Nutrition assessment methods

Valid dietary intake recording is key for nutritional intervention. The methods used for collecting food intake data can be classified in a number of ways. Based on the time of the collection, retrospective methods, such as the 24-h recall (24HR), require memory for recollection of foods eaten. In contrast, prospective methods require diet reporting as food is consumed, acting as food diaries.

An important and long-established retrospective nutrition assessment method is the FFQ. A FFQ is based on a list of popular food items in a specific region, and typically contains about 100 items⁽Reference Molag⁴⁾. FFQ have some accuracy limitations (e.g. energy estimation) but are more appropriate for initial dietary assessment for the purposes of personalised nutrition, because this method captures longer-term patterns in the diet. It has been used in epidemiological studies for decades, for example, the European Prospective Investigation into Cancer and Nutrition (EPIC) study, one of the largest cohort studies in the world with more than half a million participants recruited across ten European countries. The UK version of this FFQ is known as the EPIC-Norfolk FFQ⁽Reference Welch, Luben and Khaw⁵⁾, which was adapted for use in the Food4me project⁽Reference Forster, Fallaize and Gallagher⁶⁾, the largest study (n 1540) on personalised nutrition at the time of writing this review.

Diet quality scores

Most public nutrition campaigns aim to improve diets at population level through qualitative or simplified quantitative approaches, such as the Eatwell guide in the UK⁽⁷⁾. Similarly, food pyramid guides have been used to indicate the relative proportions of food groups (e.g. vegetables, salad and fruit v. fats, spreads and oils) that make up a balanced diet⁽Reference Davis, Britten and Myers⁸⁾. These tools are important but mainly designed to offer a visualisation of a balanced diet, and do not easily lend themselves to a computational analysis of the diet. In order to propose a more quantitative analysis of a dietary intake, indexes of diet quality have been developed over the past decades. Some focus on local guidelines and others on specific target groups or diseases. A comprehensive report containing existing indexes of diet quality was published in 2005⁽Reference Waijers and Feskens⁹⁾.

The Healthy Eating Index (HEI), revised in 2010⁽Reference Guenther, Casavale and Reedy¹⁰⁾, was developed by the United States Department of Agriculture. In some cases, specific indexes have been created to address particular diseases, such as the alternative HEI (AHEI) that predicts the risk of chronic diseases⁽Reference Chiuve, Fung and Rimm¹¹⁾. The AHEI calculations are clearly defined and quantitative⁽Reference Chiuve, Fung and Rimm¹¹⁾, making them well-suited for implementation in software. Furthermore, this index can be used with the data derived from different dietary assessment methods such as 24HR or FFQ.

Online personalised nutrition

Tailoring plays an important role in health interventions⁽Reference Noar, Benac and Harris^12, Reference Hawkins, Kreuter and Resnicow¹³⁾ and there is an interest amongst end-users in receiving personalised nutrition advice⁽Reference Wendel, Dellaert and Ronteltap^14, Reference Livingstone, Celis-Morales and Navas-Carretero¹⁵⁾. Some studies are investigating how to personalise nutritional guidance for more effective behaviour change⁽Reference Celis-Morales, Lara and Mathers¹⁶⁾. The Food4Me study was designed to determine the efficacy of different levels of personalised nutrition: based on diet intake alone; diet and phenotypic biomarkers; and diet, phenotype and genotype. Results showed that personalisation of food-based dietary advice based on an individual's dietary intake was more effective than generic public health advice, although in that study, no further benefit resulted from the addition of phenotypic or genetic advice⁽Reference Celis-Morales, Livingstone and Marsaux^17, Reference Ryan, O'Donovan and Forster¹⁸⁾. The Food4Me dietary advice was based on a decision tree executed manually by researchers during the study to create a report that was sent via e-mail. This decision tree was subsequently automated⁽Reference Forster, Walsh and Donovan¹⁹⁾, but their details have not been published. The lack of transparency in reporting nutrition interventions makes the reproducibility and comparison between studies more challenging and the improvements in this field slower. This difficulty was reported in 2017 by Warner et al. in a systematic review of telehealth-delivered dietary intervention trials⁽Reference Warner, Kelly and Reidlinger²⁰⁾.

