Validity of an AI-based Program to Identify Foods and Estimate Food Portion Size: Complete Statistical Analysis

1. Trial at a Glance

This page covers the registry results for Validity of an AI-based Program to Identify Foods and Estimate Food Portion Size (NCT05343585), a na na study. The primary outcome is Identification of Food Plated Using the Openfit Mobile App.

2. Clinical Question

The question is about the app’s identification performance in a controlled laboratory exercise, not about dietary intake during ordinary meals. The primary outcome description focuses on agreement at the food-item level and across the tested items. It also describes comparing the proportion correctly identified by Nutrition AI with the proportion correctly identified by a human rater.

The registry record states that 255 food items were tested across participants. That item count describes the material assessed; it does not establish how many participants contributed usable observations to each analysis, or how observations were distributed among participants. Those details are not reported in the registry record.

The record names portion estimation in the study title and describes the app’s purpose as identifying foods and estimating food portion size. However, the primary outcome described in detail is identification of food plated. The registry record does not report a numerical result for identification accuracy or a primary result for portion-size estimation. The clinical question therefore remains one of intended evaluation rather than a documented estimate of performance.

3. Trial Design

This was a completed, single-group pilot study listed under Nutrition Assessment. The registry record lists enrollment of 24 and one experimental group. Allocation is listed as not applicable, masking as none, and the study model as single group. Accordingly, the design is not a randomized comparison between separate intervention and control groups.

The intervention is named PortionSize AI and classified as a device. The arm description specifies training and use of Openfit, use of the app to estimate food intake from simulated laboratory meals, and ratings of usability and satisfaction. The meals took place at PBRC or LSU, and participants did not eat food during them. The registry record describes the setting as controlled and gives the visit timeframe as approximately two hours.

The listed study start date is 2022-04-27, and the primary completion date is 2022-06-03. The study status is completed. The sponsor is Pennington Biomedical Research Center, and the phase is listed as NA. These administrative details describe the registered study; they do not provide information about the results or the conduct of each analysis.

The primary outcome concerns identification of food plated using the Openfit mobile app. The description distinguishes Nutrition AI without correction from Nutrition AI with user correction and refers to a human rater as the criterion method. A food name did not need to be identical to count as a match: a close match could qualify under the stated definition.

Although the record has an experimental-group label, the design does not support treating that group as a like-for-like control group. The registry record does not report a randomized allocation, a masked assessment, or a separate concurrent comparison arm. It also does not report detailed procedures for participant selection, the order of meal tasks, or how the human rating was conducted. These aspects should not be inferred from the single-group designation or from the brief study description.

4. Participant Flow and Accounting

The registry reports how participants moved through each stage of the study. All counts below are exactly as posted; missing counts are not inferred from other stages.

Overall Study

5. Baseline Characteristics

The table reproduces the baseline characteristics posted in the registry, which describe the participants analysed at baseline and help show whether the randomized groups were comparable.

6. Primary Endpoint

The registry lists the primary outcome as Identification of Food Plated Using the Openfit Mobile App, measured over One study visit of ~2 hours.

The primary endpoint assesses identification of foods and beverages plated for simulated meals using the Openfit mobile app. It is expressed in food items and is described as agreement between the app’s identification and known identification. The outcome description considers identification item by item and across the tested items.

Two app-based approaches are specified: Nutrition AI without correction, described as automated, and Nutrition AI with user correction, described as semi-automated. The description also refers to a criterion method using a human rater. It says proportions will be used to assess whether the percentage of items correctly identified by Nutrition AI differs from the percentage correctly identified by that criterion method. Descriptive data are also intended to characterize how often food items were correctly identified across participants.

The matching rule matters for interpretation. A food name could count as an exact match if it matched or was a close match to the food served; the registry gives fruit cocktail identified as fruit salad as an example of an acceptable match. Thus, the endpoint does not necessarily require identical wording. The record does not report additional detail about how close matches were adjudicated.

