Machine Learning Models Can Predict Performance Outcomes In A Military Lift-To-Place Task

The lift-to-place movement is commonly performed in military environments by loading objects, such as ammunition boxes, onto vehicles. The Australian Defence Force requires active personnel to lift a minimum of 20 kg to a 1.5-metre-high platform to test this ability. PURPOSE: To determine if a machine learning (ML) model could predict lift-to-place task performance from anthropometry and physical performance tests. METHODS: 25 healthy participants (male: 20, female: 5, age 27.2 ± 7.9 years, height 1.74 ± 0.08 m, mass 75.7 ± 12.6 kg) completed a physical test battery and lift-to-place task to failure. Data from VALD, Styku, and Polar (biomechanical, physiological, and anthropometric data) systems were cleaned and split into training (80%) and test (20%) sets. To assess the feasibility of 8 models, a leave-one-out cross-validation recursive feature elimination was performed on the training set. Feasibility was determined by the minimum features selected and the root mean square error (RSME). The ability to optimise and tune the model was also considered. The most feasible model was optimised via hyperparameter tuning to determine prediction accuracy for the testing set. RESULTS: The XGBRegressor (XBG) model was most feasible, with 5 features (jump kinetics, shoulder flexion, and forearm circumference), an RSME of 3.85 kg, and greater optimisation tuning parameters than the other models (Figure 1). After hyperparameter optimisation, the RSME was reduced to 0.18 kg for the training set. The testing set resulted in a RMSE of 4.06 kg. CONCLUSION: The XBG model is a useful tool to predict the lift-to-place task from simple physical tests. By applying a 4 kg buffer to the model, the minimum military standards of the task can be predicted. As models are optimised like the XGB, military organisations may be able to rely on outputs to stream individuals into roles based on simple measures that predict job task capacity.Office of Naval Research Grant and VALD Perfromance Grant