Predicting parental Human papillomavirus vaccine hesitancy: development and internal validation of machine learning models

Objectives: Cervical cancer is Cameroon's second most common cancer among women. However, high parental Human papillomavirus (HPV) vaccine hesitancy impedes elimination, and no validated predictive tool exists. This study aimed to develop and internally validate machine learning models predicting parental HPV vaccine hesitancy in the Buea Health District and identify key predictors. Methods: This secondary analysis used cross-sectional survey data from 1,156 parents of children aged 9-18 years. Hesitancy was derived from self-reported vaccination status and intention. Twenty-six predictor groups were screened via group LASSO (17 retained for logistic regression; all 26 for tree-based models). Data were split 80/20 into training (n=925) and test (n=231) sets. Logistic regression, random forest, and XGBoost were tuned via repeated cross-validation while thresholds were set by Youden's J statistic. Discrimination, calibration, and Brier score were assessed on the test set. Predictions were interpreted using SHAP. Results: All models showed comparable discrimination (AUC-ROC 0.841-0.870). XGBoost had the highest AUC (0.870, vs 0.869 for logistic regression, DeLong p=0.965); logistic regression had the highest sensitivity (73.1%), F1-score (74.9%), and lowest Brier score (0.1435). Calibration slopes exceeded 1 (1.12-1.46) across models. Insufficient vaccine information, perceived vaccine unsafety, and distrust in the Ministry of Health were the dominant predictors. Discussion: The three models achieved comparable, clinically meaningful discrimination, and algorithmic complexity did not improve prediction over standard regression. Hesitancy was more strongly predicted by information access and institutional trust than sociodemographic disadvantage. Conclusion: These findings support trust-building engagement over broad demographic campaigns.