Comparing SHAP and distributed-lag model summaries of environmental associations with reported malaria cases in three Nigerian states

Explainable AI is increasingly used to interpret machine-learning models in public health, but is rarely checked against models that estimate environmental exposure–lag relationships directly. We compared the two on identical inputs. We analysed ten years (2015–2024) of routine monthly confirmed malaria cases from three Nigerian states spanning a low-to-high transmission gradient. A regularised LightGBM model trained on lagged satellite rainfall, temperature and vegetation was interpreted with SHAP; a distributed-lag non-linear model fitted to the same predictors provided a reference. The two agreed on which drivers mattered. Agreement on direction was only partial: the mean-absolute SHAP summary is sign-free, and a signed SHAP check matched the reference model on the leading driver (vegetation, positive) but not on rainfall. Agreement on the timing of effects across monthly lags was not reproducible—across rolling-origin resamples it was centred near zero for every driver and, against a permutation null, was not distinguishable from chance. This is consistent with strong collinearity between neighbouring monthly lags, though with only three collinear drivers we cannot isolate collinearity as the cause. Environmental variables did not improve out-of-sample prediction over a seasonal baseline, suggesting gains lie in finer spatial resolution rather than more covariates.