Evaluating the Validity of the PortionSize Smartphone App for Estimating Dietary Intake in Free-living Conditions: A Pilot Study

Abstract

Objective:

Evaluate the validity of the PortionSize^™ app.

Methods:

In this pilot study, 14 adults used PortionSize^™ to record their free-living food intake over 3 consecutive days. Digital Photography was the criterion measure and main outcomes were estimated intake of food (g), energy (kcal), and food groups (servings). Equivalence tests with ±25% equivalence bounds, and Bland-Altman analysis were performed.

Results:

Estimated gram intake from PortionSize^™ was equivalent (P<0.001) to Digital Photography estimates. PortionSize^™ and Digital Photography estimated energy intake (kcal); however, were not equivalent (P=0.08), with larger estimates from PortionSize^™. PortionSize^™ and Digital Photography were equivalent for vegetables intake (P=0.008), but PortionSize^™ had larger estimates of fruits, grains, dairy, and protein intake (all P>0.07; error range:11–23%).

Conclusions and Implications:

Compared to Digital Photography, PortionSize^™ accurately estimated food intake and had reasonable error rates for other nutrients; however, it overestimated energy intake, indicating further app improvements are needed for free-living conditions.

INTRODUCTION

Assessing food and nutrient intake accurately is essential for improving health and nutrition as well as reducing the risk of chronic diseases. Current food image-based methods allow capturing dietary data in real-time;^1–5 however, these methods rely on human raters to estimate food intake, meaning they cannot provide real-time and detailed feedback about an individual’s diet. Given that food intake is strongly influenced by decisions made before eating begins,^6,7 providing people with feedback on food selection (before food is eaten) can make it easier to modify their dietary behavior when the behavior occurs.^8,9 Thus, the PortionSize^™ app was developed to provide users with real-time feedback on dietary selection, intake, and adherence to the United States Department of Agriculture (USDA) MyPlate recommendations.¹⁰

The PortionSize^™ app uses specific formulas to calculate the user’s energy and food group requirements based on their sex, age, weight, and height.^11,12 The app integrates food templates and other techniques that allow users to estimate their portions in real time based on images they capture of their food with the app. As the user consumes meals throughout the day, the app provides immediate feedback on the energy, food group, and nutrient content before and after each eating occasion, as well as cumulative feedback on their intake for the entire day.^11,12 The app dashboard is updated in real time and easily allows the user to see if they have achieved their energy, food group, and nutrient targets (Figure 1). These innovative features highlight the potential of the PortionSize^™ app to advance the science of dietary assessment and nutrition counseling. However, evaluation of its usability and validity is crucial to determine if it is useful in various settings.

Figure 1.

The PortionSize^™ app relies on users capturing images of their food selection (before meal) and waste (after meal; if applicable). The Portion Summary can be viewed per meal/snack or overall day summary (as shown).

Preliminary validity data from the PortionSize^™ app has been reported by Saha et al.,¹² who tested the accuracy of the app in laboratory conditions in a small sample of 15 adults. The results showed that PortionSize^™ overestimated energy intake by ~13% when compared to weighed intake. Additionally, PortionSize^™ produced equivalent estimates for consumption of food (g), sugars, fruit, and dairy compared to weighed intake.¹² The results from the study reported herein were obtained from the same sample of participants who subsequently used the PortionSize^™ app over 3 days in free-living conditions. In the current pilot study, PortionSize^™’s validity was tested against a criterion measure, the validated Digital Photography of Foods Method^1,2,13,14. Further, we quantified ratings of participants’ satisfaction and usability of PortionSize^™. We hypothesized that food (g), energy (kcal), and food groups (servings) intake estimates from the PortionSize^™ app would be equivalent to the criterion measure and that the app’s bias would not differ over levels of intake. The results of this free-living pilot study will identify ways to improve the PortionSize^™ app, which may eventually assist individuals in accurately managing their dietary intake.

