Predictive Modeling of Secondary Microvascular and Macrovascular Complications in Type 2 Diabetes Patients: A Longitudinal Cohort Analysis Utilizing Optimized Gradient Boosted Trees

• Abilly Elly

  Texas University

  Author

Type 2 diabetes mellitus (T2DM) affects millions globally, with microvascular and macrovascular complications representing the primary drivers of morbidity and mortality. Current predictive approaches often focus on single complications or employ conventional statistical methods that fail to capture the complex, co-occurring nature of diabetic vascular complications. This study addresses this gap by developing and validating an optimized Gradient Boosting Decision Tree (GBDT) framework for predicting secondary microvascular and macrovascular complications in T2DM patients using a 10-year retrospective longitudinal cohort from a national diabetes registry. The proposed framework integrates multidimensional clinical and laboratory indicators, including coagulation profiles, cardiac enzyme panels, lipid profiles, and renal function markers, with advanced ensemble learning techniques. The optimized GBDT model achieved a superior predictive accuracy of 89.4% with an AUC of 0.92, significantly outperforming baseline methods including logistic regression (82.1%) and random forest (85.6%). Feature importance analysis identified urea, fibrinogen, prothrombin time, triglycerides, and fasting blood glucose as the most influential predictors, while SHAP analysis revealed distinct risk hierarchies for microvascular versus macrovascular outcomes. The framework demonstrated robust generalization across validation datasets with consistent calibration performance. This research provides a replicable, interpretable predictive tool that enables early risk stratification and targeted intervention strategies for diabetic complications, offering significant implications for clinical decision support systems and population health management.

• PDF

Copyright (c) 2026 Abilly Elly (Author)

This work is licensed under a Creative Commons Attribution 4.0 International License.