Designing Training Sample Size for Support Vector Machines Based on Kinematic Gait Data

INTRODUCTION Traditionally, biomechanical gait data are analyzed using discrete variables and a univariate statistical approach. In contrast, the multivariate and complex pathoaetiology of a running-related musculoskeletal injury often demands a more robust analysis [1]. Support Vector Machines (SVMs) have recently been applied to analyze biomechanical data due to its ability to identify complex associations amongst many discrete gait variables [2]. However, the number of observations required to train the SVM classifier to solve biomechanics classification problems is still unknown. For this reason, generic rule-of-thumb approaches have been used suggesting that a very large sample size is required to provide appropriate training for optimal classification. Unfortunately in clinical studies, large samples are usually not readily available. In addition, large sample sizes may be wasteful when fewer samples may result in optimal classification. Furthermore, a large sample size does not guarantee that one will obtain optimal classification since the groups may considerably overlap in their distribution, even in high-dimensional space. Therefore, it is paramount to a priori approximate the sample size in biomechanical studies. Hence, the aim of this study is to provide guidance to plan the appropriate sample size for SVM training classification based on discrete kinematic gait variables.