CNN-Based Self-Attention Weight Extraction for Fall Event Prediction Using Balance Test Score.

Area of Science:

• Gerontology
• Biomedical Engineering
• Data Science

Background:

• Falls are a major cause of injury, hospitalization, and death in the elderly.
• Early identification of individuals at risk of recurrent falls is critical for effective prevention.
• Current methods may lack interpretability or require extensive labeled data.

Purpose of the Study:

• To evaluate an interpretable semi-supervised approach for identifying elderly fall risks using ankle-mounted IMU sensor data.
• To leverage the relationship between fall events and balance ability for precise risk identification.
• To develop a model that can train on unlabeled data and provide interpretable insights.

Main Methods:

• Utilized an interpretable semi-supervised learning framework with data from ankle-mounted Inertial Measurement Unit (IMU) sensors.
• Employed a visual-based self-attention model to infer the link between fall events and balance loss.
• Focused on identifying moments where vertical acceleration exceeded 5 m/s² within a short duration.

Main Results:

• The model successfully identified the relationship between fall events and balance impairment.
• High attention weights were assigned to critical moments of vertical acceleration, indicating fall risk.
• Achieved 71% detection of potential fall risk events within a 1-second window, outperforming threshold-based methods.
• Demonstrated reduced annotation costs for fall prevention using wearable devices.

Conclusions:

• The developed semi-supervised approach effectively identifies elderly individuals at high risk of falling.
• The interpretable nature of the model provides valuable insights for healthcare professionals.
• This adaptive tool can significantly enhance large-scale, cost-effective fall prevention efforts.