Established technologies

In terms of online FFQ, the first online versions presented simple tables with food names in the rows and frequencies in the columns⁽Reference Forster, Fallaize and Gallagher²⁶⁾, very similar to the paper-based versions⁽Reference Mulligan, Luben and Bhaniani²⁷⁾. Some paper-based FFQ present images for the participants to select the portion sizes. This same strategy was replicated in the digital domain during the development of the Graphical Food Frequency System⁽Reference Kristal, Kolar and Fisher²⁸⁾.

The presentation of many frequency options and a few images to the users is challenging on small screen devices such as smartphones. The completion process is very repetitive, which may decrease the user satisfaction with the system. Another potential usability challenge is the presentation of the frequency options, since they do not necessarily all use the same units (e.g. two monthly, three weekly, one daily), making the decision-making process more difficult. The online Food4Me FFQ presented three food images together with seven radio buttons for the user to select the portion size (i.e. very small, small, small/medium, medium, medium/large, large, very large)⁽Reference Forster, Fallaize and Gallagher⁶⁾. Both the small radio buttons and the number of options presented simultaneously on the screen are not appropriate for smartphones. In addition, the Food4Me web design was not responsive, that is, it did not adapt to the screen size of the device. The visual design and layout of the FFQ options and their impact on usability was a research gap explored during the current project.

One project using 24HR is the ASA24⁽Reference Subar, Kirkpatrick and Mittl²⁹⁾, an automated self-administered system developed by the American National Cancer Institute. Three UK-based projects also used the 24HR method: Oxford WebQ⁽Reference Liu, Young and Crowe³⁰⁾, Intake24⁽Reference Delve, Simpson and Adamson³¹⁾ and MyFood24⁽Reference Carter, Albar and Morris^32, Reference Carter, Albar and Morris³³⁾. This method is able to capture how a dietary intake is distributed across meals but does not contain much information about food patterns and frequencies over a longer time period. A 24HR system can be adapted to work as a food diary, collecting dietary intake prospectively. For instance, MyFood24 can be used either retrospectively or prospectively.

A number of popular nutrition-related mobile apps, currently available for download, use a nutrition assessment strategy based on food diaries⁽Reference Franco, Fallaize and Lovegrove²⁵⁾. The aim of collecting detailed information about food consumption sounds promising, but it also demands time and discipline from users⁽Reference Thompson, Subar and Loria³⁴⁾. In order to simplify this process, some investigators have proposed alternatives such as the POND (pattern-oriented nutrition diary), which is a Food Index-Based Nutrition Diary⁽Reference Andrew, Borriello and Fogarty³⁵⁾. This study explored some alternatives to the food lookup (text search), using for example the ‘ + 1 button’ for items previously inserted (i.e. common items).

Different nutrition assessment methods can be combined during dietary intake recording, in order to increase the nutrient accuracy and also collect different aspect of the information needed (e.g. time of the meals and consumption patterns). A recent publication (2017) shows a new web-based tool (Foodbook24) which combines a 24HR, a FFQ and a supplementary questionnaire⁽Reference Timon, Blain and McNulty³⁶⁾. To improve the food and drink search, tags were applied in 484 out of the 751 items, so that brand names and misspellings could be linked to a similar food item. With these searchable tags, a specific food item may return in the search even if its name was not inserted correctly. For instance, the tag ‘Coca-Cola’ could be attached to coke, making both terms indexed in the search.

Inclusivity is another important aspect to be considered in public nutrition strategies. The NANA (Novel Assessment of Nutrition and Ageing) project included older adults in the design of a system specific for this population. The existing version of this system was successfully tested with older adults and validated from a nutritional perspective. It was a custom software app intended to be used off-line, and it involved the use of a camera and was not designed to provide feedback or advice to the users⁽Reference Astell, Adlam and Hwang^37–Reference Gilhooly³⁹⁾.

These studies propose some alternatives to established digital methods that still face many usability challenges. Most of the published studies in this field concentrate on the nutritional validity (accuracy against a gold standard method), but there is a lack of studies exploring how to increase the user acceptance of these digital methods. The amount of effort required to use these tools in relation to the perceived benefit from the user (self-monitoring or personalised nutrition advice) is one of the trade-offs that motivates this current work.

