Dataset for publication: "Exploring Large Language Models for Automated Gait Analysis"

This project has received funding from the European Union’s HorizonEurope research and innovation programme under grant agreement No[SHARESPACE, 101092889] as well as the BMBF project OrthoSuPer(project ID 13GW0564A-F and 13GW0564E). TITLE Project OrthoSuPer at German Research Center of Artificial Intelligence Kaiserslautern, Germany Related Paper: Generative AI for Health: Exploring Large Language Models for Automated Gait Analysis (incoming link: ) License: Open Data Commons Attribution License (ODC-BY 1.0)Link: http://opendatacommons.org/licenses/by/1.0/Please cite the corresponding Paper (incoming cite: ) SUMMARY This dataset accompanies a proof-of-concept study on integrating large language models (LLMs), specifically GPT-4o, into clinical gait analysis for patients undergoing total knee arthroplasty (TKA). We collected spatiotemporal motion data from 11 preoperative TKA patients and to establish ground-truth clinical descriptions, 41 expert practitioners reviewed the walks in independent pairs. We designed the prompt for GPT-4o to generate automated gait analysis due to prompt engineering techniques. VERSION & DATE Version: 1.0 Release date: 14.08.2025 CONTENTS -Walking trajectories as csv file -Gait description provided by physiotherapists -Prompt -Skeleton videos (.MP4) generated by SIMI MOTION, http://www.simi.com/de/home.html -Description of patients as csv file CONTACT Maintainer: Rebecca Keilhauer Organization: DFKI Kaiserslautern, Germany Email: [e-mail], [e-mail] ------------------------------------------------------------------------DATA COLLECTION------------------------------------------------------------------------ WALKING SEQUENCES: SETTING Location: Orthopädisches Krankenhaus Schloss Werneck, Germany Conditions: 8 view camera setup for markerless tracking, Patients wear normal clothing, Patients undergoing total knee arthroplasty PARTICIPANTS N: 11 Demographics: average age: 68.09 years, standard deviation: 5.13 years; gender distribution: 4 female, 7 male HARDWARE & SOFTWARE provided by SIMI Motion, http://www.simi.com/de/home.html PROTOCOL Tasks: 6 minute walking test Trial: Each recording session consisted of nine repetitions of a walking task, wherein the patient walked in a straight line (3meters), turned, and walked back to the starting point. Only the forward walking sequences were retained for analysis. SAMPLING Sampling rates: 80 Hz FILE FORMATS: CSV, Joint angles in degrees were captured in all three planes: [flexion/extension, abduction/adduction, and internal/external rotation] JOINTS & MARKER Ankle, right Knee, right Hip, right Ankle, left Knee, left Hip, left Pelvis Torso Skullbase Wrist, right Elbow, right Shoulder, right Wrist, left Elbow, left Shoulder, left Heel right Foot lateral right Foot tip right Heel left Foot lateral left Foot tip left Ear left Ear right Nose Eye left Eye right Heel right Index finger base right Little finger base right Index finger base left Little finger base left EXPERT GAIT-ANALYSIS PROTOCOL EXPERT PANEL N experts: 41 (primarily physiotherapists) Years of professional experience: - Mean: 13.28 - Standard deviation: 10.12 ANNOTATION UNITS - Per patient/per trial, each view (posterior / left lateral / right lateral) - Output types: clinical descriptions (transcripts) sorted in bullet points SOURCE MATERIAL SHOWN TO EXPERTS - Privacy-preserving avatar and skeletal renderings only (see Fig. 1 in the paper for examples). TRANSCRIPTS - Each session (two experts) was audio-recorded and transcribed. - Transcripts are anonymized and contain no personal identifiers. - summarized in bullet points Session flow (per pair): 1) Expert A provides a free-form clinical interpretation of the gait. 2) Expert B analyzes, reflects, and challenges/clarifies A’s description. 3) Audio of the session is recorded and later transcribed. Personal identifiers are not included in transcripts. FILE FORMATS: TXT ------------------------------------------------------------------------DATA DESCRIPTION------------------------------------------------------------------------ LABELS & ANNOTATIONS keypoint format: SIMI keypoint definition, see http://www.simi.com/de/home.html Annotations: Gait Analysis collected by interviews with Physiotherapy ------------------------------------------------------------------------FOLDER STRUCTURE------------------------------------------------------------------------ Root/ README.txt LICENSE.txt Prompt.txt Patient_desciption.csv Data/ Patient<PatientID>/ Patient<PatientID>_<TimeStep>_min_walking_out_<trialID>/ joint centers.csv joint angles.csv Marker_position.csv Gait_Analysis_P<PatientID>.txt P<PatientID>.mp4 Patient<PatientID + 1>/ ... ... ------------------------------------------------------------------------QUALITY & LIMITATIONS------------------------------------------------------------------------ KNOWN LIMITATIONS - Experts viewed only avatar/skeletal renderings (no raw videos), which can omit subtle appearance cues. - Inter-rater reliability ------------------------------------------------------------------------ETHICS & RIGHTS------------------------------------------------------------------------PRIVACY & ANONYMITY - only privacy-preserving avatar/skeletal renderings were shown. - Expert transcripts are anonymized. - recordings were approved by the ethical commitees of German Research Center of Artificial Intelligence