Tailoring LLMs for Multimodal Athlete Recovery Planning: Combining Wearable, Survey, and Training Schedule Data for Personalized Planning

The integration of multimodal athletic data, including wearable monitoring, self-reports, and training schedules presents a growing opportunity for machine learning to support individualized recovery decision-making. We present a feasibility study that evaluated the capability of large language model (LLM) pipelines to generate personalized athlete recovery recommendations using multimodal inputs. A prototype system (RecoveryGPT) was implemented across three technical environments: (1) Azure OpenAI with Heroku web deployment, (2) Azure AI Foundry, and (3) Microsoft Copilot Studio with trigger-based multimodal integration. All systems employed Retrieval-Augmented Generation (RAG) to ground responses to a Literature Corpus. We expanded the Copilot Studio implementation to incorporate user-triggered surveys, wearable data, and structured training schedules to deliver personalized, evidence-based guidance. Systems were evaluated using nine standardized recovery queries and fifteen simulated athlete case scenarios. Results indicated comparable literature grounding across platforms but revealed that Copilot Studio offers a more intuitive environment and a more capable deployment approach, making it the most operationally viable for embedding into existing team workflows like Microsoft Teams. This feasibility study offers a proof-of-concept pathway for integrating generative AI into high-performance recovery planning, with implications for use across other domains.