Generative AI models for type 2 diabetes mellitus risk prediction

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare
• Data Science

Background:

• Type II Diabetes Mellitus (T2DM) is a widespread chronic condition with significant global health implications.
• Accurate prediction of T2DM onset and risk is crucial for timely intervention and management.
• Challenges in T2DM prediction include data scarcity and class imbalance, hindering model performance.

Purpose of the Study:

• To enhance the accuracy of T2DM prediction using Generative Artificial Intelligence (GenAI) for synthetic data generation.
• To implement innovative feature selection methods for identifying optimal predictors of T2DM.
• To evaluate the performance of various Machine Learning (ML), Ensemble Learning (EL), and Deep Learning (DL) models on T2DM prediction tasks.

Main Methods:

• Utilized GenAI models, including Deep Tabular Augmentation (DTA) and Large Language Models (LLM), to generate synthetic data, addressing class imbalance and data scarcity.
• Employed the Representative Instances-based Fuzzy Rough Set Feature Selection (FRS-RI) method for effective feature selection.
• Trained and evaluated ML, EL, and DL models on diverse diabetes datasets (Sylhet, Obesity, Diagnostic Features) and benchmark datasets (PIMA, LMCH).

Main Results:

• Achieved high accuracy, precision, and recall scores across multiple datasets, with specific models reaching 100% accuracy on the Obesity dataset.
• Demonstrated strong generalization capabilities with enhanced test accuracies of 98.37% (PIMA) and 97.33% (LMCH) after applying the proposed techniques.
• Highlighted the effectiveness of GenAI-driven synthetic data and FRS-RI feature selection in improving T2DM prediction models.

Conclusions:

• The integration of GenAI for synthetic data generation and advanced feature selection significantly improves T2DM prediction accuracy.
• The developed methodology offers a robust approach for identifying individuals at risk of T2DM, with potential for clinical application.
• Emphasis on model explainability is vital for clinical trust and the justification of T2DM risk predictions.