Evaluation Strategies for Large Language Model-Based Models in Exercise and Health Coaching: Scoping Review

Large language model (LLM)-based artificial intelligence (AI) coaches show promise for personalized exercise and health interventions. However, the unique demands of ensuring safety and real-time, multimodal personalized feedback have created a fragmented evaluation landscape lacking standardized frameworks.

This scoping review systematically maps current evaluation strategies for LLM-based AI coaches in exercise and health, identifies strengths and limitations, and proposes directions for robust, standardized validation.

Following PRISMA-ScR (Preferred Reporting Items for Systematic reviews and Meta-Analyses extension for Scoping Reviews) guidelines, we conducted a systematic search across 6 major databases (eg, PubMed, Web of Science) for original research on LLM-based exercise and health coaching. Studies were included if they explicitly reported on evaluation methods. We extracted and synthesized data on model types, application domains, and evaluation strategies and developed a 5-point Evaluation Rigor Score (ERS) to quantitatively assess the methodological depth of the evaluation designs.

We included 20 studies, most using proprietary models like ChatGPT (75%). Evaluation strategies were highly heterogeneous, mixing human ratings (80%) and automated metrics (40%). Crucially, the evidence was limited by low methodological rigor: the median ERS was 2.5 out of 5, with 55% of studies classified as having low rigor. Key gaps included limited use of real-world data (40%) and inconsistent reliability reporting (45%).

The current evaluation of LLM-based health coaches is fragmented and methodologically weak. Future work must establish multidimensional validation frameworks that integrate technical benchmarks with human-centered methods to ensure safe, effective, and equitable deployment.