Artificial intelligence model as a tool to predict prediabetes

Abstract

Abstract Prediabetes is characterized by elevated blood glucose levels that are higher than normal but below the threshold for diabetes mellitus. While AI models have been used for prediabetes prediction, most rely solely on standard clinical and biochemical markers. This study introduces a novel Pattern Neural Network (PNN) model that uniquely integrates total antioxidant scavenging potential with traditional risk factors, providing new insights into the role of oxidative stress in prediabetes risk stratification among Indian adults. A total of 199 individuals aged 18 to 60 years were recruited and classified based on HbA1c levels into Control ( n  = 99) and Prediabetes ( n  = 100) groups. Fourteen input features including age, gender, total antioxidant status, HbA1c, FBG, OGTT, TGL, HDL, LDL, VLDL, TC, WC, Hb, and BMI were used to train a PNN with 14 input nodes, 10 hidden nodes, and one output node. The dataset was randomly divided into training, validation, and testing subsets. Model performance was compared against SVM, KNN, and LR classifiers. Feature importance analysis was conducted to interpret the model’s clinical relevance. The PNN model achieved superior validation performance with an accuracy of 98.3%, outperforming SVM (96%), KNN (83%), and LR (71%). Notably, antioxidant scavenging potential and waist circumference emerged as the most influential predictors. The model’s output value of 0.8770 (threshold > 0.5) effectively identified individuals at increased risk of developing diabetes. By incorporating oxidative stress markers, this study provides the first AI model in an Indian cohort to link antioxidant status to prediabetes risk. The PNN model demonstrates excellent accuracy and interpretability, offering a clinically actionable tool for early disease detection and personalized intervention.