Standardised and Objective Dietary Intake Assessment Tool (SODIAT): Protocol of a dual-site dietary intervention study to integrate dietary assessment methods

Study population

In randomised cross over study SoDiat-1 thirty male and female participants aged between 18-70 years were recruited by the research teams at Imperial College London and University of Reading, with an equal distribution per study location. Participants of all ethnicities with a body mass index (BMI) 20-30 kg/m² were eligible. Exclusion criteria were as follows:

Recruitment

Various methods of recruitment were employed, including distributing posters around the university campuses, emails sent to the respective clinical unit’s volunteer databases and university mailing lists, and social media advertisements for groups within the universities and surrounding areas. Recruitment started in December 2023 and finished in May 2024, when required sample size was achieved. Recruitment materials included a link and QR code that took interested participants to REDCap, a secure web application for administering online surveys and recording datasets in research studies (https://www.project-redcap.org/), where they could view and download the participant information sheet.

Study design

Screening

Interested participants completed an online screening questionnaire on REDCap, after which the respective research teams determined their eligibility and/or requested further information from the interested volunteers.

Consent and pre-study visit

Prior to starting the study, participants attended the NIHR Imperial Clinical Research Facility at Hammersmith Hospital or the Hugh Sinclair Unit of Human Nutrition (HSUHN) at the University of Reading (depending on their preferred location as specified on the screening form). Researchers explained the study in full, reconfirmed eligibility (such as ensuring all food items on the menu can be consumed) and allowed the participant to ask questions before informed written consent was taken. During the pre-study visits, participants were provided with a study handbook and the technologies used during the study were explained and demonstrated to them. Their self-reported BMI was also confirmed by measuring height and weight using a bioelectrical impedance analyser (Tanita MC780 MA P (Imperial) and BC-418 (Reading), TANITA UK Ltd, UK) to ensure the participant was within the correct BMI range. If the volunteer was happy to proceed, they were invited to schedule the two 4-day study visits. They were also provided with urine kits to take home and a reminder checklist/log form for the evening/morning prior to their first visit. A schematic diagram (Figure 1) illustrates the study process from recruitment through to completion.

Figure 1. A schematic diagram of SoDiat Study-1.

Randomisation

After the pre-study visit, participants were assigned a study ID code and randomisation was undertaken using REDCap. Participants were randomly allocated to one of two diet orders: Diet 1 followed by Diet 2 or Diet 2 followed by Diet 1. Randomisation was stratified by study centres. The research team and participant was not be blinded to the randomisation, as study menus were provided in advance, making it clear which diet they would follow each week.

Study diets

Participants consumed two controlled diets, one per study period, provided in a random order: Diet 1: non-compliant with UK dietary guidelines (e.g., high in saturated fat, free sugars, and salt and low in fibre); Diet 2: compliant with UK dietary guidelines, e.g., within the dietary reference values for saturated fat (≤10% total energy (TE)), free sugars (≤5 %TE), salt (≤6 g/d) and fibre (≥30 g/d) ( Table 1). The diets were matched for energy, protein, total fat and carbohydrate. Foods and drinks selected for each diet were selected to allow investigation of specific biomarkers.²¹ Each diet consisted of a 2-day repeating menu (e.g., menu A served on days 1 and 3, and menu B served on days 2 and 4) as shown in Tables 2 and 3. Bottled spring water was also be available to drink throughout the day. Meals and snacks were consumed at 2-hour intervals throughout the study days (9 am breakfast, 11 am morning snack, 1 pm lunch, 3 pm afternoon snack, and 5 pm dinner) and were served using identical tableware at each research unit (white crockery and clear glass on days 1 and 2 and patterned crockery and coloured glass on days 3 and 4); no foods/drinks were consumed directly from their packaging. Participants were also provided with a snack and bottled water to take home to consume before 8 pm and instructed not to eat or drink anything else before returning to the research unit the following day.

