Predicting caregiver reported antibiotic administration among febrile children under five in Bangladesh using explainable machine learning

The use of antibiotics in febrile children is common practice in developing nations; however, the potential of routine household survey measures to predict antibiotic use based on caregiver reports remains unknown. This study identified factors associated with antibiotic use and investigated the transportability of caregiver-reported antibiotic use among febrile children aged 0–59 months in Bangladesh. The dataset used for this study was taken from the Bangladesh Demographic and Health Survey 2022; data included 2,540 children who had a fever as reported by caregivers in the past two weeks. The associations with antibiotic use, based on caregiver report, were analyzed using complex survey logistic regression with sampling weights, primary sampling units (PSUs), and strata. Six feature selection methods and seven pre-specified classifiers were tested across various feature sets using weighted, stratum-respecting, and PSU-disjoint development approaches. The final models were evaluated in a held-out PSU-disjoint sample of unseen PSUs. The DHS-weighted prevalence of caregiver-reported antibiotic administration was 12.2%. Recent cough (adjusted odds ratio [aOR] 2.30, 95% CI 1.62–3.25) and rapid breathing (aOR 1.55, 95% CI 1.02–2.35) were associated with increased odds of receiving antibiotics, while child age (aOR 0.987 per month, 95% CI 0.978–0.997) and perceiving money as a major barrier to healthcare (aOR 0.67, 95% CI 0.49–0.93) were associated with decreased odds. Logistic regression using a consensus feature set from seven domains was selected during model development (ROC-AUC 0.676; average precision 0.223). In the frozen PSU-disjoint holdout sample, the predictive performance was weak (ROC-AUC 0.561, 95% PSU-bootstrap confidence interval 0.480–0.649; average precision 0.144, 0.106–0.207), with a calibration slope of 0.249. The ROC-AUC holdout confidence interval contained 0.5, suggesting that discrimination above chance could not be established. Antibiotic use, as reported by the caregiver, was associated with respiratory complaints, the child’s age, and financial barriers to accessing healthcare. The predictive discrimination and calibration were poor for the previously unseen PSUs, suggesting that the model’s transportability under the BDHS 2022 sampling design is limited. Given that the outcome measure was caregiver-reported and that there was no access to clinical indications, microbial verification, or validated information on prescriptions, these results cannot confirm the clinical need for antibiotics.