The dietary feedback system was developed, for manual use by researchers, for the delivery of personalized dietary advice in the pan-European Food4Me Proof-of-Principle (PoP) study. The Food4Me PoP study that aimed to deliver Web-based personalized dietary and physical activity (PA) advice was designed to emulate a real-life Internet-delivered, PN service [25]. As described in detail elsewhere, 1607 participants were randomized to 1 of the 4 groups receiving different levels of PN advice: Level 0 (control group) receiving conventional, non-PN advice; Level 1 PN advice based on dietary intake and PA data alone; Level 2 PN advice based on dietary intake, PA, and phenotypic data; Level 3 PN advice based on dietary intake, PA, phenotypic, and genotypic data [25]. The aim was to recruit a total of 1540 participants (220 participants per center) to allow for a potential 20% dropout rate, planned using a priori power calculation [25]. Power calculations were conducted in Minitab and based on glucose and omega-3 fatty acid concentrations within European adult populations [25]. The study had ethical approval from the corresponding committees of all participating centers (Germany, Greece, Ireland, the Netherlands, Poland, Spain, and the United Kingdom) and was conducted from August 2012 to March 2014. The Food4Me trial was registered as a randomized control trial (NCT01530139) at Clinicaltrials.gov.

The dietary feedback system, described in this paper, was used to deliver PN advice to participants randomized to receive Level 1 PN only (n=414). All participants of the Food4Me study received dietary feedback advice via email without face-to-face contact with researchers [25]. Thus, the feedback system was designed and developed to ensure that delivery of personalized dietary feedback was consistent across all the 7 countries. The key stages in the development of the dietary feedback system are illustrated in Figure 1 and described in detail in the following section.

Nutrient intake analysis was automated for use in the Food4Me study, with the nutrient-related goals and feedback messages derived manually (using the priority system and decision trees) by Food4Me researchers using their own judgment to overrule the feedback system when appropriate. After completion of the Food4Me intervention study, the feedback system was automated by Creme Global (Dublin, Ireland). Therefore, a comparison of the automated and manual-based approaches was possible. The system was automated by firstly capturing the manual decision process for selecting priority nutrients, based on nutrient intake analysis, in a computer algorithm (written in the PHP programming language). Second, the decision trees for selection of feedback messages were encoded in table-like data structures and stored in a relational database (MySQL). Additional computer algorithms for traversing these data structures were then developed (also using PHP), which enabled feedback messages to be generated automatically from nutrient intake analysis.

The first part of the analysis examines the manual use of the feedback system during the Food4Me PoP study. Descriptive statistics were computed to describe the general characteristics of participants across each of the 7 centers using general linear model analysis with least significant difference post hoc. Nutrient-related goal selection frequency was examined across each of the 7 countries. All statistical analyses were conducted using IBM SPSS Statistics version 20, and P<.05 was considered statistically significant.

The second part of the analysis compares the dietary feedback provided by researchers during the Food4Me PoP study with feedback generated automatically by the computerized algorithms and messaging system for participants in Level 1 of the intervention. The level of agreement between the manual (researcher) and automated systems was assessed for both nutrient-related goal selection and feedback advice using baseline data for Level 1 participants, as illustrated in Figure 3. Agreement between the manual and automated systems, in selecting the 3 nutrient-related goals, was investigated by comparing each individual nutrient-related goal (1, 2, and 3) selected by both systems (manual vs automated) for each participant in the 7 countries. To evaluate the agreement for nutrient-related goal selection, we also examined whether the same 3 nutrients were selected as the 3 nutrient-related goals (in any order 1, 2, or 3) by both systems.

As outlined in Figure 3, to compare if the same feedback advice message(s) were derived from each decision tree by both systems, we selected only participants given the same nutrient-related goals (1, 2, and 3) by both the manual and automated systems and cross-compared the messages given by both systems for each nutrient-related goal (1, 2, and 3). Agreement was categorized as either “complete agreement,” “complete disagreement,” or “partial agreement” (data not shown). Overall agreement for feedback advice message was computed by summing the number of participants categorized into the 3 agreement groups for each nutrient-related goal across the countries.

A total of 414 participants, from the Level 1 group, across the 7 countries were available for inclusion in the analysis. After removing dropouts immediately after being randomized (n=41) and participants who did not have nutrient-related goals recorded at baseline (n=4), 369 participants were included in the analysis (Germany, n=52; Greece, n=51; Ireland, n=55; Netherlands, n=56; Poland, n=49; Spain, n=54; and the United Kingdom, n=52). There were no significant differences in age for men and women; however, self-reported BMI was significantly lower for women compared with men (P=.002). When comparing demographic characteristics across all the countries, no significant differences were observed for bodyweight or BMI; however, significant differences were observed for age (P=.02) and height (P<.001), as summarized in Table 1.

