Enhancing Slip, Trip, and Fall Prevention: Real-World Near-Fall Detection with Advanced Machine Learning Technique

Area of Science:

• Occupational Safety and Health
• Biomedical Engineering
• Machine Learning in Healthcare

Background:

• Slips, trips, and falls (STFs) are a significant cause of workplace injuries and financial losses.
• Existing fall detection methods often rely on simulated falls, limiting real-world applicability.
• There is a need for advanced algorithms that can accurately detect near-falls using ecologically valid data.

Purpose of the Study:

• To systematically evaluate machine learning architectures for near-fall detection using real-world kinematic data.
• To improve the accuracy and robustness of fall detection algorithms in physically demanding work environments.
• To assess the effectiveness of different neural network models for classifying near-fall incidents.

Main Methods:

• Utilized the Prev-Fall dataset, containing high-resolution inertial measurement unit (IMU) data from 110 workers experiencing near-fall incidents.
• Trained and evaluated Convolutional Neural Networks (CNNs), Residual Networks (ResNets), convolutional Long Short-Term Memory networks (convLSTMs), and InceptionTime models.
• Employed neural architecture search to optimize models across various temporal window lengths for near-fall detection.

Main Results:

• Achieved high F1 scores, demonstrating the effectiveness of CNNs and InceptionTime models in near-fall classification.
• Identified recurrent false positives during testing on unobserved data, particularly during activities like bending and squatting.
• Highlighted the necessity of incorporating additional contextual variables to enhance algorithm robustness.

Conclusions:

• Machine learning-based STF prevention systems show promise for workplace safety monitoring and fall mitigation.
• The study's findings support the use of kinematic data and advanced ML models for improved fall detection.
• Future research should focus on multimodal data integration and enhanced classification techniques to improve accuracy and generalizability.