The Influence of Motion Data Low-Pass Filtering Methods in Machine-Learning Models

This study assessed the effect of filter parameters on gait characteristics and the performance of machine-learning models. Overground walking trials (n = 99) with and without perturbations (slips, trips) were collected for 33 healthy older adults. Kinematics were collected by a motion capture system. Different Butterworth low-pass parameters were applied to the raw data, including three orders (2–6) and nine cutoffs (4–20 Hz). Spatiotemporal gait outcomes were then calculated to develop classification models to automatically identify the trial type (gait, gait–slip, or gait–trip) using Logistic Regression, Support Vector Classification, and Random Forest Classification. A 3 × 9 ANOVA showed main effects of order and cutoff (p < 0.01 for all) on gait characteristics during both perturbed and regular walking trials. However, the gait characteristics were different between them. The filter parameters significantly affected the performance of classification models using different classifiers, with significant main effects of the filter order (p < 0.05) and cutoff (p < 0.01) on AUC and overall accuracy for all of the models. Our results suggest that the standard Butterworth filter (fourth-order, cutoff: 6 Hz) is suitable for the development of classification models with low–medium complexity, while for models with high complexity (i.e., ensemble models), a filter with a higher order and cutoff (sixth-order, cutoff 10–12 Hz) might yield better performance.