Study visits

The study visits took place at the NIHR/Imperial Clinical Research Facility at Hammersmith Hospital or the HSUHN at the University of Reading and included two study periods each consisting of four full-day (8 am to 6 pm) visits from Monday to Thursday with a short visit on the fifth study day (Friday) to return final urine samples and all study equipment. A washout period of at least one week between study periods was required (menstruating women attended study visits at the same phase of their menstrual cycle).

Study visit procedures

The day before starting each study period, participants were asked to restrict their caffeine and alcohol intakes and exercise levels to amounts that were usual for them and fast for 12 hours overnight (not consuming any food or drink, except water). Upon waking, participants also collected a first morning void (FMV) urine sample.

Participants attended the research unit at 8 am on day 1 (Monday) of each study period. Upon arrival, blood pressure, height and body weight were measured. A fasted capillary blood sample (OneDraw) was also self-collected and participants were set up with the wearable camera (which was worn continuously except if participants used bathroom) before being provided with breakfast. Habitual diet was recorded using the eNutri tool (4-week version). With the exception of mealtimes, participants had the rest of the day as free time but remained in the research unit. At the end of the study day (6 pm), participants were provided with bottled water and a snack for the evening, urine kits (for last evening void (LEV) and FMV samples) and a reminder checklist/log. Days 2-4 repeated day 1 except: 1) Intake24 was used on days 2-4 to record dietary intake during the previous 24 hours, 2) 4-day version eNutri was repeated at the end of day 4 (Thursday) to record dietary intake during the previous 4 days, and 3) capillary blood samples were not collected on day 3 (Wednesday). On day 5 (Friday), participants collected a final FMV, completed Intake24 and returned samples/equipment to the research unit.

Participants’ compliance to the study protocol was recorded by study investigators during the times when they were in the controlled environment. For the times spent outside the research unit, compliance was measured using sample/data collection records and self-reported deviations to study menus.

Objective dietary assessment tools

Spot urine samples

Participants collected spot urine samples using previously described methods.²¹ For each collection, participants were provided with four additive-free vacuum collection tubes (4 ml) (plus two spares), urine transfer straws and a disposable collection cup ( Figure 2). Participants collected their FMV urine and LEV urine for all four study days in each study period as well as a FMV sample on day 5 and LEV on evening before day 1 using the collection cup. Participants would then transfer samples to four tubes via the transfer straw and store at 4 °C. During each study day, samples were processed in the research unit using previously described methods to render them acellular then they were stored at -80 °C until the end of the study.²⁶

Figure 2. Urine transfer and straw kit.

Urinary biomarkers of dietary intake will be measured at Aberystwyth University using a combination of Ultra-High Performance Liquid Chromatography (UHPLC) Triple Quadrupole Mass Spectrometry (QqQ-MS) and high-resolution mass spectrometry (HRMS). Previous studies have determined a list of dietary intake biomarkers that reflect intake of common UK diet components and are sufficiently robust and reproducible in spot urine samples from dietary intervention studies.²⁷ Spot urine samples are normalised based on specific gravity prior to extraction. Specific gravity correction factors are calculated for urine samples as a fold change of each urine specific gravity to a value of 1.006. Global dietary patterns will be assessed by measuring the urine samples using Proton Nuclear Magnetic Resonance (¹H-NMR) at Imperial College London. A global dietary score will be generated from the 1H-NMR urinary metabolic profiles following a previously validated methodology²⁸ that will indicate the quality of the diet in combination with a complementary set of urinary dietary biomarkers.

Capillary blood samples

Capillary blood samples were self-collected by participants prior to breakfast on days 1, 2 and 4 using a OneDraw kit (Drawbridge, Thorne Research, Summerville SC, US) as shown in Figure 3. The single-use device attaches to the upper arm or thigh via a hydrogel adhesive and vacuum and collects 150 μl of capillary blood with little discomfort for the participant. The capillary blood is directly collected onto two paper strips.²² When the collection is finished, the cartridge containing the blood samples is placed in the transport sleeve then left at room temperature for at least 48 hours to allow the blood to dry prior to storage at -80 °C.

Figure 3. OneDraw blood collection kit.

