Introduction
Diet plays significant direct and indirect roles in the etiology and prevention of chronic diseases including type 2 diabetes, coronary heart disease, cancer, and obesity [1]. It has been estimated that an unhealthy diet was the leading cause of premature death in the United States, contributing to more than 500,000 deaths in 2016 [1]. Despite these statistics, we lack a clear understanding of how patterns of dietary intake affect health through the life span because of dietary measurement limitations. Current dietary assessment methods including 24-hour dietary recalls, food frequency questionnaires, and dietary records have advantages and disadvantages in research settings [2]. The 24-hour dietary recalls are considered the gold standard in dietary assessment because they produce the highest quality data, but they rely heavily on participant memory and can require the greatest amount of time to acquire when administered by the interviewer. Food frequency questionnaires are more easily administered and provide good estimates of patterns of intake; however, they are also subject to participant memory and are most likely to underestimate energy intake because of limitations including a lack of cultural tailoring or limited food lists. Diet records or diet diaries minimize reliance on memory when foods and beverages are recorded when consumed; however, participants require a high level of motivation as well as training to improve recording accuracy. Furthermore, research shows that people pay little attention to when and what they eat [3], and factors such as age, sex, and weight status can influence the accurate recall of food and estimation of portion size [2,4-8]. Finally, methods that do not use unannounced recalls (eg, diet records) may be prone to biases such as social desirability or reactivity bias, which may lead to participants underreporting or omitting of foods or beverages consumed or changing their usual dietary behaviors because of the awareness that they are being observed [9]. Such errors in reporting are known to create conflicting evidence linking diet to health outcomes that could be addressed if we could more robustly measure diet [10,11]. Furthering our understanding of the connection between diet and disease will require improvements in the dietary assessment methodologies. For this reason, the research community has recognized the need for new dietary assessment methods that can reduce misreporting and recall biases [12,13].
Recent advancements in digital technology and computational sciences have laid the foundation for emerging dietary assessment solutions. These advancements have catalyzed the development of new methods aimed at automating the assessment of dietary intake, thereby limiting or eliminating the need for self-report. In particular, 2 such new dietary assessment methods have been developed: image-based dietary assessment [14-17] and the detection of food intake by biomechanical sensors [18-29]. Single image-based assessment methods use photos of foods and beverages along with fiduciary markers before and after consumption. The time-stamped images are either reviewed and coded into nutrition software by trained research staff (image-assisted assessment) or are analyzed by software designed to identify the type and volume of foods in the image (image-based assessment). Alternatively, approaches using gyroscopes, microphones, and mechanical or electrical impedance sensors have been integrated into wearable devices such as watches and headsets or are designed to be mounted on teeth to detect wrist or hand motion or patterns of chewing or swallowing indicative of food intake (eg, number of bites). However, the automation of dietary assessment using these mobile-based approaches remains in the proof-of-concept stage. There is a lack of large-scale validation studies demonstrating their utility to assess dietary intake in community-dwelling populations. For instance, the mean detection accuracy of image detection and wearable devices has been acceptable in controlled, laboratory settings (range 73%-99%) [19,21,23,24,26-35], but limited testing has been done in natural settings. The use of mechanical sensors in research is further hindered by poor battery life, having to remember to wear or use the device, needing to turn the device on or off to avoid detection errors, and the conspicuousness or general discomfort of having to wear collars, wires, or harness accessories. Substantial work will be needed before these methods can accurately quantify energy or nutrient intake for research purposes. Therefore, novel dietary assessment methods addressing limitations of the traditional dietary assessment methods and methods that bridge the gap between traditional and newer methods of dietary assessment are needed [12,13].
