REPOSITORY > RESULTS

Abstract

Food Frequency Questionnaires (FFQs) are widely used tools for dietary assessment in
epidemiological studies and digital health applications. While efficient and low-cost, FFQs
often suffer from incomplete responses, high respondent burden, and limited adaptability
to individual users. These limitations reduce data quality and compromise the reliability
of downstream dietary estimates. This PhD investigates how statistical and machine
learning methods can be used to improve the efficiency, robustness, and adaptability of
FFQs without sacrificing predictive accuracy.
The thesis focuses on three interconnected problems: reducing the number of required
questions, mitigating the impact of missing responses, and dynamically selecting the most
informative questions during questionnaire completion. First, feature selection and feature
ranking methods are applied to identify compact subsets of FFQ questions that preserve
predictive performance for dietary quality scores and nutrient intake estimates. Multiple
approaches are evaluated, including filter, embedded, and wrapper methods, demonstrating
that substantial questionnaire reduction is possible with minimal loss in accuracy.
Second, static question-ordering strategies are introduced to improve predictive quality
under partial questionnaire completion. By ranking questions by their predictive importance,
the most informative responses are collected early, thereby making dietary estimates
more resilient to incomplete data. This approach consistently outperforms conventional
missing-data handling techniques.
This thesis introduces a new method for dynamically ordering questionnaire items based
on expected utility. Instead of a fixed order, each next question is chosen based on the
answers given so far. The method combines predicted answer probabilities with measures
of how much each question is expected to improve predictions and reduce uncertainty,
allowing the questionnaire to adapt to each individual. The approach is further extended to
predict multiple dietary outcomes simultaneously using a single model, ensuring consistent
predictions across related targets. Weighted utilities are used to prioritise specific outcomes
when required.
The proposed methods are tested on data from the Extended Short Form Food Frequency
Questionnaire (ESFFFQ) and NHANES. Results on NHANES show the limits of
predicting laboratory biomarkers using FFQs alone, but the consistent patterns in performance
demonstrate that the approach is robust. Overall, this work shows that machine
learning can be used to make FFQs shorter, adaptive, and more reliable, and provides a
general framework for designing intelligent questionnaires in nutrition and other fields.