AI-FMS: System Development, Phase I Evaluation, and Human-AI Collaboration in FMS Video Review

Abstract

Preprint. Not peer reviewed. Functional Movement Screen (FMS) assessment depends on trained human observation, yet practical video review can be constrained by transient viewing, remote access, repeated manual navigation, qualitative judgments, and the compression of movement into an ordinal 0-3 score. We developed AI-FMS, a human-in-the-loop video-review system with implemented functionality for all seven FMS movements, including upload, repetition segmentation, looped playback, reviewer scoring, structured protocol fields, pose overlays, quantitative features, explainable first-pass score suggestions, quality warnings, abstention, adjudication, and traceable export. The system uses two-dimensional pose landmarks to preserve movement evidence such as joint angles, relative distances, side-specific trajectories, and full-cycle events while retaining human authority over protocol, pain, clearing conditions, and final scores. Building on the seven-movement system, Phase I quantitative research focused on Deep Squat, Hurdle Step, Active Straight-Leg Raise, and Rotary Stability. The corpus contained 28 unique source videos and 110 canonical repetitions. A balanced subset of 32 repetitions, eight per movement, was reviewed by two reviewers in two independent blinded rounds. Both rounds concealed AI scores, pose evidence, previous ratings, and the other reviewer's results. Human rating consistency and AI-human score comparison formed one part of the study: among 25 comparable items, the locked AI matched the human reference exactly on 16 and was within one point on 23. These are internal exploratory findings, not independent external validation. The structured movement evidence also enabled further analyses: a Deep Squat depth-and-joint-strategy continuum, ASLR bilateral repeatability, different control and deduction pathways within the same Hurdle Step score, and Rotary Stability full-cycle event structures. These analyses preserve continuous differences, bilateral relationships, and temporal information beyond the 0-3 score, providing traceable evidence for movement explanation, cross-repetition comparison, and future research. AI-FMS thus combines seven-movement system development, Phase I quantitative research, and movement analyses enabled by the system's data. Automation organizes evidence and offers inspectable suggestions while human reviewers retain final interpretation and scoring authority. The system is not intended for clinical diagnosis or injury prediction. Source code and project documentation: https://github.com/edwardzhu-HK/AI-FMS.