Assessment of Machine Learning Performance for the Detection of Activity Type in Military Training

Abstract

Recognition of activities performed during military training may benefit the identification and quantification of factors that may predispose to the high prevalence of injury. There is evidence to suggest that the use of machine learning classifiers along with features from accelerometry data can achieve accurate activity recognition; however, there is no evidence to this application within military activities. PURPOSE: To develop and determine the accuracy of decision tree (DT), support vector machine (SVM), k-nearest neighbour (KNN) and ensemble bagged tree (EBT) models to classify military training type activities. METHODS: 15 male participants (mean ± SD: age: 25.9 ± 3.0 height: 177.9 ± 6.8cm body mass: 80.9 ± 8.7 kg) completed three sessions that consisted of performing military activities (walking, running, marching, weighted marching, halt to attention, countermovement jump and sedentary) with a low cost accelerometer (Axivity AX3, UK) mounted on the distal third of the medial tibia. Accelerometer data were segmented into two-second windows with a 50% overlap to introduce activity variance. Raw data along with filtered (butterworth, chebyshev and elliptic) were processed through a variety of features and classifiers (DT, SVM, KNN, EBT). Models were trained (80%) and hold-out validated (20%) using the classification learner within MATLAB (MathWorks Ltd, UK). Accuracy was determined by the percentage of true values during validation. RESULTS: 40,207 two second episodes of activities were recognized (1340 minutes). Hold-out validation accuracy for the EBT model and raw data (no improvement through filtering) was 0.96 (95% confidence interval (CI), 0.96- 0.96). Other models demonstrated good validation accuracies [DT - 0.90 (95% CI, 0.88- 0.91), SVM - 0.94 (95% CI, 0.93-0.95) and KNN - 0.91 (95% CI, 0.90-0.92)]. Validation accuracy was moderate to excellent (>80%) for walking and excellent (>90%) for all other activities. CONCLUSIONS: All machine learning models (especially EBT) provided excellent classification accuracy with the use of a tibial mounted accelerometer. These low-cost sensors and models thus offer potential for characterising military activity and examining relationships of activity parameters with injury. Supported by EPSRC and Loughborough University Studentship 1814563