METHODS

Participants

The study had 2 phases: Phase 1 consisted of a laboratory test meal, which has previously been published,¹² while Phase 2 included 3-day intake estimation in free-living conditions. Participants were recruited from August to November 2020 by distributing flyers in the community, advertisements on the Pennington Biomedical Research Center (PBRC) website, and social media platforms. All interested individuals completed a screening questionnaire by telephone, and preliminarily eligible participants were scheduled for an in-person screening visit. Adults between 18 and 65 years of age with a BMI of 18.5–45 kg/m² who had an iPhone 6 or above were eligible to participate. Participants were excluded from participation if they self-reported having an eating disorder or serious mental illness.

It has been suggested that a sample size of 10–15 participants is sufficient for a pilot study;^15,16 thus, 15 participants were enrolled in the first phase of the study. Following Phase 1,¹² 1 participant did not complete Phase 2 (free-living phase); thus, the current analysis included 14 participants. All participants provided written informed consent before enrollment. Participants received $75 upon completion of both phases of the study. All study procedures were approved via expedited review by the Institutional Review Board at Pennington Biomedical Research Center. The study is registered at ClinicalTrials.gov (NCT04494971).

Procedures and Measurements

In Phase 1 of the study, participants visited the laboratory, where anthropometric measurements were taken by trained staff, and BMIs were calculated. The goal of this phase was to examine the validity of the PortionSize^™ app under controlled conditions, and was published previously.¹² During this first phase of the study, participants received approximately half an hour of experiential training, which provided information and instructions on how to use the PortionSize^™ app. A test meal, comprised of plastic food models, was provided during the training so participants could practice estimating food portions using the app. The training was complete once participants demonstrated mastery using the app. The training was followed by a lab-based meal that allowed participants to measure simulated intake of test meals with the PortionSize^™ app.

Within approximately 1 week of training, participants began the second phase of the study, which tested the usability and the validity of the PortionSize^™ app in free-living conditions against a criterion measure, Digital Photography of Foods. In Phase 2, participants were instructed to follow their regular eating routines but to use the PortionSize^™ app to track their food selection and consumption of all eating occasions for 3 consecutive days (2 weekdays and 1 weekend day). This second phase also included an Ecological Momentary Assessment (EMA) approach to prompt and remind participants to use the app throughout the observation period,¹⁷ with the goal of improving data quality and minimizing missing data (see the following section for the details of the EMA prompts).^2,17 To obtain the criterion measure, participants captured their typical meals using the app, and the images were transferred to a secure web-based server, where 2 trained raters quantified the amount of food in each image based on the Digital Photography of Foods Method. After recording their typical 3-day dietary intake, participants completed a satisfaction survey to evaluate the usability of the app.

PortionSize^™ App

Participants were instructed to use the PortionSize^™ app to record all foods and beverages that they consumed in real-time. Participants were instructed to follow their typical eating style and routines. When a participant selected their food for an eating occasion, they captured a before-meal image that included the amount selected for that eating occasion. To facilitate accurate estimation portion sizes/food intake from food images, participants placed a reference card next to the food items when capturing images of their meals, which standardizes the distance and angle of the image (Figure 2A). When the participants took the image, they identified each meal item by selecting the item from the onboard database within the PortionSize^™ app. This database includes a subset of the USDA’ Food and Nutrient Database for Dietary Studies (FNDDS)¹⁸ database (version 2017–2018). Therefore, the PortionSize^™ app currently contains approximately 1,150 food items linked to food codes in the FNDDS database, allowing participants to identify items (or similar items) within meals. The participants identified meal items by tapping on a meal item on the screen to bring up a drop-down list of foods and beverages. Using the search function, they could also seek out specific foods (shown in Figure 2B).

Figure 2.

To accurately estimate portion sizes, participants placed a reference card in the image. When the angle and distance were appropriate, the rectangular reference box on the screen turned green (A). To identify individual items within meals, participants tapped on a meal item on the screen to bring up a drop-down list of foods and beverages (B). For each food item, there were specific templates, and participants could adjust the size and move the template to cover the food to estimate portion sizes (C).

