Abstract
This study evaluates a reproducible framework for forecasting and retrospective monitoring of regional morbidity in Kazakhstan using harmonized annual official statistics. Ministry of Healthcare compendia were combined with demographic and living-standards tables from the Bureau of National Statistics. The panel comprises 1344 observations for 2012–2025 across 16 stable territories and six disease classes. Forty-two primary forecasting configurations and two sensitivity configurations were evaluated using expanding-window one-year-ahead validation, with selection based on 2017–2024 forecasts. Results for 2025 are exploratory because the outcome was inspected during model development. The selected specification combines the last-observation forecast with an Extra Trees residual correction, core lagged predictors, and a shrinkage weight selected within earlier temporal folds. Pooled mean absolute error (MAE) was 163.65 vs. 166.87 for the last-observation forecast, a 1.93% reduction; 2025 values were 108.05 and 110.69. The exact year-block test yielded a one-sided p-value of 0.066 and a familywise-adjusted p-value of 0.387. The hybrid reduced MAE in four of six disease classes. Grouped Shapley additive explanations (SHAP) decomposed the complete forecast. Monitoring used absolute error relative to preceding MAE; the categories are retrospective aids and were not validated against clinical outcomes or interventions.