Releases: fhstp/IntellEvent
IntellEvent Multicenter Evaluation
🚀 Release: IntellEvent Multicenter Evaluation & Generalization
📌 Overview
This release introduces a comprehensive multicenter evaluation of the IntellEvent deep learning model for gait event detection in clinical gait analysis. The goal of this work is to validate robustness, generalization, and scalability across different laboratories, acquisition protocols, and patient populations.
The release includes:
- Training and evaluation pipelines for IntellEvent
- 6 independent benchmark datasets for evaluation of IntellEvent and future gait event detection algorithms
- Data extraction pipelines
- The IntellEvent and IntellEvent-Multicenter initial contact and foot off keras models
🧠 What was done
1. Large-Scale Multicenter Evaluation
We conducted the first large-scale multicenter study for deep learning-based gait event detection:
- 4 clinical gait laboratories
- 4,300+ patients
- 26,000+ gait trials
- Diverse orthopaedic and neurological pathologies
This allowed us to systematically assess how well IntellEvent generalizes across:
- Different labs
- Marker setups
- Sampling frequencies
- Patient populations
🏥 3. Independent Benchmark Laboratory (BL) Evaluation
To test real-world generalization, we evaluated on:
6 independent benchmark datasets (BL1-BL6):
- ~700+ additional participants
- Completely unseen pathologies:
- Stroke
- Parkinson’s disease
- Hip osteoarthritis
- ACL reconstruction
Models, Pipelines and independent benchmark datasets are available: IntellEvent.Multicenter.Evaluation.zip
Note: Independent Benchmark datasets differ from the original publications as only trials with ground truth events (e.g., events where valid force plate data is available) were extracted and used for evaluation.
Independent Benchmark Datasets Overview:
- BL1 [1] comprises participants who have survived a stroke (Stroke, n = 18) and healthy controls (HC, n = 130).
- BL2 [2] includes patients with hip osteoarthritis (HOA) before (HOA-pre, n = 87) and six months after (HOA-post, n = 85) total hip arthroplasty, along with healthy controls (HC, n = 80).
- BL3 [3] investigates gait characteristics in patients with Parkinson’s disease (PD) both on (PD-on, n = 22) and off (PD-off, n = 17) medication.
- BL4 [4] contains 50 healthy adults walking at five different speeds (0–0.4 m/s, 0.4–0.8 m/s, 0.8–1.2 m/s, self-selected, and self-selected fast).
- BL5 includes participants following anterior cruciate ligament (ACL) reconstruction, measured at 6 weeks (ACL-6w, n = 27) and 12 weeks (ACL-12w, n = 27) post-operatively, as well as healthy controls (HC, n = 29).
- BL6 comprises individuals with cerebral palsy (CP, n = 56), foot deformities (FD, n = 34), and other diagnoses (OT, n = 55).
[1] T. Van Criekinge, W. Saeys, S. Truijen, L. Vereeck, L. H. Sloot, and A. Hallemans, “A full-body motion capture gait dataset of 138 able-bodied adults across the life span and 50 stroke survivors,” Sci Data, vol. 10, no. 1, p. 852, Dec. 2023, doi: 10.1038/s41597-023-02767-y.
[2] A. Bertaux et al., “Gait analysis dataset of healthy volunteers and patients before and 6 months after total hip arthroplasty,” Sci Data, vol. 9, no. 1, p. 399, Jul. 2022, doi: 10.1038/s41597-022-01483-3.
[3] T. K. F. Shida et al., “A public data set of walking full-body kinematics and kinetics in individuals with Parkinson’s disease,” Front. Neurosci., vol. 17, p. 992585, Feb. 2023, doi: 10.3389/fnins.2023.992585.
[4] C. Schreiber and F. Moissenet, “A multimodal dataset of human gait at different walking speeds established on injury-free adult participants,” Sci Data, vol. 6, no. 1, p. 111, Jul. 2019, doi: 10.1038/s41597-019-0124-4.
BL5: University of Applied Sciences St. Pölten
BL6: Heidelberg University Hospital
v2.0 IntellEvent
v1.0 - IntellEvent Paper Release
This release represents the original IntellEvent created in this paper with the underlying source code and Vicon Nexus pipelines.