Emerging technologies

The need for technological innovation in dietary assessment has been reported in academic publications⁽Reference Thompson, Subar and Loria³⁴⁾. The main challenges are combining improvement in accuracy of assessment and user acceptance. One possible way to make progress in this trade-off is to propose more pervasive technologies, to improve user acceptability, decreasing the burden for users. A comprehensive review of digital methods was published in 2013 and may serve as a reference for the main emerging technologies applied in this field in recent years⁽Reference Stumbo⁴⁰⁾.

Food image recognition is perhaps the most promising technology for bringing convenience to dietary recording in a very scalable way since smartphones with cameras have become widespread. Some studies are focused on food recognition⁽Reference Probst, Nguyen and Tran^41, Reference Martin, Nicklas and Gunturk⁴²⁾ and Google has recently announced a research project named Im2calories in this area⁽Reference Meyers, Johnston and Rathod⁴³⁾. These experiments train the systems based on limited menus, from restaurants or schools, and try to expand their applicability to new contexts, but they still have unsatisfactory accuracy. To design and develop a system that could recognise foods and estimate their weights in any context, including home-made meals, is still an open challenge.

Some of the difficulties in food recognition are related to hidden foods on the plates as well as depth and volume estimation. Some projects have used laser beams to overcome this difficulty⁽Reference Stumbo⁴⁰⁾, although the use of extra hardware makes the scalability of these solutions less promising for the immediate future. Similar scalability challenges occur with solutions proposing to use wearables for automatic dietary monitoring⁽Reference Schiboni and Amft⁴⁴⁾.

Natural language processing has also been used⁽Reference Korpusik and Glass^45–Reference Lacson and Long⁴⁷⁾, but still without significant results as a stand-alone input method and without evidence from experiments with end users recording food diaries, for example. This technology could potentially be combined with image recognition in order to increase its accuracy. For instance, a smartphone user could take a food image and record the name of the food items in the meal, perhaps with some estimation of portion sizes. As mentioned in the previous section, new assessment methods have to be validated against independent methods (e.g. weighed food records). No validation study using natural language processing was found in the literature.

Another proposal for addressing the lack of trained professionals for dietary assessment utilised crowdsourcing of untrained workers to estimate energy and macronutrients of photographs⁽Reference Noronha, Hysen and Zhang⁴⁸⁾. This strategy could increase the scalability of nutrition assessment solutions, but probably not enough to support free or low-cost applications to the final users due to the number of manual transactions.

Menu and meal planning

The use of computational methods for supporting nutritional planning has been investigated for some time. The need of meeting nutritional requirements via a vast number of possible combinations of foods opens an opportunity for the use of computers. Mixed integer linear programming was proposed as a possible solution for this in 1993, based on the Simplex algorithm, considering nutrients and prices⁽Reference Sklan and Dariel⁴⁹⁾. Besides meeting the nutritional guidelines, it is imperative to recommend meals that individuals will find acceptable. This challenge was presented by Buisson et al. in 2003⁽Reference Buisson and Garel⁵⁰⁾ and further elaborated upon in a paper presenting Nutri-Educ software in 2008⁽Reference Buisson⁵¹⁾. Based on the nutrition literature, they were aware that a solution would need to propose a well-balanced diet that was not very different from a person's current diet. Improvements would need to be introduced gradually, in order to have good user acceptance. With this in mind, they stated that the current meal is linked to similar meals via addition, deletion or portion modifications. Their solution was to consider this state space as a graph, in which nodes were possible meals and edges between nodes the acceptable transformations. The search strategy consisted of finding an acceptable meal, meeting the nutritional recommendations, as close as possible to the initial meal⁽Reference Buisson⁵¹⁾.

