Inferring high-fat dietary patterns from electronic health record data using machine learning.

Area of Science:

• Computational epidemiology
• Biomedical informatics
• Nutritional science

Background:

• Electronic health records (EHRs) lack detailed dietary information, hindering diet-disease research.
• Developing computable phenotypes for dietary patterns is crucial for leveraging EHR data.

Purpose of the Study:

• To develop machine learning (ML) computable phenotypes for identifying high-fat diets (HFD) using standard EHR variables.
• To assess the epidemiologic validity and clinical relevance of ML-derived dietary phenotypes.

Main Methods:

• Utilized National Health and Nutrition Examination Survey (NHANES) data (1999-2020) with 24-h dietary recall as ground truth.
• Trained ML models (Extreme Gradient Boosting, logistic regression, random forest) using EHR-compatible variables.
• Assessed model performance using F1-score, recall, and precision; evaluated clinical relevance by comparing cancer associations.

Main Results:

• ML models effectively classified HFD, with the random forest model achieving an F1-score of 0.79 at a cutoff of 20.
• Key predictors for HFD included race/ethnicity, triglycerides, obesity metrics, and metabolic panel results.
• ML-derived phenotypes reproduced known diet-disease relationships, demonstrating epidemiologic validity.

Conclusions:

• Dietary patterns can be inferred from routinely available EHR variables using ML-based phenotyping.
• This scalable approach enables the integration of dietary information into EHR-based research and precision medicine.
• The identified predictors align with biological pathways linking diet, obesity, metabolism, and cancer risk.