Bridging Ensemble Performance and Transparency: Knowledge Distillation for Obesity Classification on KNHANES Dataset

Area of Science:

• Machine Learning
• Public Health
• Biostatistics

Background:

• Obesity is a significant global health issue requiring accurate predictive models.
• Interpretable models are crucial for clinical applications in obesity management.
• Existing models often lack transparency, hindering clinical trust and application.

Purpose of the Study:

• To evaluate machine learning models for Body Mass Index (BMI) prediction and obesity classification.
• To develop an interpretable model for identifying key obesity-related factors.
• To enhance clinical decision-making in obesity management through transparent AI.

Main Methods:

• Utilized Korea National Health and Nutrition Examination Survey (KNHANES) data.
• Compared various machine learning models for BMI regression and binary classification.
• Employed knowledge distillation: XGBRegressor (teacher) trained a DecisionTreeRegressor (student).

Main Results:

• XGBRegressor showed high performance in binary classification (AUC).
• Knowledge distillation significantly improved the student Decision Tree's performance.
• The distilled model provided interpretable, rule-based predictions, identifying insulin resistance (HOMA-IR) as a key factor.

Conclusions:

• A distilled decision tree model offers a balance of predictive accuracy and interpretability for obesity.
• This approach aids clinicians in identifying critical obesity risk factors like insulin resistance.
• Enhanced model transparency supports targeted interventions for obesity management.