Abstract
BACKGROUND: Stroke is a major public health concern requiring valid estimates for planning and evaluating health interventions. The GBD (Global Burden of Disease) studies have become a major source of information; however, data sources have historically been a limitation. OBJECTIVES: We sought to compare stroke mortality estimates in Kazakhstan with those reported by the GBD study. METHODS: Mortality data were extracted from the Unified Electronic Healthcare System of Kazakhstan (UNEHS). We used the autoregressive integrated moving average (ARIMA), Bayesian structural time-series (BSTS), and Extreme Gradient Boosting (XGBoost) to model data from the UNEHS and forecast its trends until 2030. The accuracy metrics were mean absolute error, root mean square error, and mean absolute percentage error. We calculated the standardized difference in mortality estimates between the databases for the observed and forecasted estimates. RESULTS: The BSTS, ARIMA, and XGBoost models revealed slight variations in accuracy metrics, which depended on forecasting horizons and mostly favored XGBoost. During 2014-2030, the absolute difference in death counts was 207,108 between the GBD and UNEHS. The GBD estimates were twice as many across both the observed and predicted periods, with a moderate standardized difference (0.73) when considering their average. This study showed a systematic difference between GBD and national data. CONCLUSIONS: We found that UNEHS estimates were not comparable despite our efforts to replicate the GBD methods. Further studies are needed to explore the discrepancies between the national or regional data and GBD. Current limitations related to primary data and reproducibility require caution when interpreting GBD findings.