After finding the food item on the list, the user clicked “select” and was guided by the app to use the food templates (Figure 2C). There were specific templates for each food item; for example, a deck of cards was appropriate for steak, and a poker chip was appropriate for crackers. To estimate portion sizes, participants could adjust the size and move the template to cover the food. For example, as shown in Figure 2C, a hamburger shaped template was used to estimate the portion size of a hamburger in the app. After participants tagged and estimated the portions of the foods in the before-meal photo, they clicked the “done” button, and the app automatically calculated dietary information based on the templates and the built-in nutrient database. Immediately afterward, a meal summary screen appeared, providing the dietary information for the meal. The information provided to the users included energy, fruit, vegetables, grains, protein, and dairy intake (servings), as well as amounts (g) of saturated fat, added sugars, and alcohol. Finally, after consuming the meal, participants captured an after-meal image and indicated whether they had eaten everything or their percentage of remaining food. In cases where food images were not captured when food was eaten, participants used the “forgot meal” tab in the PortionSize^™ app to record their meals. These meals were added to their daily intake and displayed on their PortionSize^™ app Portion Summary screen.

Participants were provided with Ecological Momentary Assessment (EMA)^2,17 prompts to improve data quality and minimize missing data. Each participant received 4 to 6 automated prompts throughout the day, which reminded them to take photographs using the PortionSize^™ app. Prompts were sent through text messages and were customized to be delivered around individual participants’ mealtimes. Examples of prompts include: “Did you eat or drink anything today and forget to take a picture?” and “Can you take before and after pictures of your lunch?”. The participants were instructed to respond to each prompt with a yes or no response, or they could type in more detailed responses.

Criterion Measure

The criterion measure was Digital Photography of Foods, which was possible since the meal photos that participants captured with PortionSize^™ were transmitted wirelessly and automatically to a server at PBRC. Using existing and validated visual comparison procedures,^1,2 human raters analyzed these food images to estimate dietary intake. To determine the level of interrater reliability, we oversampled 20% of the images and used interclass correlation coefficients to compare the raters’ estimates of food selection, plate waste, and food intake. This evaluation revealed excellent interrater reliability among the 2 raters (intraclass correlation coefficients > 0.95). Ratings were conducted using a computer program built at PBRC called the Food Photography Application^©.^1,2 The raters received at least 15 hours of training by the designated master rater, in which they learned how to use the Food Photography Application^© software. The tasks of trained raters were threefold: 1) to match each food item depicted in the “before” image to a proper food code in the FNDDS; 2) to search and select a food-specific image of a standard portion-size; and 3) visually compare participants’ images to the standard portion images to estimate the amounts of foods in the before- and after-meal images and enter those values into the software. The software then calculated dietary information based on the entered amount and the FNDDS¹⁸ database (version 2017–2018), which includes 7,083 food and beverage items.

Other Measurements

Body weight and height were assessed by trained researchers and were used to calculate the participant’s BMI.¹² Participant’s demographics were assessed using a survey and their sex was self-reported by selecting male or female responses. Race or ethnicity was also self-reported by the participants from a list including non-Hispanic White, non-Hispanic Black, Hispanic, Asian or Pacific Islander, Native American (including Alaskan), biracial or multiracial (specify), or other (specify). Participants completed an 8-item survey that was adapted from prior studies^19,20 to obtain: overall satisfaction with the PortionSize^™ app, satisfaction with embedded food templates and app training, and ease of use. All items were rated on a scale ranging from 1 to 6, with 1 indicating “extremely dissatisfied” or “not at all” and 6 indicating “extremely satisfied” or “very much.”