Another less developed dietary assessment method with the potential to improve the validity and reliability of dietary assessment is the mobile device–assisted ecological momentary assessment (mEMA). mEMA is based on the foundation of ecological momentary assessment (EMA) described by Shiffman et al in 2008 [36]. EMA involves the repeated sampling of a person’s current behaviors and experiences in real time, in their natural environments. Currently, there are 2 mobile device–assisted ecological momentary diet assessment (mEMDA) approaches: event-contingent mEMDA and signal-contingent mEMDA. Event-contingent mEMDA most often occurs in real time at the time of eating (or drinking). The frequency of sampling is determined by the number of times a participant reports eating. Here, the act of initiating a meal or snack triggers either the real-time recording of dietary intake (eg, dietary records or image-assisted dietary records). The advantage of real-time diet records is that they are intended to capture all foods and beverages consumed without having to recall the events at a later time. Although this is an advantage, the key limitation of event-contingent mEMDA is that there are no unannounced sampling events. The self-monitoring of dietary intake can be influenced by psychosocial and behavioral factors that introduce reactivity and measurement bias including eating behaviors (dietary restraint or disinhibition), social desirability, body image, or depression and anxiety [5]. Furthermore, with image-assisted diet records, there is the potential for data entry bias by research staff viewing images or for intentional or unintentional reporting errors (eg, inaccurate report of portion sizes) by participants, particularly if foods are omitted from the images or the images are not taken at multiple points of a meal (eg, before and after a meal).
Signal-contingent mEMDA relies on researcher-initiated, signaled prompts to participants that trigger the recall of current or recent dietary intake. Although study participants are often prompted multiple times per day, signal-contingent mEMDA does not always allow for the real-time assessment of dietary intake. Rather, assessment surveys often include questions referencing dietary intake occurring within the most recent interval of time (eg, past 30 min). Moreover, study participants are most often asked to report the consumption of specific foods or foods from specified food groups (eg, fruits and vegetables) by means of a brief survey. The frequency of sampling using signal-contingent mEMDA is determined by the researcher and can occur randomly at fixed or semifixed times or randomly within fixed or semifixed time intervals. As with image-assisted or image-based dietary records, these momentary dietary assessments can be time-stamped and are either stored or transmitted for later database integration. Although this method benefits from random (unannounced) sampling, short recall intervals, and reduced participant burden, the commonly used sampling schemes, limited study durations, or limited food lists can hinder the quantification of energy or nutrient intakes.
Due to recent advancements made in mobile device hardware and software and the pervasive use of mobile devices, EMA-based methods leverage the capabilities of mobile technology, offering researchers an opportunity to assess the dietary intake of study participants as they are occurring in natural settings. Both event-contingent and signal-contingent mEMDA seek to reduce recall bias and thereby improve the accuracy of dietary assessment by eliminating or shortening the recall interval and reducing participant burden while maximizing ecological validity as compared with the traditional dietary approaches. However, neither approach has been well characterized nor adequately compared against objective biomarkers or other methods of energy intake assessment (eg, 24-hour dietary recalls). Focused efforts are needed to develop mEMDA methods for their consistent and replicable application in research settings. Therefore, the goals of this systematic review were to summarize the event-contingent and signal-contingent mEMDA protocols that have been used in research to date, to characterize key aspects of these assessment approaches (eg, design, data collection, data processing and dietary analysis, and dietary outcomes), and to discuss the advantages and disadvantages of each as well as implications for future mEMDA research. The focus of the review was on studies summarizing unique dietary assessment protocols using mobile devices to facilitate event-contingent or signal-contingent EMA to assess dietary intake.
Methods
Literature Search
A systematic strategy was devised by 6 authors (SMS, YL, MDH, JH, GFD, CAT, and CJB) to search the MEDLINE, EMBASE, PubMed, PsycINFO, and IEEE explore databases for all relevant literature published through February 2018. The search was limited to articles written in the English language and conducted with humans. The database search included the use of controlled vocabulary and keywords to identify studies addressing dietary assessment, mobile devices, and ecological momentary assessment. Keywords such as “nutrition assessment,” “diet surveys,” “diet records,” “energy intake,” “meals,” or “eating” combined with “text messaging,” “mobile phone,” “mobile applications,” “micro-electrical-mechanical systems,” or “wearable electronic devices” were included as MeSH search terms. In addition, non-MeSH search terms were included to be complete: “caloric intake,” “food diary,” “diet monitoring,” “food tracking,” “diet tracking,” “diet assessment,” or “calorie tracking” and “text messages,” “cell phone,” “smartphone,” “tablet computer,” “mobile health,” “eHealth,” “mHealth,” “digital health,” “mobile technology,” or “experience sampling.” Search terms synonymous with the terms above that did not produce additional references (eg, “food record”) were omitted from the final search conducted by the author (SMS). References cited in all included studies and studies citing included studies (referred to as “other sources” in the Preferred Reporting of Systematic Reviews and Meta-Analyses [PRISMA] diagram) were also reviewed to identify any additional studies to be screened for inclusion.