Artificial intelligence approaches, such as case-based reasoning, have also been proposed for planning menus⁽Reference Ravana, Rahman and Chan^52–Reference Petot, Marling and Sterling⁵⁴⁾. Some of these systems were designed to be used by dietitians⁽Reference Ravana, Rahman and Chan^52, Reference Khan and Hoffmann⁵³⁾ who predefined the menus to be selected via CBR based on a number of variables such as age and sex. Khan et al. proposed to use ripple-down rules to create the knowledge base through a direct system interaction while the domain expert is accomplishing their tasks of constructing a diet for a given client⁽Reference Khan and Hoffmann⁵³⁾. However, in consultations between patients/individuals and dietitians/nutritionists the decision-making process for dietary change advice is more complex, with consideration of usual dietary habits in the context of the patient's/individual's preferences are used for optimum advice. Yet, user preference is not necessarily the optimal goal in nutrition recommender systems, as the preferred diets are often unhealthy. That is a limitation of recommender systems which are based on recipes collected from the internet and their reviews⁽Reference Harvey, Ludwig and Elsweiler⁵⁵⁾.

The usual nutritional recommendations provided by nutrition professionals take into account a number of important variables, such as age, sex, dietary restrictions and lifestyle. Some specific groups (e.g. older adults, vegans) need to receive specialised advice. Hence, technology based systems also need to have similar levels of personalisation. An example of such a system is described in the paper ‘Nutrition for Elder Care: a nutritional semantic recommender system for the elderly’, whose recommendations take into account participants’ preferences⁽Reference Espín, Hurtado and Noguera⁵⁶⁾. This paper, published in April 2016, describes the system design and the ontology supporting the recommender system. The system takes into account demographic data (e.g. sex), physical characteristics (e.g. weight and height), nutritional state (e.g. malnourished) and also some responses from an FFQ without portion sizes. The aim of their system is to recommend variations of pre-defined diet plans, instead of modifications to the person's current diet.

These approaches are very useful when there are pre-defined meals and diets (e.g. hospitals), but they can result in proposed diets that are very different from the current diet, making the goal harder to achieve by users. These techniques applied to tailoring daily menus are not directly applicable to creating personalised nutrition advice that is comparable to a nutritional consultation. The next subsection will present dietary modelling tools, which were closer to the final aim of this project.

Dietary modelling tools

For modelling well-balanced diets, the target is generally to balance the food groups (fruit, vegetables, etc.) and distribution of nutrients. A recent study used nonlinear constraint optimisation techniques to design a tool for standardising the background diet of participants during a dietary randomised controlled trial (RCT)⁽Reference Probst, Morrison and Sullivan⁵⁷⁾. A constraint is a function resulting in a Boolean output (i.e. true/false), which is true if the combination of all values is allowed and false otherwise. An objective function results in a set of solutions that are optimal with respect to the objectives, during the constraint optimisation process. The users of this tool (likely dietitians) enter the macronutrients (fat, carbohydrate and protein) and food groups serving targets (e.g. five servings of vegetables) and participant details (height, weight, age and sex) so that the tool can calculate the estimated energy requirement. After this step, using a reference food composition database and also pooled baseline food intake data from completed trials before intervention, the tool solves the optimisation problem and returns to the user their target servings per food group to meet the trial requirements suited to each participant. This solution is possible because the macronutrient energy densities are constant (e.g. fat contains 37·66 kJ/g), such that the tool needs to minimise the difference (i.e. Euclidean distance) between the target and calculated macronutrient values, varying the possible servings per food group⁽Reference Probst, Morrison and Sullivan⁵⁷⁾.

The approaches used by most of the existing systems for checking the outputs against the references are discrete. For instance, if the reference nutrient intake for vitamin D in the UK is 10 μg and an individual is consuming 9·9 μg, a system could indicate, based on a specific rule, that this specific target was not met using a binary evaluation. However, the decision-making process for this type of evaluation is not discrete when completed by domain experts (e.g. dietitians). Human intuition would evaluate this hypothetical scenario as something similar to ‘almost meeting the recommendation’ or ‘very close to the recommendation’, but this rationale is not captured using Boolean logic. This fact motivated a number of researchers to investigate the possibility of doing this analysis using fuzzy logic⁽Reference Buisson^51, Reference Lee, Wang and Lan^58–Reference Wang, Kurozumi and Kawaguchi⁶¹⁾, in which the possible values are not only true or false, but a number between 0 and 1.