Statistical Analysis

The primary outcomes of this study were estimations of intake of food (g), energy (kcal), and food groups (oz. eq. and cup. eq.) using the PortionSize^™ app compared to the criterion method (Digital Photography of Foods) across 3 consecutive days in free-living conditions. The secondary outcomes of the study were macronutrients (carbohydrates, fat, and protein), selected nutrients (saturated fat, cholesterol, dietary fiber, sugars, and added sugars), and selected micronutrients (sodium, calcium, iron, potassium, and vitamin D). The rationale for selecting these specific nutrients was to include the nutrients that are available for consumers in nutrition fact panels.²¹ Influence diagnostics were used to identify data points exhibiting high leverage on model parameters for estimations of food (g) and energy (kcal) intake.²² The influential data points were excluded from all analyses and reported with residuals and Cook’s D values.²³

A series of analyses were conducted to test the validity of the PortionSize^™ app. First, all outcomes were analyzed by assessing the equivalence between the PortionSize^™ app and the Digital Photography of Foods method by using the Two One-Sided T-Test (TOST) method.²⁴ These analyses relied on data from the mean outcome across the 3 days of measurement. Based on a previous validation study,²⁵ the equivalence bounds were set at ± 25%, with significance indicating evidence that the measures are equivalent and their difference falls within the 25% equivalence bounds. We also estimated the mean percent error (i.e., [(PortionSize^™ - Digital Photography)/ Digital Photography] × 100) at the group level to quantify measurement error between PortionSize^™ and Digital Photography. Second, Bland-Altman analyses were used to evaluate the level of agreement between the 2 methods for estimations of all outcomes.^26,27 Differences or bias between the 2 methods on the y-axis was plotted against the mean of the 2 methods on the x-axis. The zero-bias line, 95% upper or lower confidence limits (mean difference ± 2 SDs of the differences), and the regression trend line were included on the same plot. For primary outcomes, the significance level was set at 0.05; however, for secondary outcomes, we set a P-value of 0.01 in order to control the error rate.²⁸

A linear mixed model with repeated measures was used to determine whether estimations for food and energy intake outcomes differed between the weekdays and the weekend days or across the 3 days.²⁹ These analyses relied on data from daily intakes of food (g) and energy (kcal). Standardized effect sizes (Cohen’s d) are reported for the primary and secondary outcomes. Based on Lakens et al. (2018), the hypothesized effect sizes would be small since equivalence between PortionSize^™ and criterion measure was expected.³⁰ Results from linear mixed models were reported as adjusted means ± SEM; all other results were reported as mean ± SD. All analyses were performed using SAS software (SAS version 9.4, SAS Institute, Inc., Cary, NC, 2013).

RESULTS

Subject Characteristics

A total of 14 participants were included in the current analysis. Demographic characteristics of participants are shown in Table 1. Participants were white (100%) and predominantly female (71%). Mean ± SD age and BMI were 26.4 ± 11.0 years and 22.9 ± 4.6 kg/m², respectively. Half of the participants reported some college (50%) as their highest level of education, and the majority reported a household income ≤ $50,000.

Table 1.

Characteristics of Adults Using PortionSize^™ to Record Free-Living Food Intake Across 3 Consecutive Days (n=14)

Characteristic	n (%)
Sex
Male	4 (29%)
Female	10 (71%)
Education
High school diploma or General Educational Development	1 (7%)
Some college	7 (50%)
Bachelor’s degree	4 (29%)
Postgraduate degree	2 (14%)
Employment
Unemployed	2 (14%)
Full-time employment	3 (22%)
Part-time employment	7 (50%)
Retired	1 (7%)
Other	1 (7%)
Characteristic	Mean ± SD (range)
Age (years)	26.4 ± 11.0 (20 – 57)
Height (cm)	169.5 ± 9.0 (152 – 183)
Weight (kg)	66.7 ± 19.6 (50 – 113)
BMI (kg/m2)1	22.9 ± 4.6 (19 – 34)

¹BMI, body mass index

Dietary Estimations

Participants recorded 147 eating occasions across 3 days in free-living conditions. Examination of the influence diagnostics for the model of dietary intake estimates identified 6 meals reported by 5 different participants as influential. These meal entries had extremely high residuals, leverage, and Cook’s distance (all D_i > 0.05).²³ It was evident that the estimations from these meal entries were extremely high due to overestimation caused by the misuse of portion templates in the PortionSize^™ app. Thus, we excluded the 6 influential meal entries from both methods, resulting in a total of 141 eating occasions. Detailed information regarding the excluded meal items was included in Supplemental Table 1.

