Intake24: SWIPE
TBC · TBC
Pichapat Leetongin
Dietary assessment is essential for understanding nutrition and shaping public health policy. Traditional methods such as 24-hour recalls are accurate but resource-intensive, while food frequency questionnaires are cheaper but less precise. Technology-based tools aim to bridge this gap, yet their reliance on detailed self-reporting can limit accessibility for people with low literacy or digital skills. This study developed and evaluated a machine learning–driven dietary assessment method that classifies individuals into nutrient profiles using simplified binary and portion-based questions derived from historical intake data. Using anonymised Australian Intake24 records (834,242 food entries from 49,954 participants), foods were clustered by macronutrient ratios and meal timing. The top 50 features were converted into yes/no questions with small, medium, and large portion options. XGBoost classifiers for protein, fat, and carbohydrate were trained and tested. The models achieved accuracies of 74.3% (protein), 71.2% (fat), and 73.8% (carbohydrate), with macro-averaged F1-scores around 0.70. This approach provides a low-burden alternative to conventional methods, with potential for adaptive questioning and cross-cultural use.

