Antimalarial Pharmacotherapy Gaps in Nigerian Children Under Five: A Predictive Machine Learning Analysis of Care-Seeking, Testing, and ACT Treatment Using NDHS 2023–24
Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
Background
Nigeria’s malaria burden is the highest in the world, both in terms of incidence and fatalities. Case management uses the WHO’s Test, Treat, Track (T3) approach, where failures can occur at three stages (care-seeking, diagnostic testing, and treatment appropriateness), which are often combined into a single coverage measure. In this study, we split the cascade into two outcomes and compared four predictive modeling approaches to identify determinants.
Methods
From the Children’s Recode file of the 2023–24 Nigeria DHS, we identified 3,962 children under 5 years of age who reported having a fever in the two weeks before the survey. Gap A (access) was defined as: seeking care from any source and receiving a diagnostic blood test. Gap B (quality) was limited to the 1,776 children who had received any antimalarial, and specifically defined as receipt of an artemisinin-based combination therapy (ACT) only. To predict both outcomes, we tested and compared survey-weighted logistic regression, elastic-net regression, random forest, and gradient boosting (XGBoost) using a single held-out test partition and the area under the receiver operating characteristic curve (AUC) for all four models across both outcomes.
Results
The proportion of febrile children with combined access threshold was 16.0%, and 60.2% of children treated with antimalarials received an ACT. The Gap A logistic model did not reach overall statistical significance (p = 0.189); the Gap B model did (p = 0.003), with caregiver-reported financial barriers to care significantly associated with lower odds of ACT receipt (β = −0.97, p = 0.013), a finding that replicated across three of the four modeling approaches. AUCs were between 0.48 and 0.62 for the eight models, and all 95% CIs contained 0.50, meaning that none of the models could be confidently ruled out as being no better at discrimination than chance. For both outcomes, gradient boosting had the lowest AUC of the four methods, similar to what has been previously reported for a comparable treatment-cascade outcome in Mexico’s ENSANUT survey.
Conclusion
Limited variation in Nigeria’s malaria care-seeking, testing, and treatment cascade can be explained by household and maternal characteristics captured in standard survey data. The consistent association between financial access barriers and reduced likelihood of ACT-specific treatment suggests point-of-sale cost, rather than caregiver knowledge alone, as a target for intervention within Nigeria’s pharmacy and patent/proprietary medicine vendor distribution channels.