The meal photos that were captured with the PortionSize^™ app were also utilized for the Digital Photography of Foods evaluation; therefore, both methods used the same data source for individual meal entries. Of the 141 eating occasions recorded by both methods across 3 days, 2.4 ± 0.6 occasions were entered as main meals/day, and 1.0 ± 0.7 were entered as snacks/day, where 88% included food items and 12% were beverages (10% of beverages were alcoholic beverages). Participants responded to 181 (85%) of the 213 EMA prompts (with a range of 12 – 20 per participant) that were delivered around their mealtimes. Users recorded 7% of eating occasions utilizing the forgot-meal tab of the PortionSize^™ app.

When looking into dietary estimations for each method, estimations of food intake (g/d) were not significantly different on weekdays compared to weekend days for PortionSize^™ (1081 ± 123 vs. 1155 ± 153 g/d, respectively; P = 0.64) and Digital Photography (1073 ± 121 vs. 1112 ± 121 g/d, respectively; P = 0.82). The results were similar for energy intake estimates when comparing weekdays to weekends for PortionSize^™ (1452 ± 136 vs. 1778 ± 176 kcal/d, respectively; P = 0.11) and Digital Photography (1317 ± 116 vs. 1415 ± 145 kcal/d, respectively; P = 0.53). Similarly, reported food (g) and energy intake (kcal) did not differ significantly across 3-days for both methods (all P > 0.29).

Based on PortionSize^™ and Digital Photography data, participants consumed 97 ± 3 % of the amount of food (g) they selected. PortionSize^™ estimates indicated that participants consumed 44% of their daily energy as carbohydrates, 17% as protein, and 39% as fat. Digital Photography estimates were similar: participants consumed 44% of their daily energy as carbohydrates, 18% as protein, and 38% as fat. Supplemental Table 2 shows the food group estimations compared to daily nutritional goals.

Differences Between the PortionSize^™ App and Digital Photography

The results of the equivalence test indicated that estimated food intake (g) by PortionSize^™ (mean ± SD, 1105 ± 418 g/day) was equivalent to Digital Photography estimates (1086 ± 396 g/day, P < 0.001). The mean food intake (g) difference between the 2 methods was 19 ± 189 g, and the mean percent error (MPE) for the estimation of food intake (g) was 1.8% (Table 2). PortionSize^™ and Digital Photography estimates of energy intake (1561 ± 447 vs. 1350 ± 389 kcal/day); however, were not equivalent (P = 0.08), with higher estimates from PortionSize^™ (211 ± 318 kcal/day; MPE=15.6%) compared to Digital Photography (Table 2). As shown in Figure 3A, the Bland-Altman plot for total intake in grams showed a mean difference of 19 grams, and the 95% CL of agreement located above and below the mean difference was 389 g and 351 g, respectively. The measurement error of PortionSize^™ in estimating food intake was consistent over different levels of food intake (P = 0.68). A similar pattern of error over different levels of energy intake was observed for total energy intake, as shown in Figure 3B; and the regression indicated that bias did not differ over levels of intake (P = 0.50).

Table 2.

Comparison of Portion Size, Energy, and Nutrient Intake Estimates Between PortionSize^™ and Digital Photography (n=14).

