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

• DOI: 10.1249/01.mss.0000561603.03899.ac
• Publication Date: 2019-06-25T21:52:20Z
• Authors: Elliott Fullerton; Katherine Brooke-...; Paul Sanderson; Dale Esliger; Massimiliano Zecca

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