LLM-FMS: A fine-grained dataset for functional movement screen action quality assessment

Abstract

The Functional Movement Screen (FMS) is a critical tool for assessing an individual’s basic motor abilities, aiming to prevent sports injuries. However, current automated FMS evaluation is based on deep learning methods, and the evaluation of actions is limited to rank scoring, which lacks fine-grained feedback suggestions and has poor interpretability. This limitation prevents the effective application of automated FMS evaluation for injury prevention and rehabilitation. We develop a fine-grained, hierarchical FMS dataset, LLM-FMS, derived from FMS videos and enriched with detailed, hierarchical action annotations. This dataset comprises 1812 action keyframe images from 45 subjects, encompassing 15 action representations of seven FMS actions. Each action includes a score, scoring criteria, and weight data for body parts. To our extensive knowledge, LLM-FMS is the first fine-grained fitness action dataset for action evaluation task. Additionally, a novel framework for action quality assessment based on large language models (LLMs) is proposed, designed to enhance the interpretability of FMS evaluations. Our method integrates expert rules, utilizes RTMPose to extract key skeletal-level action features from key frames, and inputs prompts into the LLM, enabling it to infer scores and provide detailed rationales. Experimental results demonstrate that our approach significantly outperforms existing methods while offering superior interpretability. Experimental results demonstrate that our approach outperforms existing methods in terms of accuracy and interpretability, with a substantial increase in the clarity and detail of the rationales provided. These findings highlight the potential of our framework for fine-grained action quality assessment with the aid of LLMs.