Respiratory symptoms drove caregiver-reported antibiotic use in febrile Bangladeshi children
Caregiver-reported antibiotic use among febrile children under five in Bangladesh clustered around respiratory complaints, but the prediction problem proved far less tractable than the associations alone suggest. In a nationally sampled analysis of 2,540 children with fever reported in the previous two weeks, recent cough and rapid breathing were both linked to higher odds of antibiotic administration, while older child age and reporting money as a major barrier to healthcare were linked to lower odds. The association pattern is clinically intuitive: respiratory symptoms often prompt families and frontline clinicians to suspect bacterial illness, even when viral syndromes are more likely, and that behavioral response can drive antibiotic exposure in settings where diagnostic confirmation is limited. Yet the same dataset also shows how easily symptom-driven decision-making resists reliable modeling when the outcome itself is based on caregiver report rather than verified prescribing records.
The DHS-weighted prevalence of caregiver-reported antibiotic administration was 12.2%. Recent cough carried an adjusted odds ratio of 2.30, with a 95% confidence interval of 1.62 to 3.25, and rapid breathing carried an adjusted odds ratio of 1.55, with a 95% confidence interval of 1.02 to 2.35. Age worked in the opposite direction, with an adjusted odds ratio of 0.987 per month and a 95% confidence interval of 0.978 to 0.997, meaning the odds of reported antibiotic use declined gradually as children grew older. Financial strain also mattered: perceiving money as a major barrier to healthcare was associated with lower odds of antibiotic administration, with an adjusted odds ratio of 0.67 and a 95% confidence interval of 0.49 to 0.93. In practical terms, the model is capturing a mix of illness severity cues and access constraints rather than any direct microbiologic signal.
Explainable machine learning identified signal, then lost it in unseen clusters
Investigators tested six feature selection methods and seven pre-specified classifiers across multiple feature sets using weighted, stratum-respecting, and primary-sampling-unit-disjoint development strategies. Logistic regression built on a consensus feature set drawn from seven domains was selected during model development, where it reached a ROC-AUC of 0.676 and an average precision of 0.223. Those values might appear modestly encouraging in an internal development frame, but the decisive test came in a frozen, PSU-disjoint holdout sample of previously unseen clusters. There, the model’s discrimination dropped to a ROC-AUC of 0.561, with a 95% PSU-bootstrap confidence interval of 0.480 to 0.649, and average precision fell to 0.144, with a 95% confidence interval of 0.106 to 0.207.
The calibration slope of 0.249 was particularly revealing. A slope that low indicates the model’s predictions were too extreme relative to observed outcomes, a classic sign that the algorithm learned structure specific to the development sample rather than a stable relationship that carried across new sampling units. The holdout ROC-AUC confidence interval also included 0.5, meaning discrimination above chance could not be established in the unseen PSUs. For a prediction model meant to work under a survey design with clustering, that is a hard stop rather than a minor statistical blemish. It means the model may describe the derivation data reasonably well while failing to deliver dependable ranking of risk in new communities.
Caregiver report is a narrow lens on antibiotic exposure
The biological and social drivers behind febrile illness management in young children are intertwined, and this analysis captures only part of that picture. Respiratory complaints plausibly prompt antibiotic use because cough and rapid breathing can signal pneumonia, the febrile syndrome most likely to trigger treatment when clinical assessment is brief and laboratory testing is absent. At the same time, rapid breathing can also accompany viral lower respiratory infections, bronchiolitis, or even nonspecific distress, so symptom recognition alone cannot separate bacterial from nonbacterial disease. Without clinical indications, microbial verification, or validated prescription records, the observed outcome remains a proxy for household recall of treatment rather than proof that antibiotics were medically indicated.
The access barrier finding is also important. Families facing financial obstacles may be less likely to reach care, purchase medications, or remember precisely what was given during an illness episode, which can pull reported antibiotic use downward for reasons that have little to do with true need. That combination of symptom recognition, care-seeking behavior, and recall uncertainty creates a prediction target that is inherently noisy. Explainable machine learning can still map which features seem to matter inside one dataset, but the external transportability of those signals depends on whether the same social and clinical pathways operate in the next cluster, district, or survey wave.
From an antimicrobial resistance perspective, the result is a reminder that misuse surveillance in low- and middle-income settings cannot rely on elegant models alone. If the outcome is measured indirectly, the algorithm may confuse exposure, access, and reporting behavior with prescribing decisions. The lack of specific funding and the use of publicly accessible secondary survey data do not weaken the analytic value of the work, but they do frame it as a study of prediction limits as much as a study of antibiotic use itself. The data support one clear conclusion: caregiver-reported respiratory symptoms and financial barriers are associated with reported antibiotic administration, yet the model did not generalize well enough across unseen PSUs to support dependable prediction in Bangladesh’s 2022 survey structure.