Domain-Adaptive Fall Detection Using Deep Adversarial Training.

Area of Science:

• Assistive technologies
• Machine learning
• Healthcare

Background:

• Fall detection (FD) systems are crucial for alerting caregivers to emergencies.
• Acquiring large, diverse datasets for training accurate FD models is challenging.
• Existing machine learning models struggle with performance drops due to domain mismatch (e.g., different sensor positions or configurations).

Purpose of the Study:

• To propose a domain-adaptive fall detection (DAFD) method using deep adversarial training (DAT).
• To enable knowledge transfer from source domains to target domains in FD systems.
• To address cross-domain challenges like varying sensor positions and configurations.

Main Methods:

• Implemented deep adversarial training (DAT) for domain adaptation.
• Developed a DAFD model to minimize domain discrepancy between source and target data.
• Evaluated DAFD performance against conventional FD models in cross-domain scenarios.

Main Results:

• DAFD achieved an average F1-score improvement of 1.5% to 7% in cross-position scenarios.
• DAFD showed an average F1-score improvement of 3.5% to 12% in cross-configuration scenarios.
• Significant performance gains were observed compared to models without domain adaptation.

Conclusions:

• The proposed DAFD method effectively overcomes domain mismatch issues in fall detection.
• DAFD enhances the reliability and performance of FD systems in new, unseen environments.
• This approach facilitates the development of more robust and adaptable fall detection technologies.