Village-Disjoint Validation and Parsimonious Feature Selection for Climate-Sensitive Child Mortality Risk Prediction in Uganda

Entries to the AI4EAC Climate and Health Risk Prediction Challenge are scored by a fixed composite of the F1 measure at a decision threshold of 0.5 and the area under the receiver operating characteristic curve. Because the official test partition is dominated by villages absent from the training data, random cross-validation overstates generalisability. This paper describes a regularised tri-model gradient boosting pipeline evaluated under a fifteen-fold, village-disjoint protocol. A distributional audit shows that meteorological covariates are closely matched across partitions whereas terrain covariates are not, and a coordinate-only nearest-neighbour baseline performs at chance level. The selected ensemble obtains a composite score of 0.8234 locally and 0.8354 on the public leaderboard. A first rainy-season indicator is the only candidate feature whose gain replicates across four independent village partitions. Reanalysis daily shocks, district target encoding and denoising autoencoder embeddings are rejected on the same evidence. Because the protocol's own seed-to-seed standard deviation is 0.00114, the size of any claim supported by a single partition is bounded, and that bound is respected throughout.