The dried blood samples will be extracted at the University of Cambridge using a standard protocol for dried blood spots²⁹ and the lipid profile will be analysed using a combination of UHPLC and HRMS, and lipids will be quantified against internal standards as published previously.²²

Wearable camera technology

To effectively capture the dietary habits and food-related activities of individuals in UK households, a comprehensive passive dietary assessment system has been meticulously designed for this study. This system is a fusion of both hardware and software components, each with distinct functionalities to enhance the accuracy and efficiency of dietary data collection.

Hardware

For this pilot study, prototype of the camera was developed - wearable camera capable of recording up to 20hrs called M2.1. The wearable camera device is a high-definition camera with a maximum resolution of 2592x1944 pixels, mounted on the side arm of lens-less eye glasses frames to align with the user’s viewpoint and is connected to a rechargeable powerpack whilst in use ( Figure 4). It is designed to be used during daytime, capturing images of the eating process at frequency of one image every 1.5 seconds. The camera is powered by a built-in STM32 microcontroller with a 32-bit arm processor. The device begins recording upon the insertion of an empty 128 GB SD card and the camera is turned on. The camera is turned on automatically once is connected to the external battery.

Figure 4. Wearable camera equipment.

Prior to breakfast on day 1, participants received a camera device mounted on a glasses frame (or they can mount this on their own glasses frame) and were instructed to wear this until they went to bed, with the exception of when it is not suitable for the camera to be worn (such as when getting dressed and using the bathroom) in which case the glasses were be temporarily removed and details noted on the camera log. Upon arrival to the research unit on days 2-5, the SD cards were removed from the camera device by the research team and uploaded in duplicate to two encrypted external hard drives. Whilst the cameras were not in use overnight, participants were instructed to fully charge the power packs and start wearing the cameras the following morning before returning to the research unit (days 2-4).

Software

To ensure the anonymity of individuals, all footage captured by wearable cameras will undergo pre-processing prior to analysis. Initially, a large foundation model known as the Recognize Anything Model (RAM)³⁰ will be employed. RAM specializes in image tagging and has been developed through extensive training on a large number of general images. This model will play a crucial role in identifying images captured by our customised wearable camera. Its function will be to detect the presence of food items within these images. Upon detection of food items, RAM will assign a ‘food tag’ to the relevant images. This tagging mechanism is both efficient and precise, ensuring that only images with clear food content are marked for inclusion. The images that receive a ‘food tag’ from RAM will then be segregated from the rest and earmarked for further analysis, forming the core dataset for the study. Meanwhile, we have also designed ‘excluding tags’ for our model, which include around 20 categories of items such as bathrooms, mobile phone screens, and PC screens, with the flexibility to add or remove items as needed. This design is intended to prevent sensitive items from appearing alongside food in the collected data. Furthermore, for the retained images, an additional layer of protection will be implemented by blurring the faces of the participants and other individuals residing with them, as well as any other visible phone and computer screens that were unintentionally missed in the previous step. This step will use YOLOv8,³¹ a deep learning technique renowned in the field of image recognition, to prevent the inadvertent disclosure of identities and personal information. Only after this pre-processing will the images be subjected to further analysis.

Following pre-processing is the food recognition phase, where we will leverage large language models (LLMs)³² capable of processing both text and images. These models have been extensively trained on vast datasets, allowing them to recognize a wide range of food types without the need for additional training or fine-tuning. Portion size estimation also plays a pivotal role in dietary assessment, and our approach is tailored to address the unique challenges associated with using a wearable camera that captures only red-greed-blue (RGB) images, without the depth information provided by stereo imaging. This brings us to the issue of scale ambiguity, which is a significant problem given that our system cannot rely on stereoscopic methods to estimate the volume of food items. To circumvent the need for users to place a reference object next to their food - which would be an inconvenience and could disrupt a natural eating environment - we are exploring the potential of leveraging large-scale artificial intelligence (AI) foundation models to learn the general context of objects and their environmental surroundings in relation to food. By understanding these contextual relationships, the model can make more accurate inferences about the portion sizes of the food being consumed.

