Validity of an Artificial Intelligence-Based Application to Identify Foods and Estimate Energy Intake Among Adults: A Pilot Study.
Curr Dev Nutr 2023;
7:102009. [PMID:
38026571 PMCID:
PMC10656219 DOI:
10.1016/j.cdnut.2023.102009]
[Citation(s) in RCA: 0] [Impact Index Per Article: 0] [Reference Citation Analysis] [Abstract] [Key Words] [Track Full Text] [Figures] [Journal Information] [Subscribe] [Scholar Register] [Received: 05/03/2023] [Revised: 09/12/2023] [Accepted: 09/18/2023] [Indexed: 12/01/2023] Open
Abstract
Background
The commercial application Openfit allows for automatic identification and quantification of food intake through short video capture without a physical reference marker. There are no known peer-reviewed publications on the validity of this Nutrition Artificial Intelligence (AI).
Objectives
To test the validity of Openfit to identify food automatically and semiautomatically (with user correction), test the validity of Openfit at quantifying energy intake (kcal) automatically and semiautomatically, and assess satisfaction and usability of Openfit.
Methods
During a laboratory-based visit, adults (7 male and 17 female), used Openfit to automatically and semiautomatically record provided meals, which were covertly weighed. Foods logged were identified as an "exact match," "far match," or an "intrusion" using Food and Nutrient Database for Dietary Studies (FNDDS) codes. Descriptive data were stratified by meal, food item, and FNDDS group, and presented with or without beverages. Bland-Altman analyses assessed errors over levels of energy intake. Participants completed a User Satisfaction Survey (USS) and the Computer Systems Usability Questionnaire (CSUQ). Open-ended questions were assessed with qualitative methods.
Results
Exact matches, far matches, and intrusions were 46%, 41%, and 13% for automated identification, and 87%, 23%, and 0% for semiautomated identification, respectively. Error for automated and semiautomated energy estimates were 43% and 33% with beverages, and 16% and 42% without beverages. Bland-Altman analyses indicated larger error for higher energy meals. Overall mean scores were 2.4 for the CSUQ and subscale means scores ranged from 4.1 to 5.5. for the USS. Participants recommended improvements to Openfit's Nutrition AI, manual estimation, and overall app.
Conclusion
Openfit worked relatively well for automatically and semiautomatically identifying foods. Error in automated energy estimates was relatively high; however, after excluding beverages, error was relatively low (16%). For semiautomated energy estimates, error was comparable to previous studies. Improvements to the Nutrition AI, manual estimation and overall application may increase Openfit's usability and validity.This trial was registered at clinicaltrials.gov as NCT05343585.
Collapse