This endpoint is relevant to the stated goal of evaluating food identification because it compares app classifications with known food identity. It does not, by itself, establish accuracy in estimating portion size, performance during meals in everyday settings, or effects on eating behavior. The registry record reports that 255 items were tested, but it does not post the observed identification proportions, an uncertainty estimate, or the comparison result.

7. Why No Re-analysis Is Shown

8. Statistical Interpretation

9. Statistical Methods Explained

The study’s stated outcome is a count-type measure centered on food items. The description translates the app’s classifications into whether an item was correctly identified under a matching rule, then proposes use of proportions to compare app identification with the criterion method. That makes the target of interpretation a rate of correct identification among assessed items, rather than a continuous measurement such as a mean portion-size error. The registry record does not post those proportions or the underlying counts.

What does agreement measure?

Agreement describes whether two approaches give matching classifications under a defined rule. In this study, the app’s identification is considered against known identification, and a human rater is also named as a criterion method. Agreement is not identical to overall usefulness or validity in every setting. Its meaning depends on what qualifies as a match, which items are included, and whether app and criterion assessments are made on comparable material.

The matching rule allows a close food-name match to count, not just an exact wording match. This can be sensible when different terms refer to sufficiently similar foods, but it also means the definition influences the resulting proportion. General statistical principle: measurement rules should be specified before results are interpreted, because changing a classification rule can change which observations count as correct. The registry record supplies an example but does not report a detailed adjudication procedure or whether assessments were made independently.

Why use proportions?

A proportion puts the number of correctly classified items in relation to the number assessed. It is a natural descriptive measure when the outcome for each item is classified as correct or not correct. Comparing proportions can describe whether app output and the criterion method differ in the fraction of items identified correctly. However, the record does not state which inferential test or model was used, and it posts no comparison estimate or uncertainty interval.

General statistical principle: a proportion is only interpretable with its denominator and the population or collection of observations to which it refers. The record reports that 255 food items were tested across participants, but does not report how many items were eligible for each app condition, whether all items received all assessments, or the per-participant distribution. Those details matter when determining the denominator for a particular estimate.

Why might items not be independent?

General statistical principle: observations from the same participant can be more alike than observations from different participants. If each participant contributes multiple food items, an analysis that treats every item as independent could understate uncertainty. Methods that account for clustering or repeated observations can address this dependence, but the appropriate method depends on the design and data structure. The registry record does not report whether or how dependence among food items from the same participant was handled.

A related consideration is variation in item difficulty. Some foods may be easier to recognize than others, and results can depend on the mix of items tested. General statistical principle: an overall proportion averages over the particular items and participants included, so it may not represent performance for a different collection of foods or users. The record does not give the food-item composition or item-level results, limiting assessment of this issue.

What does uncertainty add?

A point estimate, if reported, would describe the observed performance in the assessed sample. A confidence interval is a general statistical tool for showing the uncertainty around an estimate under the assumptions of the analysis. It does not guarantee that the true value lies within the interval, and it does not correct for design limitations or a mismatched analysis method. No confidence interval for the primary outcome is reported in the registry record.

For a comparison of proportions, an uncertainty interval around the difference could help assess both the direction and range of plausible differences under the model. A non-significant result, by itself, would not establish that methods are equivalent. Equivalence requires a prespecified margin and an analysis designed to evaluate that question; those details are not reported in the registry record. The record also does not report a p-value or a conclusion from the planned comparison.

What can be said about the planned analysis?

The registry outcome description says proportions will be used to assess whether the percentage of items correctly identified by Nutrition AI differs from the percentage correctly identified by a criterion method. It also says descriptive data will characterize the frequency of correct identification across items and participants. These statements identify broad analytical aims, but do not specify the statistical test, model, handling of missing item assessments, or adjustment for participant-level clustering.

The registry lists four outcome measures and zero analyses. That means analytical results are not posted in the registry record; it does not establish that no analysis was performed outside the record. The record does not report an analysis population, exclusions, or a statistical analysis plan. Consequently, no estimate should be reconstructed from the design description alone.

Planning calculations, when presented for a study, answer a different question from observed results. General statistical principle: a sample-size illustration based on an observed effect is a planning scenario, not observed power and not a substitute for an outcome estimate. The registry record does not report the numerical planning outputs for this study. No sample-size conclusion or claim about adequacy can be drawn from the enrollment figure alone.

