Classification of knee osteoarthritis severity using markerless motion capture and long short-term memory fully...

Area of Science:

• Biomedical Engineering
• Orthopedics
• Artificial Intelligence in Healthcare

Background:

• Knee osteoarthritis (OA) diagnosis and severity grading traditionally rely on clinical assessments and imaging, which can be subjective and resource-intensive.
• Gait analysis offers a non-invasive method to assess functional impairment in knee OA, reflecting disease-related biomechanical changes.

Purpose of the Study:

• To investigate the efficacy of integrating markerless motion capture with deep learning for automated knee OA severity classification based on gait kinematics.
• To compare the performance of a Long Short-Term Memory Fully Convolutional Network (LSTM-FCN) model using random versus subject-based data splitting strategies.

Main Methods:

• Utilized markerless motion capture to collect gait data from individuals with varying knee OA severity.
• Employed an LSTM-FCN deep learning model to analyze gait patterns and classify severity according to Kellgren-Lawrence grades.
• Evaluated model generalizability using two data-splitting strategies: random split and subject-based split.

Main Results:

• The LSTM-FCN model achieved high accuracy (0.91) with random data splitting.
• Classification performance decreased significantly (accuracy 0.76) with subject-based splitting, highlighting challenges in inter-subject generalizability.
• The model effectively distinguished between severe OA and healthy gait patterns, but showed higher misclassification rates for early and moderate OA due to overlapping gait characteristics.

Conclusions:

• Deep learning analysis of gait kinematics presents a promising, scalable, and accessible alternative for automated knee OA severity classification.
• Inter-subject variability in gait patterns poses a significant challenge to model generalizability, necessitating further research into advanced feature extraction and multimodal data integration.
• Future studies should explore longitudinal data to evaluate the predictive potential of these models for disease progression and treatment efficacy.