When Digital Health Falls Short: Limited Validity of an AI-Powered App for Dietary Assessment in Females with Obesity

Artificial intelligence (AI) and computer vision are rapidly transforming dietary assessment, yet few tools have been clinically validated against physiological reference methods. We evaluated the real-world validity of SNAQ, an AI-powered mobile application that estimates energy intake from food images, against doubly labelled water (DLW) in female adults with obesity. In this validation study, adult females with BMI ≥ 30 kg/m² were recruited at the University Hospital Zurich between September 2020 and January 2023. Participants completed a 7-day protocol including baseline and follow-up visits. Total daily energy expenditure (TDEE) was measured by DLW and total daily energy intake (TDEI) by SNAQ and a 24-hour dietary recall (24HR). Agreement, bias, and day-to-day reliability were assessed using Bland-Altman analysis, intraclass correlation, and Goldberg cut-offs. Twenty participants completed the study (age = 37.9 ± 13.4 years; BMI = 39.3 ± 5.4 kg/m²). Compared with DLW-derived TDEE (3004 ± 481 kcal day⁻¹), SNAQ underestimated energy intake by 25% (2187 ± 1401 kcal day⁻¹; p = 0.01) and 24HR by 50% (1464 ± 554 kcal day⁻¹; p < 0.0001). Individual variability was high (limits of agreement − 3707 to 2073 kcal day⁻¹) and within-subject reliability was negligible (ICC = 0.00). No significant associations were observed with body composition or weight change. Despite its advanced AI architecture, SNAQ showed poor accuracy and reproducibility in free-living adults with obesity. These findings illustrate the translational gap between algorithmic performance and clinical feasibility and highlight the need for standardized validation frameworks to ensure the safe integration of AI-powered nutrition tools into healthcare practice.