The top 3 most frequently selected nutrient-related goals (1, 2, and 3) across each of the countries are summarized in Table 2. Saturated fatty acid was most frequently selected as nutrient-related goal 1 across all centers except for Spain where PUFA was most frequently selected as nutrient-related goal 1 in feedback given to 35% (19/54) of participants. Salt was most frequently selected as nutrient-related goal 2 in all 7 centers, with the exception of Germany, where both folate and salt were most frequently selected. As summarized in Table 2, greater variation was observed across the 7 countries for nutrient-related goal 3. As a result of the prioritization process, salt was most frequently selected as both nutrient-related goals 2 and 3, in both Spain and the United Kingdom.

The number of times each of the 17 nutrients was selected as a nutrient-related goal (1, 2, or 3) across each of the 7 countries is summarized in Table 3. Overall, SFA, salt, and dietary fiber were the top 3 most frequently targeted nutrient-related goals provided to 72%, 72%, and 39% of participants, respectively. Saturated fatty acid was the most frequently chosen nutrient-related goal overall for Germany, Ireland, and Poland. Salt was the most frequently selected nutrient-related goal overall for the Netherlands, Spain, and the United Kingdom, and in Greece, both dietary fiber and salt were most frequently selected.

A summary of the key issues encountered during the Food4Me study regarding the use and applicability of the decision trees and feedback advice messages is summarized in Table 4. Most of these issues were related to the SFA and salt decision trees, owing to their more complex design.

The level of agreement between the manual and automated systems for the selection of the 3 nutrient-related goals in all the countries is summarized in Table 5. Good agreement was observed between both methods in all the countries for nutrient-related goals 1 and 2 with average agreement of 92% and 87%, respectively. Agreement between the 2 methods ranged from 100% (the United Kingdom) to 82% (Greece) for nutrient-related goal 1 and from 98% (Spain) to 80% (the Netherlands) for nutrient-related goal 2. Lower agreement was observed between the manual and automated systems for nutrient-related goal 3, mean 63% across centers. For nutrient-related goal 3 selection, agreement was highest for Spain (85%) and lowest for Greece (45%). Agreement between the 2 systems for all 3 nutrient-related goals in random order ranged from 83% (Spain) to 47% (Greece), averaging 66% across all countries.

Having selected the priority nutrients, the next stage was the selection of feedback messages. The level of agreement between the manual and automated systems for the feedback advice messages is summarized in Table 6. Complete agreement between the 2 systems for feedback advice ranged from 90% (Greece and Ireland) to 82% (Germany and Poland), with an average of 87% across the 7 countries.

This paper presents a novel system for providing consistent personalized dietary advice, automatically generated from dietary intake data submitted via the Internet by participants in a multi-country study. To our knowledge, this paper is one of the first to describe, in detail, the steps involved in developing a dietary feedback system to translate food and nutrient intake data into automatically generated personalized feedback advice and to compare the automatically generated advice with advice manually provided by nutrition researchers.

Development of the feedback system consisted of a 4-step process: designing a gradation system to categorize nutrient intakes, creating a priority system to enable 3 nutrients to be further selected as nutrient-related goals for particular emphasis in the feedback report, constructing decision trees to link nutrient intakes to feedback advice, and finally, developing feedback report templates. The findings of this paper demonstrate that automation of the feedback system is feasible and more superior compared with manual use of the system by result of removal of human error.

The feedback system described in this paper is unique as it was developed to provide personalized feedback on intakes of 5 food groups and 17 nutrients, 3 of which were prioritized and selected as “nutrient-related goals” for subsequent targeted dietary advice. As the feedback system was developed to deliver relevant PN advice for all European adults, it was important that the system considered all nutrients for which changes in intake would be recommended to improve the health of significant proportions of the adult population. To address this requirement, the system that we developed included key food groups and a large number of nutrients so that feedback could be generated and personalized for a complete nutritional profile, rather than limited personalized information on one or several nutrient(s) as previously applied in other tailoring studies [20,37,38]. Providing extensive information, guaranteed that the dietary feedback would include personalized advice, which could be translated into food groups and nutrients of importance for a healthy diet (including fruit and vegetables, SFA, salt, fiber, whole grains, red meat and dairy products, omega-3 fatty acids, vitamin C, and iron) [39]. Several studies have briefly alluded to the application of a similar multistep process, linking nutrient intake data to feedback messages [1,13,21,24,37,40,41]. However, most of these studies have focused mostly on tailoring advice for selected nutrients or food groups only (eg, SFA, fruit and vegetable intakes) [1,6,13,16,21,24,37,42-44], with the exception of the advice generated for adolescents by Maes et al [40] which included fiber, vitamin C, calcium, iron, and fat and the reports developed by Kannan et al [23] that provided personalized feedback for 13 nutrients.

In addition to providing personalized food-based messages, feedback was also displayed graphically in reports by comparison of the individual’s nutrient intakes with dietary recommendations, similar to many other studies [1,6,15,16,42,45]. Nutrient-related goal selection was relatively consistent across the countries, with SFA and salt being the most frequently assigned nutrient-related goals targeted, although some North–South gradient differences are observed. This result was expected given the prevalence of diets high in SFAs and salt across Europe [39,46,47].

