Comparison of optimized machine learning approach to the understanding of medial tibial stress syndrome in male military personnel

PURPOSE: This study investigates the applicability of optimized machine learning (ML) approach for the prediction of Medial tibial stress syndrome (MTSS) using anatomic and anthropometric predictors. METHOD: To this end, 180 recruits were enrolled in a cross-sectional study of 30 MTSS (30.36 ± 4.80 years) and 150 normal (29.70 ± 3.81 years). Twenty-five predictors/features, including demographic, anatomic, and anthropometric variables, were selected as risk factors. Bayesian optimization method was used to evaluate the most applicable machine learning algorithm with tuned hyperparameters on the training data. Three experiments were performed to handle the imbalances in the data set. The validation criteria were accuracy, sensitivity, and specificity. RESULTS: The highest performance (even 100%) was observed for the Ensemble and SVM classification models while using at least 6 and 10 most important predictors in undersampling and oversampling experiments, respectively. In the no-resampling experiment, the best performance (accuracy = 88.89%, sensitivity = 66.67%, specificity = 95.24%, and AUC = 0.8571) was achieved for the Naive Bayes classifier with the 12 most important features. CONCLUSION: The Naive Bayes, Ensemble, and SVM methods could be the primary choices to apply the machine learning approach in MTSS risk prediction. These predictive methods, alongside the eight common proposed predictors, might help to more accurately calculate the individual risk of developing MTSS at the point of care.