Evaluating the Predictive Performance of Metabolic Biomarkers for 20- Year Incidence Diabetes in the Yazd Healthy Heart Cohort: A Machine Learning Analysis
TL;DR
Conventional machine learning models using baseline data are insufficient for predicting diabetes incidence over long-term horizons, and the development of dynamic models that incorporate temporal changes in variables and complex pathophysiological interactions is essential for acceptable accuracy in long-term prediction.
Abstract
Background: Type 2 diabetes is a major global health challenge, necessitating the development of accurate prediction models for early intervention. This study aimed to assess the performance of machine learning models in predicting the long-term incidence of diabetes and to identify predictive metabolic biomarkers based on data from the Yazd Healthy Heart Cohort. Methods: This retrospective study was conducted on 906 non-diabetic individuals from the Yazd Healthy Heart Cohort with a 20-year follow-up. Five machine learning models (Logistic Regression, Linear Discriminant Analysis (LDA), Support Vector Machine (SVM), K-Nearest Neighbors (KNN), and Random Forest) were implemented based on 17 demographic, clinical, and biochemical variables. The data were randomly split into 70% for training and 30% for testing. Model performance was evaluated using 10-fold cross-validation and metrics including accuracy, sensitivity, specificity, F1-Score, and Area Under the Curve- Receiver Operator Characteristic (AUC-ROC). Results: During the study, 340 participants (37.5%) developed diabetes. None of the models achieved satisfactory performance (AUC> 0.8). The best performance was observed for the LDA model in the 10-year prediction, with an accuracy of 73% and an AUC of 0.70, although its sensitivity was low (47%). The composite indices triglyceride-glucose (TyG) and atherogenic plasma index (AIP) were identified as the strongest predictors across most models. Conclusion: Although composite biomarkers such as TyG show significant predictive potential, conventional machine learning models using baseline data are insufficient for predicting diabetes incidence over long-term horizons. The development of dynamic models that incorporate temporal changes in variables and complex pathophysiological interactions is essential to achieve acceptable accuracy in long-term prediction.