Using Non-Linear, Machine Learning Methodology to Assess the Potential Metabolomic-Based Biomarkers of Total Fat and Percentage Fat Intake - A Thesis

Aug 5, 2026

Upload by Aurora

Non-linear tree-based machine learning models are evaluated for predicting daily intake of total fat and percent energy from fat using high-dimensional metabolomic profiles from fasting serum and 24-hour urine in a controlled feeding study. The dataset includes 153 post-menopausal women and combines serum and urine metabolites with demographic and Food Frequency Questionnaire information. Cross-validated comparisons with linear regression using LASSO show low predictive performance, with no significant differences between linear and non-linear models.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages51

Reading Time2 h 9 min

AccessFree to download

Abstract

Background

Bias and the search for objectivity in Nutritional Science research

Nutritional Metabolomics

Machine Learning: a promising modeling tool for Nutritional Epidemiology

Using non-linear, machine learning methodology to assess the potential metabolomic-based biomarkers of total fat and percentage fat intake using a controlled feeding study. Caroline Lea Nondin A thesis Submitted in partial fulfillment of the Requirements of the degree of Master of Science University of Washington 2023 Committee: Marian Neuhouser Eardi Lila Program Authorized to Offer Degree: Nutritional Sciences ©Copyright 2023 Caroline Lea Nondin University of Washington Abstract Using non-linear, machine learning methodology to assess the potential metabolomic-based biomarkers of total fat and percentage fat intake using a controlled feeding study. Caroline Lea Nondin Chair of the Supervisory Committee: Marian Neuhouser Nutritional Sciences Department Background: Understanding and identifying objective dietary biomarkers is a crucial component of nutrition research today. By investigating the relationship between biomarker profiles and dietary intake using machine learning methodologies, there could be a way to more objectively assess study participant nutrient profiles and better understand the relationship between nutrient intake and disease. Our aim in this thesis is to assess the utility of non-linear tree-based models in predicting daily intake of total fat and the percent of energy from fat from serum and 24-h urine high dimensional metabolites. Methods: Our analysis used the dataset from a 2-week controlled feeding study mimicking the participants’ habitual diets among 153 post-menopausal women from the Nutrition and Physical Activity Assessment Study Feeding study, conducted in the Women’s Health Initiative (WHI) . Fasting serum metabolite profiles, urine metabolites, as well as demographic and Food Frequency Questionnaire (FFQ) data, were used to predict total fat and percent energy from fat using four cross-validated tree-based machine learning models. A LASSO model for regression was used as a way to compare the models to a linear model. Results: The highest cross-validated multiple correlation coefficients (CV-R2) for total fat intake and percent energy intake were 10.2% and 10 .4%, respectively. None of the models had a CV-R2 of over 36% . There were no significant differences found between the performance of linear and non-linear models in predicting fat intake. Conclusion: Both linear and non-linear models were shown to be unable to predict total fat and dietary fat intake using serum and urinary metabolites accurately and reliably. Variable importance suggests that tree-based, machine-learning models have the potential to help understand non-linear interactions between biomarkers and dietary intake. 1. Introduction 1.1. Bias and the search for objectivity in Nutritional Science research Understanding the relationship between diet and the risk of disease is at the forefront of nutrition research today. However, research in Nutritional Science regarding these associations is hindered by our current methodological approaches. One crucial issue is the reliance on participant selfreport to assess their diet, such as the commonly used Food Frequency Questionnaire (FFQ) . The reliance on self-report in these methodologies can lead to systemic bias and measurement error that can influence our understanding of the diet-disease relationship 1. Self-report bias can be due to participants inability to fully recall their dietary intake, intentional misreporting, and/ or due to the natural variations in food preparations and quantities that are not reflected by the food composition database used to calculate energy and nutrient intake from the FFQs 1. There is, therefore, a critical need to incorporate objective measures, such as metabolomic biomarkers, into these epidemiological methodologies. Previous work from the Women’s Health Initiative (WHI) has provided a strong foundation for correcting self-report biases using regression calibration methods. These methods us

FAQs

What study goal does the thesis address?

To assess whether non-linear tree-based models can predict daily total fat intake and the percent of energy from fat from serum and 24-hour urine metabolite profiles.

What data and models are used for prediction?

A 2-week controlled feeding study dataset from 153 post-menopausal women is used. Four cross-validated tree-based machine learning models predict total fat and percent energy from fat, and a LASSO regression model is used to compare against a linear model.

How accurate were the models in predicting fat intake?

The highest cross-validated multiple correlation coefficients were about 10.2% for total fat and 10.4% for percent energy from fat. No model achieved CV-R2 above 36%, and there were no significant performance differences between linear and non-linear approaches.

Aug 5, 2026

Upload by Aurora