A Leakage-Free and Statistically Validated Machine Learning Evaluation Framework for Diabetes Risk Prediction Across Multiple Datasets
Year: 2026
Type: preprint
Abstract: Machine learning models for diabetes classification frequently encounter issues with data leakage and inconsistent validation methodologies, rendering performance evaluations unreliable. This study presents a machine learning framework that mitigates data leakage and integrates fold-contained pre-processing with multi-seed stratified 10-fold cross-validation to guarantee fair evaluation. We examined six models: Logistic Regression, Random Forest, Feedforward Neural Network, Backpropagation Neural Network, General Regression Neural Network (GRNN), and Artificial Bee Colony Optimised Feedforward Neural Network. We conducted analyses on the Pima Indians Diabetes Dataset and the Early-Stage Diabetes Risk Dataset. In Dataset 1, models attained ROC–AUC values ranging from 0.76 to 0.84, with the GRNN model demonstrating the highest performance (AUC = 0.837 ± 0.002). Logistic Regression (AUC = 0.832) and Random Forest (AUC = 0.829) produced similar outcomes. Dataset 2 exhibited superior performance, with GRNN achieving an AUC of 0.998, followed by ABC-FFNN (AUC = 0.991) and FFNN (AUC = 0.982). The calibration results indicated that GRNN excelled in probability estimation. Statistical analysis confirmed notable performance differences among selected model pairs. The proposed framework enhances the reliability and reproducibility of diabetes classification research, facilitating a more comprehensive machine learning evaluation in healthcare environments.
Source: Research Square
Authors: Amresha Soomro, Oguz Ata
Institutions: Altınbaş University, Atlas Üniversitesi
Language: English
Subfield: Health Information Management
Field: Health Professions
Domain: Health Sciences
Open Access status: green
OpenAlex ID: W7163027958