Feature Selection, Construction, and Validation of a Lightweight Model for Foot Function Assessment During Gait With In-Shoe Motion Sensors

Abstract

Identifying and monitoring overpronation and oversupination in the long term during activities of daily living are essential for people's ambulatory health. Using an in-shoe motion sensor (IMS) with power-saving functions is a potential solution. In this study, we challenged the development of an estimation model of foot function using the foot center of pressure excursion index (CPEI) as an index via linear multivariate regression, which is sufficiently light for this type of IMS. Data collected from 65 and 17 participants were involved in model construction and validation, respectively. We validated ten scenarios simulating daily living activities, including walking on different surfaces, using different shoes, with or without carrying a bag, and indoors and outdoors. We applied statistical parametric mapping (SPM) to determine significant predictors and performed our original feature selection algorithm, leave-one-subject-out least absolute shrinkage and selection operator (LASSO), to compress the volume of the predictors. We successfully discovered significant sex-specific predictors for foot function estimation from foot motion and constructed large effect-sized sex-specific foot function estimation models that achieved high-precision CPEI estimation. In the validation, the model successfully estimated a maximum of 99.0% and 100.0% males' and females' data under the same experimental conditions with the training data and 92.8%-100.0% and 85.8%-100.0% data in different scenarios. The constructed models are effective and possible to provide applications for long-term foot function monitoring by using an IMS.