Abstract
Background
The epidemics of type 2 diabetes mellitus and obesity pose major public health challenges in New Zealand and globally. While driven by global changes to the quantity, quality, availability, and marketing of the food supply, these conditions manifest uniquely in individuals, limiting the effectiveness of generic dietary advice. The efficacy of many dietary interventions in type 2 diabetes mellitus remains inconsistent, which may be due to individual variation in glycaemic responses to food, which traditional approaches do not account for. Personalised nutrition, which tailors dietary recommendations to an individual’s metabolic responses, offers a promising alternative. The potential of personalised nutrition has been unlocked by a series of significant technological advances. Continuous glucose monitoring has enabled high resolution glucose sampling that was previously impractical. Advances in artificial intelligence have allowed this high resolution glucose data to be processed, enabling the development of models that can predict post-prandial glycaemic responses to dietary intake. However, the accuracy of these models is constrained by the quality of dietary intake data used for training, which makes them more reliant on expensive and invasive genetic and microbiome data. Improving the accuracy of dietary intake data is an important step in unlocking the potential of personalised nutrition tools. Dietary intake data are gathered using dietary assessment methods. These methods are limited by under-reporting, where a participant’s reported intake is less than true intake, among other inaccuracies. Several promising strategies to reduce under-reporting have emerged, enabled by the widespread adoption of highly capable, camera-enabled smartphones: the use of images as the main recording method—termed image-based dietary assessment and software that automatically recognises food items in images. These strategies may enhance the usability, or user-friendliness, of the methods, an attribute which is under-emphasised in many research-grade dietary assessment methods. Improving the usability of dietary assessment methods may improve the accuracy of dietary intake data, thus bringing us closer to fulfilling the potential of personalised nutrition.
Objectives
Building effective personalised nutrition tools thus requires first addressing under-reporting in dietary intake data. This thesis begins with a mixed-methods study that tested if the use of personalised prompting—reminders that were sent based on the individual’s eating habits—improved the use of an image-based dietary assessment method, and began exploring the participant’s experience of using dietary assessment methods. The findings of this mixed-methods study informed the design and development of our own novel automated image-based dietary assessment App, designed with an emphasis on usability. Next, we conducted a pilot study comparing the energy intake estimates from the App against energy expenditure estimated using indirect calorimetry and accelerometry and against a 24-hour recall, a validated dietary assessment method. The participant experience of using the App was also explored, and was used to inform the next iteration of the App, including additional features and altering the design and layout of the App. We then integrated continuous glucose monitoring into the App, converting the iterated image-based dietary assessment App into a personalised nutrition tool. The personalised nutrition tool was then tested in a feasibility study to identify further iterative improvements to the App in preparation for future studies.
Results
The mixed-method randomised crossover trial described in Chapter 3 demonstrated that text prompts tailored to an individual’s typical meal times improved the completeness of image-based dietary records, suggesting that individual tailoring of prompts will improve accuracy of energy intake assessments. Participants preferred image-based to traditional text-based dietary assessment, but highlighted that an ideal dietary assessment App would have both inputs available. The pilot study described in Chapter 5 found that, compared to estimated energy expenditure, the energy intake recorded by the novel automated image-based dietary assessment App had an estimated mean bias of -1814 kilojoules (p=0.005). Compared to the 24-hour recall, the energy intake recorded by the App had an estimated mean bias of 783 kilojoules (p=0.33). Both comparisons had wide limits of agreement, limiting their utility at the individual-level. The participants had mixed responses to the user experience of the App. The most frequently cited limitation and source of frustration was the food composition database within the App, which either did not feature the food item required by the participant or featured it under a different name. The feasibility study described in Chapter 7 identified a number of iterative improvements to the personalised nutrition tool. Notably, participants expressed preference for more guidance to enable dietary behaviour change. This could be implemented in future iterations with the use of large language models. Compared to estimated energy expenditure, the personalised nutrition tool had an estimated mean bias of -1099.5 kilojoules (p=0.22), albeit with wide limits of agreement.
Conclusions
Tailored prompting can improve dietary record completeness, and the novel image-based dietary assessment App yields energy intake estimates comparable to validated methods, with accuracy improving through design iterations. Accurate dietary intake data enables participants to correlate their intake with real-time glycaemic responses, and study findings build towards an accurate, usable personalised nutrition tool to help optimise blood glucose levels for those living with pre-diabetes, type 2 diabetes mellitus, or any person seeking to improve their nutrition and metabolic health.