Recognizing knee pathologies by using gait dynamics via kernel principal component analysis and deterministic learning theory

Abstract

This study concentrates on dynamical motion characteristics of knees during treadmill walking, and describes the application of kernel principal component analysis and deterministic learning theory to the detection of abnormal gait dynamics induced by knee pathologies. First, twelve-dimensional knee joint rotation and translation parameters are measured by a marker-based motion analysis system during treadmill walking. Second, the dimension of the gait parameter signal is reduced based on kernel principal component analysis. Third, gait dynamics underlying time-varying gait parameter signal is captured by using localized radial basis function neural networks through deterministic learning mechanism. This kind of dynamics information represents the frame-to-frame temporal change of kinematics and kinetics modifications during treadmill walking, which is shown to be sensitive to the variance during knee pathologies. Gait patterns are represented as the gait dynamics underlying time-varying gait parameters. The training patters under different knee pathologies further constitute a uniform training dataset, containing all kinds of gait dynamics under different knee pathologies. By comparing the training patterns with a test knee gait pattern to be recognized, a set of recognition errors are generated. The average $L_1$ norms of the errors are taken as the classification measure between the dynamics of the training gait patterns and the dynamics of the test pattern. A rapid recognition scheme can be achieved according to the smallest error principle. Experimental results show that encouraging recognition accuracy can be achieved. The proposed method is sensitive for the detection of knee pathologies and is capable of discharge of patients without knee pathologies.