Toward intelligent clinical support for personalized sport training rehabilitation via large language models

Effective sport-training rehabilitation demands exercise prescriptions that adapt to each patient's changing symptoms and adherence patterns, yet most recommender systems rely on either numerical logs or handcrafted rules, failing to exploit the rich information embedded in free-text feedback. How can we unify linguistic self-reports with temporal behaviour to produce more accurate, interpretable training recommendations under real-world data sparsity? We introduce ReLite, a three-stage architecture that (i) encodes exercise reviews with a fine-tuned large language model, (ii) refines session histories via a lightweight Transformer, and (iii) aligns semantic embeddings with ordinal ratings through multi-head cross-attention before a dual-head MLP outputs continuous and ordinal tolerance scores. We conduct a series of experiments on two public dataset. The experimental results demonstrate that integrating large-scale language understanding with task-specific sequence modelling and adaptive text-rating alignment yields a robust, data-efficient foundation for intelligent clinical support in personalised sport-training rehabilitation.