Using Machine Learning and Wearable Inertial Sensor Data for the Classification of Fractal Gait Patterns in Women and Men During Load Carriage

Abstract

Ambulating while carrying a mission specific load is one of the most frequently executed occupational tasks for the military, especially for individuals in combat roles. Prolonged ambulation is a naturally dynamic and complex process, characterized by highly multi-dimensional interactions within the gait mechanics of the lower extremity. Recent wearable sensors studies, like inertial measurement unit (IMU)-related gait studies have demonstrated that machine learning (MLN) algorithms and fractal analysis can successfully discriminate between classes, such as movement patterns, injury, age and sex. This study attempts to classify fractal gait patterns of women and men using IMU-based signal data obtained from accelerometer, gyroscope and magnetometer during a 2 km loaded (20 kg) march. Random Forest (RF) MLN algorithm was used to generate a model that can measure the accuracy and identify the importance of IMU-based signal-related fractal variables. A total of 18 variables were calculated using 2 fractal methods, detrended fluctuation analysis (DFA) and wavelet transform-based power spectral density (PSD), from 3 IMU-based signals in their 3 axes (medial-lateral, vertical and anterior-posterior). A total of 33 healthy adults (17 men [26.7±5.9 years] and 16 women [25.2±4.5 years]) volunteered for this study. A 9-axis IMU sensor was attached to each participant at each of the following locations: feet, shanks, thighs and lumbar spine. An independent training-testing approach, called one-vs-one ( i.e., variables from one IMU-based signal were trained and tested using another IMU-based signal) was applied to determine the classification accuracy ( i.e., similarities between IMUs) and variable importance (score ranges: 0.0-1.0) measures. These values were then used to select the variables that best independently describe the rank in classification margin. The results from each IMU sensor placement based on the fractal values showed ‘moderate’ accuracy (50-75%), with the exception of two cases: the left shank yielded ‘good’ accuracy (80.1%) compared with the right shank, and the right thigh generated ‘poor’ accuracy (48.9%) compared with the left foot. No IMU location showed excellent accuracy (>90%). The results indicate that each IMU placement location has their own fractal patterns that are not similar to another IMU location in terms of sex classification. The analysis of the variable importance in the classification margin showed that most of the PSD resulted variables were classified as ‘most important’ compared with the DFA resulted variables. IMU sensors, and the associated analyses, could be used during military load carriage to evaluate changes in gait resulting from injury, fatigue or overtraining.