jpsicilia/climate-mortality-uganda-zindi

# Climate-Sensitive Mortality Classification — Iganga-Mayuge HDSS (Uganda) Supervised classification of whether a recorded death falls into a **climate-sensitive category**, using demographic, geographic and climate data from the Iganga-Mayuge Health and Demographic Surveillance System (Uganda, 2007–2022). [AI4EAC Climate & Health Risk Prediction Challenge, Zindi] ## TL;DR — the finding In a challenge framed as *climate*-health, **climate variables carry no predictive signal beyond age and secular trend.** The target is, in effect, a proxy for the age structure of cause of death (child vs. adult/elderly mortality). This is shown three ways: - **Ablation (spatial + random CV):** dropping every climate feature does not lower the score. - **Nested likelihood-ratio test:** adding 5 climate variables to a demographic+temporal logistic model is non-significant (LR = 6.63, df = 5, **p = 0.25**); AIC *worsens*. - **Residual test:** after the demographic model, no climate variable correlates with the residuals (all |r| < 0.03, p_FDR ≈ 0.88). The performance ceiling (~0.83 composite / ~0.845 AUC) is an **information ceiling, not a model ceiling** — confirmed empirically (gradient boosting, neural net and stacking all converge to the same AUC) and consistent with the verbal-autopsy literature, where high accuracy requires clinical/symptom data absent here. ## Method - **Model:** LightGBM, seed-averaged (10 seeds × 5 folds), `scale_pos_weight` tuned so the fixed 0.5 threshold sits at the F1 optimum (a modelling choice, not post-hoc thresholding). - **Features:** age (dominant, non-monotonic), secular trend (year), seasonality (day-of-year sin/cos), plus a small set of empirically-mined age×climate interactions (infant×rainfall, elderly×NDVI) that survive Benjamini-Hochberg correction. - **Validation:** random *and* village-grouped (spatial) 5-fold CV — they agree, confirming the demographic signal generalises to the spatially-disjoint test villages. - **Metric:** 0.6·F1( …