An Effective Deep Learning Framework for Fall Detection: Model Development and Study Design.

Area of Science:

• Wearable sensor technology
• Human motion analysis
• Artificial intelligence in healthcare

Background:

• Fall detection systems (FDS) are crucial for health monitoring.
• Existing FDS often overlook variable data segment contributions, impacting accuracy.
• Deep learning (DL) offers potential for improved fall detection by analyzing complex motion patterns.

Purpose of the Study:

• To develop and validate a DL framework for accurate fall detection using wearable sensor data.
• To identify essential features for distinguishing falls from daily activities.
• To enhance FDS by creating a weighted feature representation for better differentiation of fall events.

Main Methods:

• Proposed a dual-stream convolutional neural network self-attention (DSCS) model using 3-axis acceleration and gyroscope data.
• Incorporated a self-attention module to weight feature contributions and improve classification.
• Trained and tested the DSCS model on public datasets (SisFall, MobiFall) and conducted practical validation with 10 participants.

Main Results:

• Achieved high accuracy on public datasets: 99.32% (SisFall) and 99.65% (MobiFall).
• Demonstrated superior performance on MobiFall, achieving the best accuracy, recall, and precision.
• Showcased robust practical validation accuracy of 96.41%.

Conclusions:

• The DSCS model significantly enhances fall detection accuracy.
• The model demonstrates robust performance in real-world scenarios.
• This research offers a promising advancement in wearable-based fall detection systems.