MCDCD: Multi-Source Unsupervised Domain Adaptation for Abnormal Human Gait Detection.

Area of Science:

• Biomedical Engineering
• Machine Learning
• Gait Analysis

Background:

• Gait analysis faces challenges in detecting subtle abnormalities due to high inter-subject variability in biometric traits and movement.
• Existing methods often suffer from limited detection accuracy and poor generalizability across different subjects.
• Unsupervised Domain Adaptation (UDA) offers a potential solution by leveraging labeled data from source domains to improve performance on unlabeled target domains.

Purpose of the Study:

• To propose a novel deep multi-source Unsupervised Domain Adaptation (UDA) approach, Maximum Cross-Domain Classifier Discrepancy (MCDCD), for enhanced gait abnormality detection.
• To improve classification performance on unseen subjects (target domain) by effectively utilizing information from multiple labeled subjects (source domains).
• To minimize domain shift between multiple source domains and the target domain for robust gait analysis.

Main Methods:

• Developed a deep UDA model (MCDCD) comprising a shared feature extractor and domain-specific classifiers for each source domain.
• The feature extractor learns discriminative gait representations.
• Minimization of cross-entropy loss for source sample classification and maximization of a novel cross-domain discrepancy loss between classifiers to reduce domain shift.

Main Results:

• Collected high-quality Motion capture (Mocap) and noisy Electromyography (EMG) data from eighteen subjects exhibiting normal and imitated abnormal gaits.
• The proposed MCDCD approach demonstrated superior performance in classifying abnormal gaits compared to baseline deep models.
• MCDCD outperformed state-of-the-art UDA methods across both Mocap and EMG data modalities.

Conclusions:

• The MCDCD approach effectively addresses the challenge of inter-subject variability in gait analysis.
• This novel deep multi-source UDA method significantly enhances the accuracy and generalizability of abnormal gait detection on novel subjects.
• MCDCD shows promise for real-world applications requiring robust gait abnormality identification from diverse data sources.