Outcome	PortionSize™ App		Digital Photography		Difference		Equivalence at ±25%	Effect Size1	Mean Percent Error2
	Mean	SD	Mean	SD	Mean	SD
Portion Size (g)	1105.2	418.4	1086.1	396.0	19.1	188.7	0.0001 3	0.05	1.8
Energy (kcal)	1560.6	446.8	1349.7	389.2	211.0	317.7	0.080	0.50	15.6
Fruits (cup eq.)	1.1	1.0	1.0	0.8	0.1	0.6	0.2683	0.15	13.8
Vegetables (cup eq.)	1.3	0.9	1.4	0.9	−0.02	0.4	0.008 3	−0.02	−1.4
Grains (oz. eq.)	5.2	2.3	4.7	2.0	0.5	1.7	0.0797	0.23	10.6
Dairy (cup eq.)	1.3	0.9	1.1	0.7	0.2	0.6	0.413	0.29	21.7
Protein (oz. eq.)	5.6	3.6	4.5	2.6	1.0	1.5	0.3985	0.33	22.7
Saturated Fat (g)	21.4	8.4	18.0	8.5	3.4	4.0	0.1669	0.40	19.0
Added Sugars (tsp)	6.5	4.6	5.8	3.5	0.8	2.2	0.1282	0.18	13.1
Protein (g)	68.4	23.0	61.7	20.5	6.7	10.5	0.0042 3	0.31	10.9
Total Fat (g)	68.6	23.9	58.0	20.6	10.6	16.1	0.1884	0.47	18.2
Carbohydrates (g)	174.3	66.4	147.8	26.5	39.6	55.9	0.2476	0.44	17.9
Dietary Fiber (g)	19.2	12.5	14.7	7.1	4.5	10.1	0.6158	0.44	30.5
Sugars (g)	60.0	29.9	49.4	22.7	10.6	15.8	0.3457	0.40	21.5
Cholesterol (mg)	266.0	203.0	223.6	149.6	42.5	81.4	0.2742	0.24	19.0
Sodium (mg)	2724.5	807.8	2172.7	593.4	551.8	498.5	0.5251	0.78	25.4
Calcium (mg)	655.8	335.6	619.5	295.8	36.3	200.4	0.0226	0.11	5.9
Iron (mg)	20.6	32.2	9.6	2.6	11.1	31.5	0.8403	0.48	116.1
Potassium (mg)	2029.7	820.6	1847.2	637.2	182.5	493.4	0.027	0.25	9.9
Vitamin D (μg)	4.2	3.3	3.0	2.3	1.2	2.1	0.7537	0.40	37.9

¹Standardized effect sizes are represented by Cohen’s d.

²Mean percent error (Mean ± SD) = ([PortionSize^™ − Digital Photography]/ Digital Photography) × 100.

³Indicates significant results of Equivalence tests. For primary outcomes, estimated intake of food (g), energy (kcal), and food groups (servings), the significance level was set at 0.05; however, for secondary outcomes (all other outcomes), we set a P-value of 0.01.

Figure 3.

Bland-Altman plots comparing food intake (A) and energy intake (B) estimates using the PortionSize^™ app and Digital Photography in 14 adults. UCL: upper confidence limit; LCL: lower confidence limit; PS: PortionSize^™. The measurement error of PortionSize^™ in estimating food intake was consistent over different levels of food intake (R² = 0.01; P = 0.68). A similar pattern of error over different levels of energy intake was observed for total energy intake, and this bias was not statistically significant (R² = 0.04; P = 0.50).

PortionSize^™ and the Digital Photography were equivalent for vegetables (P = 0.008), but the PortionSize^™ app had larger estimates of fruits, grains, dairy, and protein intake (all P > 0.07). While the magnitude of food group error was relatively low for fruits (0.13 cup eq.) and dairy (0.23 cup eq.), the errors were more meaningful for grains (0.49 oz. eq.) and protein (1 oz. eq.). Among the 5 food groups (Table 2), estimations of vegetables had the lowest mean percent error (−1.4%), with those for protein having the highest (22.7%). Bland-Altman analysis for all food groups had wide limits of agreement, and no significant trends were found (all R² < 0.22, P > 0.09) except for protein. The regression line was significant (R² = 0.42, P = 0.01), with PortionSize^™ underestimated protein intake at lower levels of intake, but overestimation occurred and increased with higher levels of intake.

