Cross-Country Transportability of Machine Learning Models for Childhood Malaria Infection Prediction: Reciprocal Validation and Recalibration in Nigeria and Mali

Background: The cross-country transportability of childhood malaria prediction models remains unclear across populations with different prevalence and environmental conditions. Nigeria and Mali share correlates of malaria infection. We evaluated correlates of malaria rapid diagnostic test (RDT) positivity and the transportability of four machine learning algorithms using nationally representative 2021 Mali and Nigeria Malaria Indicator Surveys Methods: We included children aged 6–59 months with valid RDT results (Mali, n = 8,730; Nigeria, n = 10,717). Survey-weighted multivariable logistic regression identified correlates. We assessed regularised logistic regression, random forest, XGBoost, and an artificial neural network using five-fold cluster-aware cross-validation, reciprocal zero-shot external validation, and cross-fitted local recalibration. Results: Weighted RDT positivity was 19.4% in Mali and 39.6% in Nigeria. Increasing age, lower wealth and maternal education, larger household size, and unprotected or surface water were shared correlates. Internal discrimination was moderate (ROC-AUC 0.690–0.719), with complex models consistently and moderately outperforming regularised logistic regression. External discrimination was lower (ROC-AUC 0.668–0.689). Mali-trained models underpredicted Nigerian RDT positivity by 18.9–26.5 percentage points, whereas Nigeria-trained models overpredicted Malian positivity by 24.5–27.1 percentage points. Recalibration reduced mean prediction gaps to approximately zero with slightly improved discrimination Conclusions: Cross-country prediction requires external validation and representative local recalibration before wider public health use and should not replace confirmatory testing.