Fuzzy logic for nutritional analysis

Lee et al. proposed an ‘Adaptive personalized diet linguistic recommendation mechanism based on type-2 fuzzy sets and genetic fuzzy markup language’⁽Reference Lee, Wang and Lan⁵⁸⁾ in 2015. Taking into account nutritional guidelines (e.g. recommended percentage of energy from fat in a diet) and knowledge from domain experts, it was possible, for instance, to model the fuzziness of percentage of energy from fat. When combining opinions from different domain experts, it was possible to model how those words (low, medium and high) were perceived differently by each expert.

Using the percentage of macronutrients (fat, carbohydrates and protein), the energy ratio and food group balance obtained from the food pyramid, it was possible to compute the diet health level (i.e. very low, low, etc.) from the domain expert and the fuzzy system, as well as the degree to which they matched. For the whole diet sample (n 160) the matching accuracy was 55 % before training, but after learning using genetic algorithms, it increased to about 75 %. This important study only evaluated one variable (diet health level), but this approach could be expanded to more complex systems as long as evaluation from domain experts could be collected also for other aspects of the diet, including nutrients.

One important aspect of the fuzzy logic approach is that real-life nutrition recommendations are more based on words (linguistic variables) than numbers. It seems more sensible to compute with words than numbers in order to construct a mathematical solution for the nutrition decision-making process⁽Reference Mendel, Zadeh and Trillas⁶²⁾. That is the main motivation for fuzzy logic in this context. Regarding evolutionary algorithms (e.g. genetic algorithms), an article published in 2014 confirms the small influence of this field on nutrition recommendation systems and indicates how promising this application could be⁽Reference AlFayoumi, Hegazy and Belal⁶³⁾.

Reproducibility and scalability

The motivations for using the internet for encouraging healthier diets are global. With this in mind, the system deployment should be inexpensive and suitable for many populations, including low income communities. One implication of this requirement is that the proposed solution must not depend on any significant financial investment to be deployed by other institutions or organisations and be built with commercially available technologies that could be replicated in other countries and scenarios. In order to be used in public health campaigns or in similar large-scale initiatives, the solution should be scalable. In other words, it should be able to be used in population interventions with many participants using it simultaneously. Because this digital health solution would collect personal data, information security was mandatory.

User acceptability and usability

As sections of the targeted population may not be familiar with technology (e.g. older adults), the proposed system needed to be easy to use and take into account specific requirements of these individuals, to increase its user acceptance⁽Reference Fisk, Czaja and Rogers⁷²⁾. Another challenge was to develop an app that was designed to run effectively on mobile devices (smartphones and tablets) as they represent the most common devices used to access the Internet currently⁽⁷³⁾.

Validity and effectiveness of the treatment

The aim of the project was to develop a system that was a validated offering from the nutritional science's perspective (i.e. a real improvement in the diet quality) that would encourage positive dietary behaviour change via the internet. The validity of the system could be based on existing reliable publications from the nutrition field (e.g. widely referenced by the scientific community) and should be validated by experts (i.e. nutritionists and dietitians) by comparison with their specific knowledge. The effectiveness of the treatment could then be measured with robust controlled nutrition intervention studies, for example via a RCT.

Scope definition

There are many opportunities for using technologies to improve nutrition assessment and advice. Due to these extensive possibilities, it was necessary to limit the scope of this project, considering its requirements and challenges. From a technological perspective, the nutrition assessment stage (i.e. dietary intake recording) only considered technologies which had been validated in the field. In other words, an acceptable level of accuracy was required in order to provide valid advice. Emerging technologies not yet validated in this field, such as image recognition, were not considered in this project, particularly if they did not meet the basic requirements mentioned in the previous section.

In order to increase the reproducibility of the system, it was limited to popular technologies and programming languages, so that it could be improved by other researchers and developers who would potentially have access to its open source code. Furthermore, bespoke devices or specific hardware were not considered as alternatives, and the use of public cloud services was a key aim of the project. eNutri is independent of software requirement or plugin installation.

