Assessment of Recommendations Provided to Athletes Regarding Sleep Education by GPT-4o and Google Gemini: Comparative Evaluation Study.

Abstract

Background: Inadequate sleep is prevalent among athletes, affecting adaptation to training and performance. While education on factors influencing sleep can improve sleep behaviors, large language models (LLMs) may offer a scalable approach to provide sleep education to athletes.

Objective: This study aims (1) to investigate the quality of sleep recommendations generated by publicly available LLMs, as evaluated by experienced raters, and (2) to determine whether evaluation results vary with information input granularity.

Methods: Two prompts with differing information input granularity (low and high) were created for 2 use cases and inserted into ChatGPT-4o (GPT-4o) and Google Gemini, resulting in 8 different recommendations. Experienced raters (n=13) evaluated the recommendations on a 1-5 Likert scale, based on 10 sleep criteria derived from recent literature. A Friedman test with Bonferroni correction was performed to test for significant differences in all rated items between the training plans. Significance level was set to P<.05. Fleiss κ was calculated to assess interrater reliability.

Results: The overall interrater reliability using Fleiss κ indicated a fair agreement of 0.280 (range between 0.183 and 0.296). The highest summary rating was achieved by GPT-4o using high input information granularity, with 8 ratings >3 (tendency toward good), 3 ratings equal to 3 (neutral), and 2 ratings <3 (tendency toward bad). GPT-4o outperformed Google Gemini in 9 of 10 criteria (P<.001 to P=.04). Recommendations generated with high input granularity received significantly higher ratings than those with low granularity across both LLMs and use cases (P<.001 to P=.049). High input granularity leads to significantly higher ratings in items pertaining to the used scientific sources (P<.001), irrespective of the analyzed LLM.

Conclusions: Both LLMs exhibit limitations, neglecting vital criteria of sleep education. Sleep recommendations by GPT-4o and Google Gemini were evaluated as suboptimal, with GPT-4o achieving higher overall ratings. However, both LLMs demonstrated improved recommendations with higher information input granularity, emphasizing the need for specificity and a thorough review of outputs to securely implement artificial intelligence technologies into sleep education.