A time-sequenced approach to machine learning prognostic modelling with implementation on running-related injury prediction

The use of machine learning (ML) methods in medical prognostic modelling is gaining popularity, yet all currently available source models were designed from a general mathematics perspective. These models encounter limitations when embedding discipline-specific information, which restricts model practical interpretability. This study aimed to introduce two novel prognostic ML source models designed with an area-specific approach by testing their performance against commonly used ML methods, and explored their interpretability. Measurements associated with multidisciplinary risk factors (genetics, history, neuromuscular capacity, biomechanics, body composition, nutrition, training) were collected on competitive endurance runners, who were subsequently monitored weekly over 12 months for running-related injuries (RRIs). Data was fitted with commonly used ML methods and the novel models using a stratified 10-fold cross validation framework for performance comparisons. Interpretable feature interactions were tested for statistical significance, and extracted feature importance scores were tested for correlation with Shapley Additive Explanation (SHAP) values. 6,181 valid weekly samples were collected from 142 competitive endurance runners. The novel methods’ performances (AUC 0·736-0·753, Accuracy 0·822-0·849, Sensitivity 0·376-0·455, Specificity 0·859-0·896) matched those of commonly used ML methods (AUC 0·649-0·784, Accuracy 0·662-0·857, Sensitivity 0·337-0·568, Specificity 0·671-0·904). Pairwise feature interactions revealed stable patterns (p<0·001). Method-specific computationally efficient feature importance scores moderately correlated with SHAP values (r=0·12-0·72), showing increases as model parameterization increased. The novel methods showed comparable performance and better interpretability against common ML methods. Interpretability improved with increasing parameterization, suggesting performance may further improve with larger datasets and more features. Future research should perform higher quality validations using larger datasets before these methods can be widely adopted in prognostic modelling research. This research is supported by the Alan Turing Institute Enrichment Scheme and the China Scholarship Council.