An Explainable Prediction for Dietary-Related Diseases via Language Models

Area of Science:

• Nutritional Epidemiology
• Computational Public Health
• Dietary Pattern Analysis

Background:

• Metabolic health outcomes are influenced by dietary patterns.
• Understanding diverse eating habits is crucial for public health.
• Existing methods may not fully capture diet-disease relationships.

Purpose of the Study:

• To explore the link between dietary patterns and metabolic health in Korean adults.
• To apply natural language processing (NLP) for dietary pattern identification and disease prediction.
• To assess the impact of NLP-derived indices on predicting obesity and dyslipidemia.

Main Methods:

• Utilized Latent Dirichlet Allocation (LDA) for dietary pattern discovery from KNHANES VII data.
• Developed NLP-based indices, including sentiment scores and identified patterns.
• Integrated these indices into machine learning models (XGBoost, LightGBM, CatBoost) for prediction.

Main Results:

• Identified three distinct dietary patterns: 'Traditional and Staple', 'Communal and Festive', and 'Westernized and Convenience-Oriented'.
• NLP integration significantly enhanced the predictive accuracy for obesity and dyslipidemia.
• Machine learning models demonstrated improved performance with NLP-enhanced features.

Conclusions:

• Dietary patterns are critical indicators of metabolic diseases.
• NLP offers a novel and effective approach to nutritional epidemiology and disease prediction.
• Findings support personalized nutrition strategies and targeted public health interventions.