The estimates for protein were equivalent (P = 0.004); however, the estimates of fat (P = 0.19), and carbohydrates (P = 0.25) were not equivalent between PortionSize^™ and the Digital Photography (Table 2). All other nutrient intake estimates were not equivalent between the 2 methods (all P > 0.02). Among all nutrients, the estimations of calcium had the lowest mean percent error (5.9%), with those for iron having the highest (116.1%). Bland-Altman analysis for all nutrients had wide limits of agreement, and no significant trends were found (all R² < 0.42, P > 0.01) except for cholesterol and iron. The regression line was significant for cholesterol (R² = 0.44, P < 0.009) and iron intake (R² = 0.98, P < 0.001); PortionSize^™ underestimated intake of cholesterol and iron at lower levels of intake, but overestimation occurred and increased with higher levels of intake.

Satisfaction and Ease of Use of the PortionSize^™ App

Participants found the PortionSize^™ app easy to use before (72%) and after (72%) a meal was consumed, as indicated by their choice of the 2 most favorable ratings (Supplemental Table 3). More than half of the participants (58%) reported that the food templates provided in the PortionSize^™ app were appropriate. In addition, a large proportion of participants rated the PortionSize^™ app with the 2 most favorable ratings for satisfaction when they captured before (57%) and after (72%) meals. Lastly, 93% of participants indicated that the iPhone training for the PortionSize^™ app helped prepare them for using the app.

DISCUSSION

In free-living conditions, the PortionSize^™ app provided equivalent estimates for food intake, but not energy intake, compared to Digital Photography of Foods. Additionally, the PortionSize^™ app provided equivalent estimates to Digital Photography only for vegetables, with relatively minimal errors for fruits and dairy intake. Following 3-days of using the PortionSize^™ app, participants were satisfied with the app and found it easy to use. The results of this pilot study indicate that the PortionSize^™ app was effective in providing similar estimates for some nutrients to the validated Digital Photography and was acceptable to participants. Nonetheless, the validity of PortionSize^™ in free-living conditions requires comparisons against gold standard measures free-living energy intake, such as the doubly labelled water method.

The PortionSize^™ app uniquely features superimposed portion templates that assist users in estimating food portions in real-time. Results from the current study indicated that the estimated food intake (g) from the PortionSize^™ app were equivalent to Digital Photography. Estimated energy intake (kcal), however, differed between the 2 methods. These results were supported by our previous findings in the controlled-feeding phase, in which participants’ dietary estimations with the PortionSize^™ app during a lunch meal were compared to weigh-back measurements.¹² Taken together, these findings suggest that, although the app users and raters estimated similar amounts of food, there may be some discrepancies between the type of food selected in the onboard nutrient database between users and raters. For example, participants may eat a burger and select the condiments and cheese options in the app’s search function but fail to provide these details in the meal description, which is critical to the Digital Photography process. Furthermore, since raters had access to the entire FNDDS database, it is possible they selected the exact food type or the available combined (mixed) food option, while participants used more generic food types or logged separate ingredients when recording their foods using the PortionSize^™ app. Based on a recent report,³¹ defining the specific food items and types has implications for the estimation of energy intake. Therefore, considering these findings, we revised our training and built-in food database to help participants more accurately find and select the foods that they are eating.

Analyzing the food group estimates for both methods, we found equal estimates only for vegetables. However, the error rate was modest for other food groups, and the effect sizes were small. The controlled phase of the study found similar results,¹² but with higher error rates for fruits and grains. Moreover, only protein was equivalent between the 2 methods, whereas iron had the highest error rate. Similarly, the controlled phase found equivalent estimates for potassium and sugars, and iron estimations had a large error rate.¹² Since food intake encompasses energy intake, nutrient intake (macronutrients, micronutrients, vitamins, minerals), and intake of various food groups (e.g., fruits, vegetables), improvements in the PortionSize^™ app that enhances the estimation of portion sizes have the potential to lead to more accurate estimation of other nutrients.

