Prediction of Low Bone Mass for Japanese Female Athletes Using Machine Learning

Attainment of optimal bone mineral density (BMD) in female athletes is a very important aspect of ensuring their well-being throughout their lives, and monitoring of BMD is essential to avoid fractures and bone-related injuries. Several tools for assessing bone health already exist, but either lack the practicality required for continuous monitoring or are not suited for the young female athlete demographic. Because of this, this study has the main objective of developing a binary classifier to discriminate between normal and low bone mass individuals among younger female athletes using features extracted from a questionnaire. The dataset consisted of data from 213 female athletes. Five different models were compared: logistic regression, decision tree, random forest, multi-layer perceptron and XGBoost. Validation was performed via cross-validation and feature importance was assessed via permutation importance. XGBoost showed the most balanced results in terms of sensitivity and specificity, achieving values of 0.93 and 0.62 respectively. It also obtained an AUC of 0.73 and an accuracy of 0.68. It was observed that the duration of the current period of amenorrhea, as well as the impact of the sport, showed the highest relevance, which is consistent with previous literature. Other features such as thinness level, number of training days in a week and age at menarche also showed high importance. The models demonstrated promising results in identifying low bone mass subjects from normal ones, indicating that features based on questionnaires can be an important source for evaluating low BMD in female athletes.