Exploring the Influence of Feature Selection Methods on a Random Forest Model for Gait Time Series Prediction Using Inertial Measurement Units

Abstract

Despite the increasing use of inertial measurement units (IMUs) and machine learning techniques for gait analysis, there remains a gap in which feature selection methods are best tailored for gait time series prediction. This study explores the impact of using various feature selection methods on the performance of a random forest (RF) model in predicting lower limb joints kinematics from two IMUs. The primary objectives of this study are as follows: (1) Comparing eight feature selection methods based on their ability to identify more robust feature sets, time efficiency, and impact on RF models' performance, and (2) assessing the performance of RF models using generalized feature sets on a new dataset. Twenty-three typically developed (TD) children (ages 6–15) participated in data collection involving optical motion capture (OMC) and IMUs. Joint kinematics were computed using opensim. By employing eight feature selection methods (four filter and four embedded methods), the study identified 30 important features for each target. These selected features were used to develop personalized and generalized RF models to predict lower limbs joints kinematics during gait. This study reveals that various feature selection methods have a minimal impact on the performance of personalized and generalized RF models. However, the RF and mutual information (MI) methods provided slightly lower errors and outliers. MI demonstrated remarkable robustness by consistently identifying the most common features across different participants. ElasticNet emerged as the fastest method. Overall, the study illuminated the robustness of RF models in predicting joint kinematics during gait in children, showcasing consistent performance across various feature selection methods.