Agreement Between an Artificial Intelligence-Based Meal Image Recognition System and the Weighed Dietary Record for Estimating Energy and Nutrient Intakes.

Objectives

In Japan, smartphone applications are increasingly used for dietary recording in healthcare settings. This study aimed to examine the agreement between energy and nutrient intake estimates obtained using an artificial intelligence (AI)-based dietary recording application and those obtained using the weighed dietary record (WDR).

Methods

The AI-based dietary recording method (FoodLog Athl method) was compared with WDR. Thirty-six university students (35 women and 1 man) simultaneously recorded their dietary intake using FoodLog Athl (FLA) and the WDR for 10 consecutive days. Energy and nutrient intakes were estimated using each method, and correlations and agreement between the two methods were evaluated.

Results

Significant positive correlations were observed between the two methods for energy and most nutrients, except for iron, vitamin B1, and sodium chloride equivalent (p < 0.01). Compared with the WDR, the FLA method showed systematic overestimation of energy and major macronutrients (protein, fat, and carbohydrate) and underestimation of total dietary fiber. Bland-Altman analysis indicated fixed bias and relatively wide limits of agreement for several nutrients.

Conclusions

The FLA method demonstrated moderate agreement with the WDR, with systematic bias observed for selected nutrients. These findings suggest that the application may be useful for monitoring overall dietary trends or relative intake over time, but caution is warranted when precise individual-level nutrient quantification is required. Professional review by registered dietitians may help improve estimation accuracy and reduce bias.