PCA sensitivity: The role of representative and outlier strides in gait sequence

Abstract

Principal component analysis (PCA) is a useful statistical technique for the reduction of data dimensionality. When applied to the accelerometer data in gait analysis PCA assigns common gait patterns among subjects or provides gait classification. In this paper, we study the results of PCA applied to datasets recorded with three-axial accelerometers placed on thigh, shank, and foot in subjects with hemiplegia. In particular, we analyze the impact of both representative stride (the most similar to all other strides in the sequence) and outlier stride (the most different from all other strides in the sequence) on PCA results. PCA sensitivity to data preparation was tested on three datasets: complete gait sequence, gait sequence without the outlier stride, and on representative stride.