Because the feedback system had not been tested previously in a multicenter cohort, the decision trees were used manually for the delivery of dietary advice within the intervention study. Manual use of the decision trees within the study combined with rigorous recording of the issues identified by the nutritionists or dietitians delivering the intervention facilitated evaluation of the advice and identification of aspects of the system, which could be improved. The decision trees were developed to comprise essential components such as contribution from supplements and food groups (SFA and salt decision trees only). To guarantee feedback advice is fully relevant and appropriately personalized at an individual level, suggested improvements include further expanding the decision trees to integrate specific food items and incorporate additional food groups (SFA and salt decision trees), for example, branches for consumption of commonly consumed food items high in SFAs (eg, ice cream and quiche) could be added to the SFA and total fat decision trees.

Automated technologies are increasingly used across many disciplines and have been shown to be as effective as human (face-to-face) systems for 3-month weight loss and exercise interventions [8,48]. Furthermore, Emerencia et al [49] recently developed an automated personalized system for schizophrenia patients providing advice similar to that given by clinicians. Although several studies have generated computer-automated dietary feedback, to our knowledge, this paper is among the first to examine whether it is possible to fully automate a dietary feedback system designed to provide PN advice to enhance healthy eating. We observed high agreement between the manual and automated systems, averaging at 92% and 87%, for nutrient-related goals 1 and 2, respectively. Good agreement was also observed for feedback advice selection averaging 87%. However, although our results show potential for future automation of dietary feedback tools, some disagreement was observed between the manual and automated systems.

Nutrient goal selection disagreement averaged 8%, 13%, and 37% across the 7 centers for nutrient-related goals 1, 2, and 3, respectively. Reasons for this disagreement included: researchers overruling the priority system (eg, if they thought it was inappropriate to give one of the nutrient-related goals selected using the priority system); researchers misreporting the nutrient-related goal in the report (eg, the correct nutrient-related goal was given in the report but misreported in the Internet); and researcher error in selecting the nutrient-related goal (eg, giving an adequate nutrient by mistake). The level of disagreement between the 2 systems was much greater for nutrient-related goal 3, especially for Greece, primarily, a result of overruling the priority system. Overruling did not always mean that the original selection was inappropriate, only that the researcher thought another option was more relevant. For feedback advice selection, complete disagreement was negligible (2%), but partial agreement was higher (12%), and was confined to the SFA and salt decision tree messages only. Owing to their more complex design, the feedback advice for these nutrients included 2-4 individual messages, which were more prone to researcher misreporting, error or overruling the decision trees to give additional messages. There was also a degree of ambiguity in the prioritization rules and the SFA and salt decision tree rules. For example, in relation to prioritization, no rule was set for selecting a positive third nutrient when only 2 risk nutrients were identified, thereby increasing the level of disagreement for nutrient goal selection between the manual and automated systems. Similarly, for both the SFA and salt decision trees, no rules were established for which food groups should be selected when several contributed equally to the nutrient intake.

Although we observed some disagreement between the manual and automated systems, overall, the agreement between the systems was excellent. Much of the disagreement we observed between the 2 systems would be reduced by automation of the feedback system, which would eliminate researcher error and researcher misreporting, and further clarification of the decision rules for both the prioritization process and SFA and salt decision trees. Furthermore, despite researchers overruling the priority nutrients, both the systems selected nutrients of high priority.

Technological innovations have evolved the delivery and management of health care through the increasing use of mobile phone health and computerized health care apps [50,51], with the personalized health field developing rapidly [52]. Consequently, our feedback system has potential to become a cost-effective approach to improve dietary behaviors at an individual level, which could be delivered by public or private health care providers. The present system was designed to provide advice on usual intake rather than actual intake; the challenge in the future will be to provide advice to consumers at the point of purchase or consumption.

The strengths of this study include the adequate sample size and multicenter recruitment in 7 European countries. In addition, to facilitate behavior change, the system was designed to incorporate numerous behavioral change techniques including goal setting, action planning, and barrier identification [11,25]. However, the system did not take into account motivation for change, and this could be a potential limitation. Feedback that is personalized to current dietary intake and stage of change may be considered as being more motivational and could have greater efficacy in promoting sustained dietary changes than advice personalized on dietary intake data only [14]. Another potential limitation of the study is that the feedback system was tested in a group of individuals who volunteered to receive PN advice and may therefore not be applicable to other population groups. However, the dietary and anthropometric characteristics of participants in the Food4Me study were broadly similar to those of the wider population of European adults so that the study tested the utility of the system to select appropriate feedback for the likely most common dietary changes. In addition, we did not follow-up with participants regarding their opinions about the advice they received, for example, if they thought it was relevant to them, appropriate, or useful, and this would have been extremely useful information to have captured for the future progression of PN.

We developed a Web-based dietary feedback system that was capable of delivering consistent personalized dietary advice to adult European participants in the multicenter Food4Me study. Outcomes from comparison of the manual and automated feedback systems provide confidence that such an automated dietary feedback system can be developed and implemented across multiple countries with the potential to contribute to scalable and cost-effective interventions to improving dietary behaviors and health across large populations.