Subjective dietary assessment tools

24-hour dietary recall

Participants completed a self-administered 24-hour dietary recall following each study day to measure both the accuracy of self-reporting and usability of repeated 24-hour dietary recalls. Participants used Intake24 (intake24.com), which is a validated, web-based, open-source computerised dietary recall system based on the multiple pass method.³³ The tool, currently maintained by the University of Cambridge, Monash University and Newcastle University, is used by the UK’s National Diet and Nutrition Survey Rolling Programme.^25,33 Participants used Intake24 to record everything they ate and drank the day before from midnight to midnight by using free text to list each food/drink item consumed per meal as part of the initial ‘quick list’. Next, the detailed pass stage involved searching Intake24’s food database for the closest match for each item then estimating their portion size using the images presented on the screen. Intake24 also prompted the participant about foods usually consumed together, e.g., if they recorded coffee then it will ask if they added milk and/or sugar, as well as frequently forgotten foods, such as condiments. Additional questions were presented if Intake24 identified long time gaps without food or very low energy intakes, and the recall ended with the participant reviewing their entries. Prior to completing Intake24 for the first time, participants were encouraged to watch the tutorial video (accessible via the Intake24 menu). At the end of the study, data was exported from Intake24 including the quantities and nutritional intakes for each food item recorded per recall, after which mean daily intakes per day per participant will be calculated.

FFQ

The eNutri web app, developed by researchers from the University of Reading, includes an FFQ based on UK diets.³⁴ The current version includes 157 food, drink and supplement items. For each food and drink item, users first select how often they consumed it during the previous 4 weeks from 10 frequency buttons (such as ‘not in the last 4 weeks’ and ‘once a day’). If consumed, they then select their typical portion size from 7 portion size photos/buttons. Certain items (n=37) also request additional details, for example, the type of milk (if any) added to their coffee and whether the item consumed was a low-fat or low-sugar variety (e.g., soft drinks, yogurts); each of these items also has an ‘I’m not sure’ option. Users also report frequency of use of salt (added at the table and/or during cooking) and 8 dietary supplements. Using this information, eNutri automatically calculates mean daily intakes (g/d) of each food item, from which it estimates a large range of food group intakes (e.g., vegetables, dairy) and nutrient intakes (e.g., protein, vitamin C). In addition to dietary intake, eNutri also records certain demographic and lifestyle information about the users (such as age, sex, ethnicity, education level, physical activity levels and smoking status).

For this study, participants used the eNutri FFQ tool on day 1 of study period 1 to measure their habitual diet by recording what they ate and drank during the previous 4 weeks. Prior to using eNutri for the first time, participants watched the short tutorial video on the web app. To measure dietary intake during the two 4-day study periods, a separate version of eNutri was created that adapted the frequency options to reflect 4 days of dietary intake (such as ‘not in the last 4 days’ and ‘1x in the past 4 days’). This was completed on day 4 of each study week.

Participant feedback and usability of technologies

Participants completed the system usability scale (SUS) questionnaire following their first use of eNutri (day 1, week 1) and Intake24 (day 2, week1) via REDCap. The SUS questionnaire is widely-used to “measure people’s subjective perceptions of the usability of a system” and comprises of 10 alternating positive and negative statements regarding the user experience ( Table 4).³⁵ For each statement, respondents rate their agreement on a 5-response-scale from “Strongly Disagree” to “Strongly Agree” and, using the method described by Brooke (1995), a SUS score ranging from 0 to 100 is calculated, with higher scores indicating better usability.³⁶ Overall usability is also evaluated with a general question: “Overall, I would rate the user-friendliness of this system as:” with 7 options from “Worst Imaginable” to “Best Imaginable”.

At the end of the study (day 4, week 2), participants also provided feedback about all of the tools used during the study. This included free text and Likert questions and were completed via REDCap.

Statistical analysis

Previous studies have shown that the misreporting rate of total energy expenditure between self-reported dietary assessment tools and double labelled water was 35 % with a standard deviation of 33 %.³⁷ To reduce the misreporting rate to <10%, an a priori sample size calculation determined that to achieve a power of 80 % with a type 1 error of 5%, a total of 27 participants are required. This was increased to 30 participants to account for potential dropouts. Two separate clinical research centres were used, with both centres recruiting participants to ensure consistent demographics between sites (sex, age, and BMI).

