Relative Validity Of Generative Artificial Intelligence Versus Weighed Food Record For Quantification Of Caloric Intake

Relative Validity of Generative Artificial Intelligence Versus Weighed Food Record for Quantification of Caloric Intake Hunter Smathers, Wyatt, Baggett, Robyn York, and Jarrett WalboltMontreat College Developing accurate and efficient methods of estimating caloric intake may result in improved health outcomes. Generative artificial intelligence (AI) is becoming more accessible to the public and may be a helpful way for people to estimate nutritional intake. PURPOSE: The purpose of this study was to assess the accuracy of publicly available AI for estimating caloric intake. METHODS: Dietary consumption (n = 39) was analyzed using both weighed food record (WFR) and generative AI. Meals were weighed using a food scale and then caloric intake calculated using the USDA nutrition database. The same meals were also described to generative AI (ChatGPT 3.5, Open AI) based off the program’s prompts to receive a caloric estimation for the meal. The differences between the WFR versus AI were assessed by a paired t-test on SPSS 28. Agreement between the two methods was assessed using Bland-Altman analysis. Linear regression analysis was used to assess proportional bias between WFR and AI. RESULTS: Paired t-test showed no significant difference between WFR and AI (t = 1.59, p = 0.120, = 0.06). Bland-Altman analysis showed a mean difference between estimates of caloric intake were -373 g to 484 g. Linear regression demonstrated there is no significant proportional bias (R2 = 0.024, p = 0.343). CONCLUSION: This analysis supports the possible use of generative AI as an accessible assessment tool for estimating caloric intake. Users will have to determine the acceptable limits of agreement for their usage and goals. As AI tools improve and validity is established, AI may be useful for approachable quantification of nutritional intake.