Development and internal validation of machine learning prognostic models of sports injuries using self-reported data in athletics (track and field): The influence of quantity and quality of features

To compare the performance of sports injury prognostic machine learning models when trained on (i) baseline data (i.e. collected once) vs. monitoring data (i.e. collected frequently over a period), (ii) raw monitoring data vs. time-integrating engineered features of the same data, and (iii) different numbers of features. Self-reported data collected during a previous randomised controlled trial in athletics athletes over 39 weeks constituted the dataset for model development. Baseline features, monitoring features, and two time-integrating feature engineering strategies were employed. Seven machine learning algorithms were trained with different groups and numbers of features and validated internally with bootstrapping. The models' discrimination was statistically compared using t-tests or Mann-Whitney tests (α = 0.00026). A dataset of 4537 cases including 149 injuries was derived from 165 athletes. Monitoring features outperformed baseline features in 5 out of 7 algorithms (<i>p</i> < 0.00026). The two feature engineering strategies showed marginal differences (1-8%) in 4 out of 7 algorithms (<i>p</i> < 0.00026). Larger numbers of features showed consistent improvements of performance for 6 out of 7 algorithms. Developing injury prediction ML models based on self-reported data in the sport of athletics seems promising but highly influenced by the quality and quantity of features.