Meta-analysis

Photo-based vs. manual-entry dietary assessment: a meta-analysis of accuracy differentials

DAI-MA-2026-03

Helena Weiss, PhD, MPH, RD; Daniel Okafor, PhD, MS; Lars Henriksen, PhD
Published March 18, 2026 · DOI: 10.5281/zenodo.dai-2026-03

1. Background

Photo-based dietary assessment — in which the user takes a photograph of a meal and a software pipeline estimates energy and macronutrient content — has proliferated alongside manual-entry applications, in which the user types or selects items from a food database. The two modalities have different error structures: photo-based estimation depends on ingredient identification and portion estimation from visual features; manual-entry depends on user recall, user diligence, and database quality. Neither has been established as categorically superior.

The Initiative’s 2025 systematic review (DAI-SR-2025-06) pooled independent per-meal MAPE across 31 studies at 18.7% without differentiating modality. The narrative review on mixed-dish portion error (DAI-NR-2025-08) argued that the single-item vs. mixed-dish distinction is more consequential than the photo vs. manual distinction in many consumer contexts. The narrative review of manual-entry database quality (DAI-NR-2026-02) documented substantial heterogeneity in manual-entry performance driven by database provenance.

Against this background, the present meta-analysis asks: across studies in which both modalities have been evaluated against a common reference, is there evidence for a consistent accuracy differential?

2. Methods

2.1 Protocol

The meta-analysis was pre-registered on 7 October 2025 (DAI-2026-03-PROT). It extends the Initiative’s 2025 systematic review (DAI-SR-2025-06) by focusing on studies with paired or otherwise comparable evaluation of both modalities. PRISMA 2020 guidance was followed.

2.2 Eligibility

Studies were eligible if they reported independent (non-vendor) per-meal or per-day MAPE for at least one photo-based and at least one manual-entry application within the same study design, against the same reference standard. Studies without extractable per-metric variance were included in narrative synthesis only.

2.3 Search and extraction

The 47 studies from the 2025 systematic review were screened for modality comparability; additional searches were run on PubMed, EMBASE, IEEE Xplore, and Google Scholar through 15 September 2025. Two reviewers extracted paired MAPE figures and study-level moderators.

2.4 Analysis

Random-effects meta-analysis (DerSimonian-Laird) pooled the difference in per-meal MAPE (photo − manual) across studies. Heterogeneity was summarised using I² and τ². Pre-specified sub-group analyses examined: publication year (≤2022; 2023 onward), reference standard, cuisine family predominance, and mixed-dish proportion of the evaluation set.

3. Results

3.1 Included studies

Of 47 screened studies, 34 contributed extractable paired data (4,812 meals; 1,287 participants). Five studies were excluded because their photo and manual conditions used different reference standards; four were excluded because variance estimates were not reconstructible.

3.2 Pooled differential

The pooled difference in per-meal MAPE on energy (photo − manual) was +1.8 percentage points (95% CI +0.3 to +3.3; I² = 79.1%; τ² = 4.2; k = 34). A positive value indicates manual-entry MAPE was lower (better) on average.

3.3 Heterogeneity

Between-study heterogeneity is substantial (I² = 79.1% overall; remained large within every sub-group except the duplicate-meal reference, which had too few studies to estimate precisely). Moderator regression identified publication year, mixed-dish proportion, and reference standard as predictors (all p < 0.05). Unexplained heterogeneity remained after moderator adjustment.

3.4 Sensitivity analyses

Leaving out the three largest studies did not materially alter the pooled estimate (Δ shifted from +1.8 to +2.1 pp). Restricting to studies at low risk of bias (n = 11 per the 2025 systematic review’s QUADAS-2 assessment) produced a pooled Δ of −0.8 pp (95% CI −3.2 to +1.6), not statistically distinguishable from zero.

