Abstract
Type 2 diabetes (T2D) is considered a significant global health concern. Hemoglobin A1c level (HbA1c) is recognized as the most reliable indicator for its diagnosis. Genetic, family, environmental, and health behaviors are the factors associated with the disease. T2D is linked to substantial economic costs and human suffering, making it a primary concern for health planners, physicians, and those living with the disease. Saudi Arabia currently ranks seventh worldwide in terms of prevalence rate. Despite this high rate, the country lacks focused research on T2D. This study aims to develop hybrid prediction models that integrate the strengths of multiple algorithms to enhance HbA1c prediction accuracy while minimising the number of significant Key Performance Indicators (KPIs). The proposed model can help healthcare practitioners diagnose T2D at an early stage. Analyses were conducted in a case-control study in Saudi Arabia involving cases (patients with HbA1c levels ≥ 6.5) and controls with normal HbA1c levels (< 6.5). Medical records from 3,000 King Abdulaziz University Hospital patients containing demographic, lifestyle, and lipid profile data were used to develop the models. For the first time, we utilized recommended machine learning algorithms to develop hybrid prediction models to reduce the number of significant KPIs while enhancing HbA1c prediction accuracy. The hybrid model combining Random Forest (RF) and Logistic Regression (LR) with only 4 out of 10 KPIs outperformed other models with an accuracy of 0.93, precision of 0.95, recall of 0.90, F-score of 0.92, an AUC of 0.88, and Gini index of 0.76. The significant variables identified by the model through backward elimination are age, body mass index (BMI), triglycerides (TG), and high-density lipoprotein (HDL). The proposed model helps healthcare providers identify patients at risk of T2D by monitoring fewer key predictors of HbA1c levels, enhancing early intervention strategies for managing diabetes in Saudi Arabia.