This study aimed to develop and internally validate predictive models for parental HPV vaccine hesitancy in the Buea Health District, Cameroon. The motivation was that cervical cancer is the second most common cancer among women in Cameroon and that high parental hesitancy impedes vaccine-based prevention efforts. The authors sought to identify key predictors and compare classical regression with tree-based machine learning algorithms.
The analysis was a secondary analysis of cross-sectional survey data collected from 1,156 parents of children aged 9–18 years. The outcome labelled “hesitancy” was derived from parents’ self-reported child vaccination status and stated vaccination intention.
Twenty-six predictor groups were assembled as candidate inputs. Group LASSO screening was used to select predictors for the logistic regression model, retaining 17 groups for that approach. Tree-based models (random forest and XGBoost) were fit using all 26 predictor groups.
Data were split into training (80%, n=925) and test (20%, n=231) sets. Models developed and compared included logistic regression, random forest, and XGBoost. Tuning used repeated cross-validation on the training set. Decision thresholds for classification were chosen using Youden’s J statistic. Model predictions were interpreted using SHAP (SHapley Additive exPlanations) to assess the influence of predictors on individual predictions and on model-level importance.
All three models produced comparable discrimination on the held-out test set, with AUC-ROC values spanning 0.841 to 0.870. XGBoost achieved the highest AUC of 0.870; logistic regression had an AUC of 0.869. The DeLong test comparing XGBoost and logistic regression returned p=0.965, indicating no statistically significant difference in AUC between those two models in this dataset.
On other performance measures, logistic regression performed well: it achieved the highest sensitivity (73.1%), the highest F1-score (74.9%), and the lowest Brier score (0.1435) among the models tested. Calibration slopes for the models exceeded 1, with reported slopes in the range 1.12–1.46.
These results indicate that algorithmic complexity (tree-based machine learning) did not meaningfully improve discrimination over standard logistic regression for this prediction task in the available data.
Model interpretation using SHAP highlighted a small set of dominant predictors. The most influential factors associated with parental hesitancy were insufficient vaccine information, perceived vaccine unsafety, and distrust in the Ministry of Health. According to the authors’ analysis, factors related to information access and institutional trust were stronger predictors of hesitancy than measures of sociodemographic disadvantage in this cohort.
Given the comparative performance of simpler and more complex algorithms, the authors suggest that implementers could rely on standard regression models for prediction while retaining interpretability. More importantly, the predictor pattern supports interventions focused on improving access to reliable vaccine information and building institutional trust—targeted engagement and trust-building may be more effective than broad campaigns based solely on demographic targeting.
This report describes internal development and validation only; no external validation was reported in the source. Details on variable definitions beyond the broad predictor groups were not provided in the summary. The authors state that all data from the present study are available upon reasonable request to the authors.
The study obtained ethics approval from the Ethics Committee and Institutional Review Board of the University of Buea (Ethics Registration number 2023/1949-02). The authors declared no competing interests. This manuscript was posted as a medRxiv preprint (posted September 20, 2026) with a DOI provided in the source. The source states adherence to reporting guidelines and that participant consent and relevant approvals were obtained.