Study Inclusion and Exclusion Criteria
Eligible studies had protocols using mobile devices and event-contingent or signal-contingent EMA approaches to assess dietary intake in research settings. These included diet assessment studies as well as behavioral trials where dietary intake was assessed. Dietary intake was defined as the quantification of energy intake, macro- or micronutrients, discrete foods, servings, or food groups. Literature returned by the search was screened first by article type then by title, abstract, and the described methods by 2 authors (SMS and YL). Only original research articles were included. Abstracts, review papers, editorials, etc, were excluded. Additionally, studies were excluded if the title, abstract, or methods indicated (1) the study was not diet-related (eg, nondiet-related papers, proof-of-concept, or technology design papers); (2) the studies were interventions with non-EMA dietary assessment methods (24-hour dietary recalls and food frequency questionnaires); (3) did not assess dietary intake (eg, binge eating lapses, availability of snack foods, and food craving); (4) used self-monitoring approaches without dietary analysis; (5) were described in an earlier study or were considered a secondary analysis; or (6) were not peer-reviewed journal articles (eg, abstracts, editorials, discussions, evaluations, reviews, reports, news, notes, surveys, or content analysis). Additional papers referencing the included studies were used to obtain methodological details not otherwise provided in the included studies.
Data Extraction and Analysis
Data were extracted into a structured coding form according to PRISMA guidelines [37] and the Cochrane Handbook for Systematic Reviews of Interventions [38]. A data extraction form developed for this review was used by 2 authors (SES and KGH) to independently extract and review characteristics from all studies. Extracted data represented details on mEMDA protocols and included but were not limited to (1) mobile platform, (2) sampling duration, (3) prompt approach (signal-contingent only), (4) prompt frequency (signal-contingent only), (5) data collection method, (6) data processing and nutrient analysis, (7) diet data outcomes, and (8) protocol adherence. A comprehensive list of data extracted from included studies is provided as Multimedia Appendix 1. Discrepancies in the extracted data were resolved by a discussion between 2 expert reviewers (SMS and YL) to complete the dataset. In several cases, studies closely related to the included studies were reviewed for additional information to resolve issues of missing or unclear data. Extracted data were descriptive in nature. The data were synthesized narratively and tabularized with the intent of summarizing available protocols for assessing diet using mobile EMA methods.
Discussion
Summary of Key Findings
This systematic review summarized the existing protocols for measuring dietary intake using 2 mEMDA approaches: event-contingent and signal-contingent mEMDA. Studies describing 20 unique mEMDA protocols were included in the review. Half of the studies used event-contingent mEMDA protocols and half used signal-contingent mEMDA protocols. Most studies used mobile phones to collect dietary data. Studies that used event-contingent mEMDA most commonly assessed diet in real time, used dietary records to collect data, and provided estimates of energy and nutrient intake for data collection purposes. All signal-contingent mEMDA studies used near real–time recalls and unannounced abbreviated diet surveys and assessed the frequency of consumption of foods or food groups most relevant to the research. Only 6 (30%) mEMDA studies directly compared mEMDA and dietary outcomes measured by the traditional dietary assessment methods (eg, 24-hour dietary recalls). As such, the evolving body of literature identified in this review supports the application of mEMDA as the next step for advancement in the field of dietary assessment, bridging the gap between traditional methods and newer, more technologically advanced methods (ie, biomechanical sensing and image-based food detection), which are currently under development.
