Qualitative evaluation of artificial intelligence-generated weight management diet plans

Abstract

Importance: The transformative potential of artificial intelligence (AI), particularly via large language models, is increasingly being manifested in healthcare. Dietary interventions are foundational to weight management efforts, but whether AI techniques are presently capable of generating clinically applicable diet plans has not been evaluated. Objective: Our study sought to evaluate the potential of personalized AI-generated weight-loss diet plans for clinical applications by employing a survey-based assessment conducted by experts in the fields of obesity medicine and clinical nutrition. Design, setting, and participants: We utilized ChatGPT (4.0) to create weight-loss diet plans and selected two control diet plans from tertiary medical centers for comparison. Dietitians, physicians, and nurse practitioners specializing in obesity medicine or nutrition were invited to provide feedback on the AI-generated plans. Each plan was assessed blindly based on its effectiveness, balanced-ness, comprehensiveness, flexibility, and applicability. Personalized plans for hypothetical patients with specific health conditions were also evaluated. Main outcomes and measures: The primary outcomes measured included the indistinguishability of the AI diet plan from human-created plans, and the potential of personalized AI-generated diet plans for real-world clinical applications. Results: Of 95 participants, 67 completed the survey and were included in the final analysis. No significant differences were found among the three weight-loss diet plans in any evaluation category. Among the 14 experts who believed that they could identify the AI plan, only five did so correctly. In an evaluation involving 57 experts, the AI-generated personalized weight-loss diet plan was assessed, with scores above neutral for all evaluation variables. Several limitations, of the AI-generated plans were highlighted, including conflicting dietary considerations, lack of affordability, and insufficient specificity in recommendations, such as exact portion sizes. These limitations suggest that refining inputs could enhance the quality and applicability of AI-generated diet plans. Conclusion: Despite certain limitations, our study highlights the potential of AI-generated diet plans for clinical applications. AI-generated dietary plans were frequently indistinguishable from diet plans widely used at major tertiary medical centers. Although further refinement and prospective studies are needed, these findings illustrate the potential of AI in advancing personalized weight-centric care.