Previous studies suggest that adherence to more burdensome methods of assessing dietary intake can be low^32,33 and we found that methods that were perceived as more burdensome were less preferred.¹¹ Thus, use of PortionSize^™ may be negatively affected by user burden since, for example, it is more burdensome than apps that require the user to only capture an image of their foods.¹¹ Specifically, the app provides users with real-time feedback on dietary selection, intake, and adherence to USDA MyPlate recommendations. In this pilot validity study, participants were instructed to follow their regular eating routines without making any attempt to align their food and energy intake with the provided real-time information. The use of real-time feedback to alter participant eating behavior may increase motivation and lead to a better overall experience. In addition, because the current study used an early version of the app, participants experienced app glitches and data transfer issues. We attempted to address these issues by communicating with participants through EMA prompts; however, these issues increased participant burden when using the PortionSize^™ app. Despite these practical issues, we found that users were satisfied with the PortionSize^™ app and found it easy to use. In addition, almost half of the participants were satisfied with the feedback provided by the app regarding their portion sizes, which indicates that users were satisfied with the real-time feedback provided by the app. The user satisfaction results, however, should be interpreted with caution as the participants only used PortionSize^™ and only for 3 days.

A strength of this pilot study is that the participants captured dietary intake over 3 days in ecologically valid free-living conditions, which included a weekend day. Another strength is that the images that were captured with the PortionSize^™ app were used for the Digital Photography method. Therefore, we pilot tested the validity of the app compared to human raters using an identical data source. The study also has important limitations that must be considered. First, we excluded influential points. Second, the PortionSize^™ app was only available for iPhones. Third, the sample was entirely White and predominantly female, which may preclude the generalization of the findings to larger, more heterogeneous samples. Last, the food items in the PortionSize^™ app contained only a subset of the FNDDS database, which may have influenced the current findings and participants’ experience in the free-living conditions.

The PortionSize^™ smartphone app was developed to provide real-time feedback on food selection, intake, and USDA MyPlate food groups to assist individuals in adhering to dietary recommendations. Based on Fogg’s behavior model, behavior change is effective only if motivation, ability, and triggers are present at the same time.³⁴ Thus, with its real-time feedback, the PortionSize^™ app has the potential to promote healthy eating across large groups. Evaluation of PortionSize^™’s validity in free-living settings found that PortionSize^™ produced accurate food intake estimates compared to the Digital Photography without requiring human rating. The PortionSize^™ app, however, overestimated energy intake, indicating additional app improvements are needed for free-living conditions.

IMPLICATIONS FOR RESEARCH AND PRACTICE

The PortionSize^™ app estimated food intake accurately and provided reasonable effect sizes for other dietary estimates compared to the Digital Photography. These findings provided insight into areas in which the PortionSize^™ app needs to be improved, such as portion templates and EMA features. The findings also informed the addition of a feature to reduce misuse of the PortionSize^™ app and to minimize influential data points. Specifically, the app now prompts participants when their portion estimate is 4 times larger than the standard portion. This change was made after the study reported herein. For a proper validity test for dietary intake estimates, the app needs to be further tested in various conditions with a more diverse and gender-balance sample to ensure the study results are more representative. Since app development is a dynamic and continuous process, for future versions, it may be possible to reduce burden and increase engagement by incorporating emerging technologies, such as automated food identification and volume estimation.^35,36

There are several unique aspects of the app that make PortionSize^™ a promising approach for use in various settings, including real-time portion estimation and personalized dietary intake guidance regarding the extent to which a person’s consumption throughout the day meets specific energy intake and nutrition goals. Considering the ongoing trials and improvements, the PortionSize^™ app has the potential to advance nutrition counseling, dietary assessment, and serve as a useful intervention tool. Decreasing participant burden while increasing engagement can help to fulfill the eventual goals of the app, and maximize clinical utility while minimizing researcher burden, costs, and time.