Comparing watershed-based spatial units and administrative boundaries for malaria prediction in the Peruvian Amazon

Malaria continues to pose a major public health threat in the Peruvian Amazon, where climate variability influences transmission patterns. This study assessed whether watershed-based spatial units (hydrobasins) outperform traditional administrative boundaries in predicting P. vivax and P. falciparum cases in Loreto, Peru. A retrospective ecological analysis was conducted using village-level geocoded malaria case data from 2009 to 2018, aggregated at the district level and three hydrobasin levels (levels 6, 7, and ANA). Environmental and meteorological variables were derived from satellite data and incorporated into Zero-Inflated Negative Binomial regression and machine learning models (Random Forest, Support Vector Machine, and XGBoost). Model performance was evaluated using Root Mean Squared Error (RMSE) and Akaike Information Criteria (AIC). Among 297,324 malaria cases analyzed (79.48% P. vivax, 20.52% P. falciparum), hydrobasin level 7 consistently showed superior predictive accuracy. For P. falciparum, hydrobasin level 7 consistently produced the lowest prediction errors across modelling approaches. Among machine learning models, XGBoost achieved an RMSE of 15.79 (95% CI: 10.42-21.03), whereas the corresponding ZINB-PCA model achieved an RMSE of 39.19 (95% CI: 17.09-62.24). District- and ANA-based ZINB-PCA models yielded RMSE values of 50.89 (95% CI: 23.65-77.87) and 59.22 (95% CI: 33.26-82.21), respectively. For P. vivax, XGBoost achieved an RMSE of 38.44 (95% CI: 26.90-49.19), whereas the corresponding ZINB-PCA model achieved an RMSE of 63.87 (95% CI: 46.38-80.95). District-, hydrobasin level 6-, and ANA-based ZINB-PCA models produced RMSE values of 102.81 (95% CI: 73.15-131.77), 175.87 (95% CI: 116.08-232.61), and 263.78 (95% CI: 144.98-389.87), respectively. These findings suggest that the choice of spatial unit affects predictive accuracy, with hydrobasin level 7 producing the lowest prediction errors; the associations captured by the models are predictive and are not interpreted as causal effects. Integrating hydrologically defined boundaries into surveillance systems may enhance malaria control strategies in environmentally complex regions.