Key Strengths and Limitations of Mobile Device–Assisted Ecological Momentary Diet Assessment
The mEMDA approaches described in this review have strengths and limitations potentially impacting the quality of estimated dietary intake. Event-contingent mEMDA protocols have several strengths based primarily on the fact that food (and beverage) consumption generally occurs as a discrete event; it serves as a cue to record intake [77]. These studies most commonly assessed diet in real time (eg, at the time of consumption) or used image-assisted or image-based dietary assessment methods to reduce recall bias (eg, dependence on memory) when intake was reported at a later time. Moreover, when integrated with food composition databases or similar software, event-contingent mEMDA protocols most often facilitate the calculation of or automatically calculated estimates of energy and nutrient intake. Finally, the mobile platform of event-contingent mEMDA enables eating events to be time-stamped. This removes the time-keeping burden from participants, allows for better specificity for eating occasions, and enhances the ability to examine the distribution and frequency of eating events across days or weeks [77]. There are also several limitations inherent to event-contingent mEMDA that are equally problematic as the similar traditional methods of dietary assessment (eg, written food records). First, event-contingent mEMDA requires participants to initiate the self-reporting of each eating event over the study period, which may be perceived as burdensome and may lead to omitted data when participants forget or decline to report [36]. Second, the routine anticipation of self-report of dietary intake (vs unannounced recalls) consistent with event-contingent mEMDA is more likely to be biased by psychological reactance or social desirability, whereby people change their usual eating behaviors or intentionally misreport intake so as to not be judged for making diet-related decisions perceived by the individual to be less healthful [5]. Finally, compared with traditional dietary assessment methods, event-contingent mEMDA may be limited by existing on a mobile platform, which might be less acceptable to some populations including people who are less comfortable with technology (eg, older adults) [78] or have low electronic health literacy [79]. For these individuals, written records might be preferred. In addition, compared with signal-contingent mEMDA protocols, which typically employ an abbreviated survey, event-contingent mEMDA more often involves more high-resource data processing to process raw video, photos, or sound files into analyzable food group or nutrient data at meal level or day level requiring software that might not be readily accessible to researchers. Overall, event-contingent mEMDA appears useful for capturing individuals’ intake as it occurs, by eliciting a time-stamped log of all eating events and their contents; however, limitations include greater participant burden (ie, recording all food and beverage consumed and remembering to do so), the increased likelihood of psychological reactance or social desirability biases, and lower acceptance levels in some populations.
The remaining studies (n=10) used signal-contingent mEMDA sampling. All the signal-contingent mEMDA studies used a near real–time recall approach with abbreviated diet surveys, whereas some also incorporated real-time prompting. Most signal-contingent studies assessed the frequency with which foods or food groups most relevant to the research were consumed. Strengths of signal-contingent mEMDA include lower participant burden related to survey brevity and unannounced sampling, which provides a random sampling of eating events throughout the day. Here, participants receive prompts to report their recent intake throughout the day, providing a representative sample of overall daily intake without having to proactively input details about each eating occasion as they occur. Additionally, the majority of existing signal-contingent studies used simplified reporting methods, asking participants whether or not they consumed certain target foods within a recent interval of time (eg, past 30 min or past 2 hours). Though signal-contingent mEMDA has several strengths, there are also limitations to note. First, it may be subject to incomplete data, particularly when sampling windows do not cover the entire day. As a result, some eating events may be omitted. For example, when a meal occurs outside the daily sampling window, such as very early or late in the day, participants may not be prompted to report it. Second, dietary intake captured by signal-contingent mEMDA versus 24-hour recalls is subject to lower specificity of timing, particularly when the recall window is longer (eg, >2 hours), as participants are typically asked to report whether or not food (or beverage) consumption has occurred, without elaboration about the specific time it occurred. Finally, although existing signal-contingent methods have typically used lower resource processing methods (eg, not requiring advanced training or specialized software), the resulting data may be limited to quantifying frequencies of intake as opposed to estimates of energy intake or nutrients due, in part, to the brevity of the surveys, a key aspect of signal-contingent mEMDA. Although survey brevity (and frequency of administration) is important for maintaining high response rates [80], it poses limitations to data quality with regard to how the data are summarized and the extent to which they can be used to describe dietary intake. Future research will be needed to determine how best to balance response rates with the collection of high-quality diet data. To summarize, signal-contingent mEMDA is able to capture a representative (ie, randomly selected) sample of individuals’ daily intake while minimizing the participant burden associated with participant-initiated reporting; however, it may not be suited for time-stamping eating events or quantifying energy or nutrient intakes.
