Fastening the future: SHAP-enhanced machine learning on child restraint system adoption

Child passenger safety is a critical public health concern, particularly in low- and middle-income countries where road traffic injuries are a leading cause of child mortality. Evidence suggests that child restraint systems (CRS) reduce the likelihood of children being injured or killed in road traffic crashes, yet CRS uptake remains low in many countries, including Pakistan. This study examines socio-demographic, experiential, and contextual factors associated with CRS use at the individual level in Lahore, Pakistan. Using survey data, a machine learning framework with nested cross-validation and hyperparameter tuning using Optuna was applied to evaluate six models: Logistic Regression, Decision Tree, Random Forest, XGBoost, LightGBM, and CatBoost. Among these, CatBoost achieved the highest mean outer-fold ROC-AUC and was further analyzed using SHAP for interpretability, although its performance was statistically indistinguishable from Random Forest. For the CRS-Yes class, CatBoost achieved a precision of 0.841 and a recall of 0.632. SHAP analysis indicated that education level, attending driving school, driving experience, minor past crash involvement, and household members contributed most to the model’s CRS-use predictions. These associations may help identify variables for further investigation and support the development of more targeted hypotheses for CRS awareness and policy research.