Predicting Military Training Outcome: A Prospective Cohort Study of Supervised Statistical Learning Classification Methods in British Army Recruits

Purpose: Investigate predictors of failure and injury during British Army basic training and test the performance of statistical learning models to predict these outcomes. Methods: 968 male and 450 female recruits had demographics, lifestyle habits, mood, physical performance, body composition, micronutrient status, and bone metabolism measured at week 1 of training. Associations between predictors and training outcome (failure; lower limb overuse musculoskeletal injury) were assessed by logistic regression. Training outcome was predicted by logistic regression, elastic net, random forests, and gradient boosting. Results: 66% of men and 76% of women passed with 20% and 34% developing an injury. For men, slower 2.4 km run time (odds ratio [OR]=1.004, p=0.011) and higher symptoms of depression (OR=1.173, p=0.001) were associated with higher risk of failing, and lower areal bone mineral density (OR=0.008, p=0.002), slower 2.4 km run time (OR=1.006, p=0.011), previous stress fracture (OR=3.267, p=0.002), lower parathyroid hormone (OR=0.783, p=0.009), and lower c-telopeptide cross-links of type I collagen (βCTX) (OR=0.255, p=0.041) were associated with higher risk of injury. For women, slower 2.4 km run time (OR=1.009, p=0.009) and higher total 25-hydroxyvitamin-D (OR=1.013, p=0.001) were associated with higher risk of failing, and slower 2.4 km run time (OR=1.007, p=0.009), higher alcohol intake (OR=1.006, p=0.017), and lower βCTX (OR=0.057, p=0.026) were associated with higher risk of injury. Training outcome was poorly predicted in men and women (area under the receiver operating characteristic curve≤0.68). All models had low sensitivity and high specificity, but application of an outcome-specific threshold determined using the Youden Index improved sensitivity. Conclusions: Interventions to improve military training outcome should address aerobic fitness prior to starting training. Prediction models may be useful in identifying some ‘at-risk’ recruits.