10. Multiplicity, Missing Data and Analysis Populations

The registry record lists four outcome measures, but does not report their complete definitions, hierarchy, or results in the information describing the primary outcome. It names one primary outcome concerning identification of food plated, while the detailed description distinguishes automated and user-corrected app results and also refers to a human-rater comparison. The record does not explain whether these were treated as separate tests or as components of one analysis.

Multiplicity arises when several outcomes, comparisons, or analyses are examined: the chance of a seemingly notable result can increase when many tests are considered. General statistical principle: interpreting multiple results is clearer when the analysis plan specifies which outcomes are primary and how comparisons will be handled. Whether any multiplicity adjustment was planned or used is not reported in the registry record. The record lists zero analyses, so it does not provide posted inferential results to evaluate for multiplicity.

Missing-data handling is also not reported in the registry record. The record does not say whether any food-item assessments were missing, how incomplete app or rater assessments were handled, or whether participants with incomplete data were excluded from a particular analysis. The overall item count does not establish the denominator used for each comparison.

The analysis populations are not reported in the registry record. It is therefore not possible to determine which participants or items contributed to the primary outcome analysis, whether all enrolled participants were included, or whether any exclusions occurred. These omissions do not establish that data were missing or that exclusions took place; they mean the registry record does not describe these matters.

11. Safety Results

The registry reports serious adverse events as affected participants divided by participants at risk.

Deaths

Other (non-serious) adverse events

Participants with at least one non-serious event at or above the registry's 0% reporting threshold.

The registry safety figures list the experimental group as having 24 participants at risk for serious adverse events and 0 affected. For other adverse events, the record lists 24 at risk and 0 affected. Deaths are listed as 0, and no serious adverse-event terms are listed.

These figures describe what is recorded for this study group; they are not a comparison of safety between randomized groups. The study is single-group, and the registry record lists no masking. The counts should not be read as demonstrating that an app or procedure has no risks. A zero count in a study record means no affected participants were recorded in that category in the registry entry, not that the possibility of an event has been ruled out.

The record does not provide event descriptions, follow-up details beyond the study context, or additional information on how adverse events were solicited or classified. It also does not report a separate safety analysis. The listed figures are therefore best interpreted narrowly, as the adverse-event counts recorded for the group rather than as a comprehensive estimate of safety in other settings or populations.

12. What the Result Does, and Does Not, Mean

The study record establishes that a pilot evaluation of an AI-based app was conducted in a controlled laboratory setting using simulated meals. It describes an intended assessment of food identification, including app output without user correction and after user correction, with comparison to known identification and a human-rater criterion method. It also reports that 255 food items were tested across participants.

What the record does not establish is the app’s observed accuracy. The primary outcome description explains how a match could be defined and how proportions were intended to be compared, but no observed proportions, item-level counts, uncertainty estimates, or inferential results are reported in the registry record. It is therefore not possible to say from this record whether identification was high or low, how the two app modes compared, or how either compared with the human rater.

The record also does not provide a numerical result for portion-size estimation, despite that subject appearing in the study title and purpose. Nor does it give results for usability or satisfaction. The primary endpoint concerns food identification, so it should not be treated as a complete measure of the app’s broader stated functions.

Even if a result were available, the design would define its scope. This was a single-group pilot using simulated meals, with participants not eating the food. Results from that context would describe performance on the tested task; they would not automatically generalize to routine use, other meal settings, different foods, or other populations. The registry record does not report an external validation study.

The endpoint is also dependent on the definition of an acceptable match. Because close matches could count, an estimate would refer to agreement under that particular rule rather than exact wording alone. Further details on match adjudication, missing assessments, and the handling of multiple items from the same participant are not reported in the registry record.

In short, the posted registry information documents the evaluation question and design, not the evidence needed to determine the result. The absence of a posted result should not be interpreted as proof of success, failure, or equivalence between methods.