Nutrition advice via the internet is still at an early stage. For that reason, it was considered important to limit the target population of the EatWellUK study, so that people with specific dietary requirements (e.g. those suffering from coeliac disease or diabetes) did not receive advice that may have conflicted with the recommendations they might be receiving from health professionals or even cause some undesired consequences (e.g. recommending inappropriate foods for individuals with serious food allergies). Since this project also evaluated the effectiveness of the nutrition advice, it did not accept participants receiving face-to-face nutritional consultations, to avoid conflicting messages and potential bias. The more these systems evolve, the wider these eligibility criteria can become.

Nutrition assessment method

This project started without a pre-defined nutrition assessment method to be used during the dietary intake recording. We were aware of the use of commercial food diaries in the app stores, but the complete reliance on this method in the most popular apps was not expected. This might be an indication that due to a small number of usability and intervention studies in this field⁽Reference Teasdale, Elhussein and Butcher⁶⁴⁾, the developers have possibly drawn inspiration for their solutions from existing apps, thereby repeating the same assessment method (food diary).

The creation of a database with product barcode information requires time and effort. It is likely that most of the apps have this feature due to the fact that there are barcode databases openly available, such as the Open Food Facts⁽⁷⁴⁾. At the time of writing, this database contains 609 621 items registered via a collaborative effort. However, since nutrition content contained in the product labels only presents the main or legally required nutrients, the food composition database of these apps may not contain reliable information for micronutrients.

Although it would be possible to create an online service for personalised nutrition advice using food diaries, this alternative could be risky due to high participant burden and associated risk of dropout. In order to provide nutrition advice, the decision engine would need to consider food diaries for multiple days (baseline), in order to detect a dietary pattern and perform the calculations, and then repeat this process after a designated period in order to measure the treatment effect. This approach could be perceived as too burdensome by the participants. The Food4Me project, which developed and validated an online FFQ for their large population-based intervention study on personalised nutrition, took the decision to use an FFQ as their nutrition assessment. In our current project, we took the same decision, and were also able to benefit from permission to use the Food4Me food list and images.

The eNutri nutrition assessment and advice relied on existing and validated methods in the nutrition field. The assessment method (Food4Me FFQ) had been previously validated in the UK (i.e. it is a representation of people's dietary intake with a certain level of accuracy)⁽Reference Fallaize, Forster and Macready⁷⁵⁾. In a similar way, the tool for measuring the quality of the diet (i.e. AHEI) is accepted by the scientific community⁽Reference Chiuve, Fung and Rimm¹¹⁾, making the measurement of treatment effect reliable.

Design and usability

Most of the usability challenges are more complex in smartphones, as compared with desktops and laptops, due to the smaller screen size. As this proposed app does not use any specific internal hardware (e.g. GPS, accelerometer or camera), it was not necessary to develop a native app (i.e. specific to iOS or Android and requiring the app to be installed) to meet the system requirements and collect usability data. However, the eNutri app can be easily converted into a native app because it takes advantage of responsive JavaScript frameworks and there are commercial tools providing this conversion for running web apps as native apps for the major mobile operating systems⁽⁷⁶⁾.

To present the food items as a list on smartphones was very challenging. Although possible, it placed constraints on the font size and required users to scroll the page repetitively. Originally, the whole of the FFQ was presented as a list that appeared on a single screen⁽Reference Franco, Alawadhi and Fallaize⁷⁷⁾. After a usability evaluation of this version, a newer version of the app was developed presenting each food item individually on the screen⁽Reference Zenun Franco, Fallaize and Lovegrove³⁾.

To increase the acceptability by older adults, some design principles were included in the app such as: reduction of the user choices on the screen, appropriate font size, reduction of hidden items, having a help button always visible, individual help links for some items and user journey serialisation⁽Reference Zenun Franco, Fallaize and Lovegrove³⁾.

Both versions⁽Reference Zenun Franco, Fallaize and Lovegrove^3, Reference Franco, Alawadhi and Fallaize⁷⁷⁾ resulted in similar completion times and user acceptance but there is evidence from the literature that a serialised design is more appropriate from a usability perspective, especially if the solution is aimed at including older adults⁽Reference Fisk, Czaja and Rogers⁷²⁾. Based on these results, the serialised FFQ design would appear to be the more appropriate one of these two options.

