALEJ; MARCOS CRESPO; DIEGO BALLESTEROS; MALCO ROSSI; JULIETA ARENA; ANDRES CERVIO; CAROLINA CUELLO ORDERIZ; ALBERTO RIVERO; MARCELO RISK; MARCELO MERELLO

Dublin

Congreso; Sixteenth International Congress of Parkinson's Disease and Movement Disorders; 2012

International Parkinson and Movement Disorder Society

Objective: To determine neural network analysis utility for gauging of kinematic gait variables and improved identification of gait patterns. Background: Applicability of kinematic gait analysis for diagnostic purposes in neurological disorders has not yet been established. Methods: 85 subjects consented to the study. A six-camera optoelectronic ELITE system (BTS Bioengineering, Milan, Italy) was used to determine time and space coordinates of 17 separate wireless markers placed on each subject. Once markers were in place, 42 absolute variables were calculated, values were also normalized according to subject height and weight. Different possible combinations were then tested, generating clusters, each of which was assigned a score. Combining variables into groups responding to clinical classification generated 16 promising variables. The network used was perception with one hidden layer and supervised learning with back propagation algorithm. Network training was performed using a group of subjects whose gait pattern was clinically known (19 subjects with normal gait and 66 patients). Results: Neural network evaluation established 4 different groups of gait patterns: normal, ataxic, parkinsonian and paraparetic. Each group presented a centroid (the central point corresponding to the 16 coordinate average for each variable) and a radius, with the radius equivalent to twice the average distance to the centroid. Groups were distributed without overlapping in 16 dimensions. Conclusions: Neural network analysis is a promising tool for biomedical research, allowing improved classification of clinically identified gait patterns. Future studies should explore whether neural network analysis of kinematic gait data would allow incipient undetermined gait disorder detection, and establish defined diagnoses sooner.