13. Why This Trial Matters Statistically

This study raises a useful statistical distinction between evaluating an algorithm on a defined task and establishing how well it performs in wider use. The primary question concerns whether foods in simulated meals are correctly identified under a specified matching rule. That is a concrete, item-level question, but interpretation still depends on the tested foods, participants, classification rules, and analysis method.

The description includes both automated and user-corrected identification. Keeping these modes distinct matters statistically because the second incorporates human input. If their results were reported separately, readers could distinguish app-only classification from performance after correction. The registry record does not report those results, so the difference between modes cannot be evaluated.

The design also highlights a common issue in validation studies: many observations may come from a relatively small participant group. General statistical principle: multiple item assessments from one person are not necessarily independent. Appropriate uncertainty estimates depend on the data structure and should account for any clustering or repeated observations where relevant. The registry record does not describe this analysis choice.

A transparent report would pair accuracy estimates with their denominators, uncertainty, and a clear description of how matches and missing assessments were handled. It would also separate identification from portion estimation and report whether each task was evaluated against an appropriate criterion. Those details are not present in the registry record’s posted results, which lists zero analyses.

The pilot design can be informative about how to frame a validation question, but the registered description alone does not answer it. Its statistical importance lies in the need to distinguish a stated evaluation plan from a reported estimate, and to distinguish task-specific agreement from broader claims about performance.

14. Limitations

• The posted summary statistics are the only data these results rest on; patient-level data, full protocols and statistical analysis plans are not available.
• Recalculated results are simple, unadjusted calculations from posted summaries. They do not replace the sponsor's prespecified analysis.

The most direct limitation for interpreting the registry record is that it does not report numerical primary-outcome results or analyses. The endpoint and comparison approach are described, but the record provides no correct-identification counts, proportions, uncertainty estimates, or test results. Performance therefore cannot be quantified from the posted information.

The study used simulated meals in a controlled laboratory setting, and participants did not eat the food. This limits what the task can say about identification during ordinary meal logging or use in other environments. The record does not report validation beyond the laboratory exercise.

The primary outcome description focuses on food identification, even though the study title and purpose also refer to food portion size. A numerical portion-estimation outcome is not reported in the registry record. Usability and satisfaction are mentioned in the arm description, but their measures and findings are also not reported.

The item-level structure raises questions that cannot be resolved from the record. It reports 255 tested food items across participants but does not report how those items were distributed, whether all were assessed in every comparison, or how dependence among items from the same participant was handled. The record also does not report missing-data procedures, analysis populations, match adjudication details, or a complete statistical plan.

Finally, the single-group design does not provide a randomized treatment comparison. The human rater is described as a criterion method, not a randomized control arm. The record’s lack of posted analytical results means the precision and scope of any observed differences remain unknown.

15. Related Tutorials

16. Related Statistical Calculators

17. Sources

• ClinicalTrials.gov: Validity of an AI-based Program to Identify Foods and Estimate Food Portion Size — NCT05343585.
• Registry record last updated 2023-11-18; status last verified 2023-11.
• Results first posted 2023-11-18.
• Data retrieved from ClinicalTrials.gov on 2026-11-09.
• Group sizes, means, event counts and sponsor-reported p-values are reproduced as posted. Effect estimates, confidence intervals and p-values not attributed to the sponsor are independently calculated from the posted summary statistics.

18. Record Summary

NCT05343585 is a completed, single-group pilot study sponsored by Pennington Biomedical Research Center. It evaluated PortionSize AI in the Openfit app with adults completing simulated laboratory meals; participants did not eat during the task. The registry lists enrollment of 24, with a visit timeframe of approximately two hours.

The primary outcome concerns identification of foods and beverages using automated app output and output after user correction, compared with known identification and a human-rater criterion method. A close match could qualify under the stated matching rule. The record reports 255 food items tested across participants.

The registry lists four outcome measures and zero analyses. It does not report numerical results for food identification, portion estimation, usability, or satisfaction. The safety record lists 0 serious adverse-event participants affected, 0 other adverse-event participants affected, and 0 deaths. Allocation is listed as not applicable, masking as none, and the model as single group.