Type 2 diabetes prediction method based on dual-teacher knowledge distillation and feature enhancement

Area of Science:

• Medical Health
• Machine Learning
• Data Science

Background:

• Diabetes prediction is vital for timely intervention, reducing health risks and costs.
• Existing methods require improvement in accuracy and reliability for effective diabetes prediction.

Purpose of the Study:

• To propose an advanced data preprocessing and feature enhancement method for diabetes prediction.
• To develop and evaluate a novel dual Convolutional Neural Network (CNN) teacher-student distillation model (DCTSD-Model) for type 2 diabetes classification.

Main Methods:

• Data preprocessing involved outlier removal and missing value imputation.
• Sparse Autoencoder (SAE) was utilized for feature enhancement, expanding data representation.
• A dual CNN teacher-student distillation model (DCTSD-Model) was developed, incorporating soft labels and weighted random samplers to address class imbalance.

Main Results:

• The SAE feature enhancement improved the expressive power of the original data variables.
• The DCTSD-Model achieved an excellent classification accuracy of 98.57% on the enhanced dataset.
• The model demonstrated superior classification ability and reliability compared to other evaluated models.

Conclusions:

• The proposed DCTSD-Model, combined with SAE feature enhancement and robust preprocessing, offers an effective solution for diabetes prediction.
• This method significantly improves accuracy and reliability in classifying type 2 diabetes.
• The findings provide a strong foundation for future research and clinical applications in diabetes prediction.