Reproducibility and scalability

The only back-end service used by eNutri was Google Firebase, which is offered by one of the most popular public cloud providers currently and available for other customers worldwide. Because it is a single page app, the content (i.e. texts, images, etc.) is initially downloaded from the hosting provider (i.e. Google Firebase) and processing occurs mainly in the browsers, making this solution scalable⁽Reference Franco, Alawadhi and Fallaize⁷⁷⁾. It also facilitates the reproducibility, because it does not require any server-side installation or maintenance. The app can be deployed using the Google Firebase free plan⁽⁷⁸⁾, meeting the low-cost requirement. Regarding security, Google Firebase is compliant with the main global standards and security certifications, including the recent 2018 General Data Protection Regulation from the EU^{(79, 80)}. From a software perspective, the security rules, detailed in a previous publication⁽Reference Franco, Alawadhi and Fallaize⁷⁷⁾, met the authentication and authorisation requirements needed for processing and storing personal data.

Decision engine

Within this project, a modified version of the AHEI (m-AHEI) was selected as the foundation of the decision engine for tailoring the personalised advice (unpublished results). Indexes of overall quality of diets have previously been used for diet assessment, but not for generating online advice. This novel app of a diet quality score (m-AHEI) was associated with some risk, as there was no prior literature on the suitability of this strategy for personalised advice that effectively changed dietary behaviour. This was tested in the EatWellUK RCT (unpublished).

Future work

The version of the decision engine used in the EatWellUK RCT derived the advice based on dietary intake (FFQ data) and sex information. In alignment with behaviour change theory⁽Reference Michie, Richardson and Johnston⁸¹⁾, the user should take more control of their diet changes. A possible solution is to evaluate the quality of a diet change using the AHEI itself, in order to identify multiple foods that can improve the AHEI score and to let the user choose from the multiple modifications. If the app also allows the users to rate the proposed modifications, that could act as further input to improve the recommendations. This type of approach seems to lend itself well to the use of node graphs, where the nodes represent the dietary intakes and edges the possible changes⁽Reference Buisson⁵¹⁾. Different from basic meal planning, it is important to have different weights (i.e. change acceptability) for the edges, similar to map routing with different traffic conditions.

It is particularly important that future versions of this app are able to consider participants’ actual willingness to change their behaviour. For instance, one could want to improve nutrition and yet be unwilling to increase vegetable consumption, so the cost of the path for increasing the vegetable component would need to be recognised as higher than an alternative method for increasing the diet quality index, such as including more fruit for example. In other words, the system would need to take into account participants’ interactions and propose different paths for increasing their diet quality.

Another key challenge for future work is how to train the decision engine. In a typical recommender system, for example a medical diagnosis system, there is available data on the inputs (e.g. symptoms) and corresponding outputs (i.e. diagnosis/treatment). One of the main differences with a nutrition recommender system is that there is no existing openly available dataset with people's case histories and the corresponding proposed diet modifications from nutrition professionals that could be used for training. Furthermore, differently from many common recommender systems (e.g. online shopping based on previous purchases), the food items that users like and consume the most are not necessarily the healthiest. In other words, the recommender system must consider factors other than users’ preferences and prior consumption.

Based on a recent systematic review⁽Reference Teasdale, Elhussein and Butcher⁶⁴⁾, an analysis of the systems used to remotely deliver nutrition interventions confirmed that there is no other publicly available decision engine that provides valid online personalised nutrition advice automatically. This publication also showed that all of these interventions were conducted in high-income countries, reinforcing the need for a reproducible and inexpensive solution for online personalised nutrition advice.

Although the latest version of the eNutri app could be useful and potentially contribute to healthier dietary habits, its decision engine is elementary computationally and deterministic (i.e. if two individuals have the same dietary intake information, they will receive the same advice, even if they have different preferences). In addition, its decision engine does not take into account important data sources, such as population data, historical data or individual's preferences. In other words, the current feedback is driven purely by the diet data and sex, and the question of how best to balance nutritional healthiness against users’ preferences in a way that maximises the likelihood of the user adopting healthier choices remains an open research challenge. The eNutri design and source code have been made publicly available⁽Reference Franco⁸²⁾ under a permissive open source license, so that other researchers and organisations can benefit from this work.