4. Discussion

The headline finding is that the pooled differential between photo-based and manual-entry dietary assessment is small in magnitude and heterogeneous in direction, with the direction itself reversing between studies published before and after 2023. The earlier literature favours manual entry; the more recent literature favours photo-based, at least for studies using current-generation pipelines. The reversal is consistent with meaningful improvement in photo-based pipelines over the past several years, but the heterogeneity within each time-stratum suggests that any specific tool’s performance cannot be inferred from the category-level pooled estimate.

Three patterns warrant attention. First, mixed-dish-dominant evaluation sets consistently favour manual entry, consistent with the Initiative’s narrative review (DAI-NR-2025-08) that mixed-dish portion estimation is the current frontier of difficulty for photo-based pipelines. Second, the heterogeneity among recent studies is itself a finding: the category is not uniform, and a meaningful minority of recent photo-based tools report performance materially below the category’s pooled estimate. Independent replication on shared evaluation sets is the appropriate next step. Third, the low-risk-of-bias subset, while small, is not inconsistent with a null differential; the quality of the underlying evidence, not the point estimate, may be the rate-limiting consideration.

Limitations of this meta-analysis include its restriction to English-language studies, the heterogeneity of the included studies, the partial overlap with the 2025 systematic review (which it is intentionally built on), and the small number of studies in some sub-groups. The meta-analysis does not rank individual applications; between-study heterogeneity is too high to support such a ranking from the current evidence.

5. Conclusions

Photo-based dietary assessment has potential to outperform manual entry when the photo pipeline is well-trained and when the meal is amenable to image-based estimation, but the current literature does not support a category-level ordering. Mixed-dish meals remain a persistent weakness for photo-based modalities. The heterogeneity among current results is the central finding, and continued independent validation on shared evaluation sets — applying the reference-construction protocol in DAI-MP-2025-07 and the reporting template in DAI-SR-2025-06 — is the appropriate programme of work for the coming period. Consumers and clinicians should treat category-level claims with appropriate caution and rely on tool-specific, independently validated accuracy figures where available.

References

1. Amin T, Nakamura Y. Head-to-head comparison of three dietary tracking modalities. JMIR mHealth Uhealth. 2022;10(7):e36729.
2. Bittner V, Weiss W. Photo-based calorie estimation in controlled cafeteria settings. Nutrients. 2024;16(9):1327.
3. Cheung L, Park Y. Manual-entry versus image-based tracking in weight-management trials. Obesity. 2023;31(4):1021-1030.
4. DerSimonian R, Laird N. Meta-analysis in clinical trials. Control Clin Trials. 1986;7(3):177-188.
5. Eriksson T, Chen H. Recent-generation photo-based dietary pipelines: a benchmark comparison. Comput Biol Med. 2024;172:108212.
6. Ferrer M, Oliveira P. Modality effects in digital dietary assessment. Br J Nutr. 2023;130(11):1897-1909.
7. Grieve A, Maler J. Risk-of-bias subset analysis in method-comparison meta-analyses. Stat Methods Med Res. 2023;32(6):1067-1079.
8. Higgins J, Thompson S. Quantifying heterogeneity in a meta-analysis. Stat Med. 2002;21(11):1539-1558.
9. Initiative of Dietary Assessment. Independent validation of image-based dietary assessment applications. DAI-SR-2025-06. 2025.
10. Initiative of Dietary Assessment. Mixed-dish portion estimation. DAI-NR-2025-08. 2025.
11. Initiative of Dietary Assessment. Manual-entry food databases: a quality and provenance audit. DAI-NR-2026-02. 2026.
12. Initiative of Dietary Assessment. A protocol for weighed-food reference meal construction. DAI-MP-2025-07. 2025.
13. Iwata S, Kwon D. Mixed-dish benchmark performance of photo-based applications. Nutrients. 2024;16(15):2478.
14. Jimenez R, Holst M. Reference-standard effects in dietary validation meta-analyses. Public Health Nutr. 2023;26(12):2931-2940.
15. Pustejovsky J, Tipton E. Small-sample methods for cluster-robust variance estimation. J Bus Econ Stat. 2018;36(4):672-683.