Summary:
In adults participating in the He Rourou Whai Painga randomized controlled dietary trial, a smartphone-based dietary assessment application using computer vision to estimate energy intake from food images over 7 days showed energy intake estimates comparable to validated 24-hour dietary recall methods (mean bias 783 kJ, p=0.33, not significant) compared to two 24-hour dietary recalls and estimated energy expenditure measured by indirect calorimetry and accelerometry, though it was associated with underestimation compared to objectively measured energy expenditure (mean bias -1814 kJ for app, -1715 kJ for recall), similar to the inherent underreporting seen with traditional methods.
| PICO | Description |
|---|---|
| Population | Adults participating in the He Rourou Whai Painga randomized controlled dietary trial in New Zealand. |
| Intervention | Smartphone-based dietary assessment application using computer vision artificial intelligence to estimate energy intake from food images over 7 days. |
| Comparison | Two 24-hour dietary recalls (EIrecall) and estimated energy expenditure (EE) measured by indirect calorimetry and wrist-worn accelerometry. |
| Outcome | Both app and 24-h recall underestimated energy intake vs EE (mean biases: -1814 kJ and -1715 kJ respectively). No significant difference between app and recall estimates (mean bias 783 kJ, p=0.33), indicating comparable performance. |
Clinical Context
Accurate dietary assessment is fundamental to nutrition research and clinical practice, yet traditional methods face significant challenges. Food frequency questionnaires require extensive recall, 24-hour dietary recalls are resource-intensive and subject to memory errors, and food diaries require high participant burden and often alter eating behavior. All self-report methods suffer from systematic underreporting, particularly of energy-dense foods and snacks.
Technology-assisted dietary assessment using smartphone cameras and artificial intelligence offers potential to reduce participant burden while maintaining or improving accuracy. Computer vision algorithms can identify foods from images and estimate portion sizes, theoretically providing objective assessment without relying on participant recall or manual data entry.
This pilot study evaluated an image-based dietary assessment application against two reference standards: the validated 24-hour dietary recall method (the current gold standard for dietary assessment) and objectively measured energy expenditure using indirect calorimetry and accelerometry (providing a biomarker of true energy intake in weight-stable individuals).
Clinical Pearls
1. App Performance Matches Traditional Methods: The non-significant difference between app-derived and recall-derived energy intake estimates (783 kJ bias, p=0.33) indicates the image-based approach performs comparably to established dietary assessment methodology. This validation supports potential clinical adoption.
2. Both Methods Underestimate True Intake: The approximately 1700-1800 kJ (400-430 kcal) underestimation compared to measured energy expenditure is consistent with the well-known underreporting phenomenon in dietary assessment. Importantly, the app did not perform worse than traditional methods in this regard.
3. Seven Days of Image Capture Is Feasible: Participants successfully used the app for 7 days, suggesting acceptable user burden. The passive nature of image capture (compared to detailed food logging) may improve long-term adherence for dietary monitoring.
4. Objective Validation Reveals Inherent Limitations: Using energy expenditure as an objective reference standard provides insights that comparison to other self-report methods cannot. Both assessment approaches miss a similar proportion of actual intake, suggesting the underreporting problem relates to participant behavior (forgetting to photograph, eating without recording) rather than technology limitations.
Practical Application
For clinical practice, image-based dietary assessment apps can be considered as alternatives to food diaries or recalls for patients willing to photograph meals. The reduced burden compared to detailed logging may improve adherence for dietary monitoring in weight management or diabetes care. However, counsel patients that underreporting is inherent to all self-monitoring approaches.
When interpreting app-derived energy intake data, apply similar skepticism as with traditional dietary assessment—expect systematic underestimation of approximately 15-20%. Use data for relative comparisons (tracking change over time, comparing eating patterns across days) rather than absolute calorie counting.
For research applications, image-based assessment may reduce cost and participant burden in large studies while providing comparable data quality to 24-hour recalls. Consider combining with objective biomarkers when precise energy intake quantification is required.
Broader Evidence Context
This study contributes to the growing validation literature for AI-assisted dietary assessment. Previous studies have evaluated food identification accuracy and portion size estimation, but validation against both established methods and objective biomarkers provides more comprehensive evidence. The finding that image-based apps perform as well (but not better) than traditional methods in capturing true intake highlights that technological advances address convenience but not the fundamental behavioral challenge of complete dietary capture.
Future developments may include prompting systems to reduce forgotten photographs, integration with other data sources (purchases, restaurant check-ins), and improved algorithms for mixed dishes and cultural foods.
Study Limitations
Pilot study with limited sample size. The specific application and its AI algorithms may not represent all available products. Participants were enrolled in a dietary trial, potentially increasing motivation and attentiveness compared to general populations. The 7-day assessment period may not capture usual dietary patterns. Specific app usability and technical issues were not detailed.
Bottom Line
A computer vision-based dietary assessment app provides energy intake estimates comparable to validated 24-hour dietary recalls, though both methods underestimate true intake by approximately 1700-1800 kJ compared to objectively measured energy expenditure, supporting app use as a convenient alternative with similar accuracy limitations.
Source: Lachlan Lee, et al. “An automated image-based dietary assessment application: a pilot study.” Read article