The FFQ data collected at baseline will be used to describe cohort demographics and serve as a covariate in analysis to account for differences in habitual dietary patterns among participants. Additionally, the data from 4-day FFQ will provide average dietary intake data for the two study weeks, which can be combined with daily aggregated data, such as Intake24. This data will be compared with metabolomic data and images to assess consistency and identify any biases. Such comparisons are crucial for evaluating the alignment of self-reported dietary data with objective biomarkers.

Bootstrapped Lin’s concordance correlation coefficient (CCC) with 95% confidence intervals will be used to test the extent of agreement between each dietary assessment tool/technology and the nutrient composition of known diets and recorded compliance.

Primary and secondary outcomes to be measured

The primary outcome will be the accuracy of dietary reporting, measured at the end of each intervention week using the known quantities of consumed foods given to participants during the intervention days and dietary data collected from wearable cameras, spot urine samples, capillary blood samples, and self-reported dietary assessments.

Secondary outcome measures include: 1) the creation of a multiplatform model of dietary intake using g/day measured from wearable cameras and self-report dietary assessments, and μg/ml of dietary exposure biomarkers from spot urine samples and capillary blood samples at the end of the study, and 2) the design of a dietary intake study in a free-living population that will be informed by the results of the current dietary intake study protocol.

Dissemination

The results of this study will be presented at medical meetings, research conferences and published open access in peer-reviewed scientific journals and lay publications, approximately six months following the end of the study. They will also be used by research students who are associated with this project in work that will contribute to their degree (BSc, MSc, PhD) or other qualification, and shared with the press and media. The datasets will be made available and deposited in public databases at the point of publication. All data will be released within 2 years of the project’s completion and will be made accessible.

Ethical approval

The study received a favourable opinion for conduct by the Camden & Kings Cross Research Ethics Committee (23/LO/0437) on 4^th July, 2023 and the University of Reading Research Ethics Committee (23/19) on 19^th May, 2023. The study will be conducted according to the principles expressed in the Declaration of Helsinki.

Ethical considerations

Prior to screening, potential participants received an ethically approved participant information sheet containing full details of the study. They had adequate time to consider taking part in the study and had an opportunity to ask questions before attending the pre-study visit where they were given written informed consent.

When participants wore the cameras, they collected images of the participant, people around them and their devices (e.g. smartphones). To ensure everyone’s anonymity and privacy, any people and device screens recorded on the images will be automatically blurred prior to analysis, as described above. In addition, any non-food related images will be removed from the dataset. Both processing steps will be achieved via an artificial intelligence methodology and only the pre-processed dataset will be analysed by the research team. Regular audits of AI pre-processing will be conducted, with results evaluated through sampling to ensure compliance and address any privacy concerns that may arise during the experiment. Participants were also advised to remove their camera when it is not appropriate to wear them (such as when in the bathroom and dressing) and to log these instances.

Data management

The data collected through the SODIAT-1 study was pseudonymised and anonymised. Pseudonymised data was shared among research partners for data analysis purposes. The confidentiality of study participants was preserved under the Data Protection Act. Acellular urine samples were transferred to Aberystwyth university and dried blood samples were sent to the University of Cambridge in compliance with the Human Tissue Authority (HTA) regulation. The data generated from the wearable cameras was stored on encrypted hard drives and transferred to Imperial College London after the data collection was completed. For subjective dietary assessment tools eNutri and Intake24, participants used pre-generated weblinks and/or login details to avoid using personal information such as email addresses and names. Other data was input on REDCap by study researchers, which was double checked by the study coordinator at each site before the records were locked.

Study status

Recruitment for this study concluded in May 2024, with data collection completed in mid-June 2024. At the time of paper submission, blood and urine samples, as well as camera images, have been transferred to the respective research teams, and data analysis is currently in progress.

Trial registration

The study was registered at ISRCTN (ISRCTN13562899).