Рақамли технологияларнинг назарий ва амалий масалалари 9-jild 3-son (2026) · 125-133-betlar
Machine learning for early maternal risk stratification in Uzbekistan: develop-ment of a gradient boosting model based on DHS data
Эсонов, Ж.Х., Фозилова, М.
Annotatsiya
Maternal mortality in Uzbekistan remains marked by substantial regional heterogeneity: differences between the capital region and several peripheral areas reach three- to fourfold levels, highlighting the need for more accurate tools for early identification of pregnant women at high risk of complications. Most existing models for predicting maternal outcomes have been developed using data from high-income countries and do not always reflect the characteristics of resource-constrained healthcare systems, where access to clinical information, laboratory diagnostics and specialized care may vary considerably across regions. The aim of this study was to develop and internally technically validate a layered explainable machine learning model for individualized maternal risk stratification under conditions of limited availability of clinical data. The synthetic MaternaUZ v1 dataset was used as the analytical basis; it includes 4,812 observations and 26 features grouped into five cumulative layers: baseline, clinical, social, environmental and dynamic. The model was constructed as a sequential M1–M5 pipeline, in which each subsequent level expands the feature space and makes it possible to assess the additional contribution of a new data source. XGBoost was used as the core algorithm; class weighting, nested cross-validation, isotonic probability calibration and SHAP-based interpretation were applied to improve model robustness and transparency. Model performance was assessed using ROC-AUC, the Brier score and decision curve analysis. Progressive expansion of the feature space was accompanied by an increase in model discrimination from a ROC-AUC of 0.66 to 0.97 and improved probability calibration, with the Brier score decreasing from 0.21 to 0.06. Model interpretation showed that anemia, arterial blood pressure, the number of previous pregnancies and distance to the nearest hospital made the largest contributions to individualized risk prediction. The obtained results suggest that the proposed approach may serve as a promising basis for subsequent adaptation to national clinical data of the Republic of Uzbekistan, external prospective validation and integration into digital clinical decision support systems.
материнский рискмашинное обучениеобъяснимое машинное обучениеградиентный бустингXGBoostSHAPDHSстратификация рискаmaternal riskmachine learning
Metadata manbasi: jurnal OAI-PMH arxivi · Sindex toʻliq matnni saqlamaydi, manbaga havola beradi.