Another limitation of mEMDA methods is that few studies have compared mEMDA protocols with the traditional dietary assessment methods. However, the few that have conducted comparison studies have generally found acceptable agreement in estimated energy intakes and reported concordance. For instance, 2 studies that compared mEMDA reports of estimated energy intake against doubly labeled water [40,44] found that mEMDA underestimated energy intake by an average of 222 kcal/day (−15.6%) [50] and approximately 11% [40], respectively. In 4 studies event-contingent [39,41] and signal-contingent recalls [62,63] mEMDA estimated intake was compared with 24-hour dietary recalls. The 2 event-contingent studies demonstrated 85% agreement for primary diet data [41] and 90% agreement for macronutrient and energy intake [39] when compared with 24-hour dietary recalls. The 2 signal-contingent studies found day-level concordance between foods and beverages reported by mEMDA and next day 24-hour dietary recalls ranged from 79% (entrees) to 94% (fruit and vegetables) in a sample of college students [62] and 2-hour concordance ranged from 65% (fruits and vegetables) to 90% (soda and energy drinks) for children [63]. Similar magnitudes of difference have been reported when comparing traditional methods of dietary assessment (eg, Web- and computer-based 24-hour dietary recalls) with a more objective reference method (eg, direct dietary observation, doubly labeled water, and controlled-feeding studies) [81]; however, it is difficult to determine the validity of any method of dietary assessment because of the inherent measurement errors that obscure differences between the observed and the true reports. Moreover, it is important to note that these studies did not report compliance rates for mEMDA versus the comparison method. Despite these limitations, the studies highlighted in this review demonstrate the feasibility and preliminary evidence of accuracy of mEMDA in research settings. It will be important for developing mEMDA methods to conduct similar comparison studies against more objective reference methods and to compare their protocol adherence rates with the traditional methods of dietary assessment.
Strengths and Limitations of This Review
Our systematic review is the first to summarize the existing literature of mEMDA for the measurement of diet in research studies and to discuss corresponding strengths and weaknesses. This study was based on a comprehensive search across multiple domains (images, biomechanical approaches, EMA, etc) using current guidance for a robust systematic review process. This study was limited in breadth as a result of many burgeoning methods not yet being applied in research settings and were, therefore, excluded from this review. Furthermore, senior reviewers made final decisions about study inclusion rather than an independent expert, which may have introduced some selection bias. In addition, because of the wide divergence of study measures and reporting of relevant items (eg, protocol adherence), we were only able to narratively describe data collection methods. Furthermore, this review was limited in its ability to describe how accurately mEMDA methods capture diet compared with traditional methods of dietary assessment because of limited reports of such comparison statistics in the existing studies.
Implications for Future Research
Currently, there is a need within the field of momentary diet assessment to maximize data quality while minimizing participant burden. Ecologically valid and reliable data on individuals’ dietary intake are essential to understand the role of diet on human health through the life span. Due to the known limitations of the existing dietary assessment methods, the research community is motivated to develop new solutions aimed at (semi)automating the assessment of dietary intake. Although the automated methods of real-time image-based detection and real-time detection of food intake by biomechanical sensors or hand-held devices have seen significant progress in terms of identifying foods and estimating portion sizes (detecting wrist or hand motion or patterns of chewing or swallowing indicative of food intake), the design and proof-of-concept data suggest the automation of dietary assessment remains out of reach for the time being. Given these limitations, mEMDA, characterized by the repeated assessment of an individual’s behaviors and experiences in real time or near-real time in their natural settings, represents a novel dietary assessment method. By reviewing the literature and identifying key patterns, strengths, and weaknesses of the existing momentary diet assessment methods, a unique topic of high relevance in the dietary assessment community and EMA community, researchers may better understand and move forward with improving and incorporating mEMDA into their own research. Compared with the traditional methods, mEMDA may reduce participant burden and recall biases while maximizing ecological validity. Therefore, mEMDA has the potential to bridge the gap between currently available methods (eg, 24-hour dietary recall) and newer methods (eg, biomechanical sensors), which are currently under development.
The strengths of event-contingent mEMDA (eg, ability to capture the full day of dietary intake and estimate energy and nutrient intakes) and of signal-contingent mEMDA (eg, lower participant burden and unannounced prompting schemes) could potentially be leveraged to design novel mEMDA methods that reduce their individual limitations. On the basis of these early studies, efforts now need to be focused on standardizing mEMDA methods with the goal of maximizing dietary data quality and ecological validity while minimizing participant burden. With improved standardization, it is likely that validated mEMDA tools will become more widely available to researchers in the future as most methods have been study-specific. Existing studies illustrate the wide range of dietary outcomes assessed through mEMDA methods, and although the ability to develop and tailor assessment items based on a particular study’s needs is an advantage, the divergence of outcome measures and lack of validation remains a major challenge.