Abstract / Summary
This cross-sectional study aimed to develop and preliminarily validate a machine learning model for preliminary forward head posture (FHP) screening and deploy the final model as an online tool. Among 825 university students, 625 comprised the development dataset (training set, n = 437; test set, n = 188), and 200 the exploratory external validation dataset. FHP was defined as a craniovertebral angle < 50°. Six machine learning models were compared through nested cross-validation. Restricted cubic splines and Shapley additive explanations analysis were used for exploratory association analysis and model interpretation, respectively. Random forest (RF) achieved the highest sensitivity in nested cross-validation (0.748, 95% CI 0.661–0.825) and was therefore selected as the final model based on the prespecified priority of reducing missed cases. In the external validation, RF achieved an AUC of 0.827 (0.751–0.896), sensitivity of 0.745 (0.615–0.865), specificity of 0.784 (0.715–0.845), and Brier score of 0.125 (0.097–0.154). Body mass index contributed most to model output and showed a significant nonlinear association with FHP (adjusted p = 0.017). The model showed moderate-to-good discrimination for FHP among university students and may offer a relatively accessible approach to preliminary FHP screening.