Type 2 Diabetes Prediction Method Based onDual-Teacher Knowledge Distillation and FeatureEnhancement

Abstract Diabetes prediction is a critical issue in healthcare, and early identification is particularly important as the number of peopleaffected by the disease continues to increase. Many patients fail to recognize their health problems in time, leading to latediagnosis, which in turn causes a large number of health problems and deaths every year. Therefore, it is crucial to developeffective methods for early diagnosis of diabetes.This study proposes a data preprocessing method that includes outlierremoval, missing value filling, and sparse autoencoder (SAE) feature enhancement. We propose a Dual-CNN Teacher-StudentDistillation Model (DCTSD-Model) for type 2 diabetes classification, aiming to improve the accuracy and reliability of prediction.SAE is used to expand the variables of the original data and enhance the feature expression ability. The proposed CNN andDCTSD-Model models are evaluated using 10-fold cross validation on the feature-enhanced dataset.The results show that afterusing SAE feature enhancement, DCTSD-Model generates soft labels through knowledge distillation of the dual teacher model,helps the student model learn rich category information, and solves the category imbalance problem through the weightedrandom sampler. Experiments show that the accuracy of DCTSD-Model in classification tasks reaches 98.57%.This model issignificantly better than other models in terms of classification performance and reliability, providing an effective solution fordiabetes prediction and laying a solid foundation for future research and applications.