MadgwickFall-Net: A Lightweight Dual-Frame Feature Fusion Network for Pre-Impact Fall Detection Using Wearable IMUs

Area of Science:

• Biomedical Engineering
• Gerontology
• Signal Processing

Background:

• Global population aging increases fall-related injuries in the elderly, a critical public health issue.
• Current wearable inertial measurement unit (IMU) based fall detection methods often fail to fully utilize sensor signal information by only using the sensor's body frame.
• Existing advanced methods rely on specific hardware and fusion algorithms, hindering replication and deployment.

Purpose of the Study:

• To develop a novel and effective fall detection system for the elderly using wearable sensors.
• To leverage the Madgwick algorithm for transforming inertial signals into a gravity-aligned global coordinate system for improved feature extraction.
• To create a computationally efficient model suitable for edge device deployment in real-world scenarios.

Main Methods:

• The proposed MadgwickFall-Net utilizes acceleration and angular velocity data.
• It incorporates the Madgwick algorithm to convert inertial signals into a global coordinate system, complementing the body frame signals.
• A four-branch parallel architecture processes signals from both coordinate frames.

Main Results:

• The system achieved an F1-Score of 0.9824 and 98.36% accuracy on the KFall dataset.
• MadgwickFall-Net outperformed all comparison models across four key evaluation metrics.
• The model has a small parameter size (59.7 KB), making it suitable for edge devices, and demonstrated a median pre-impact lead time of 390 ms.

Conclusions:

• MadgwickFall-Net provides a practical and deployable solution for wearable fall detection in the elderly.
• The dual-frame signal processing approach effectively exploits complementary information for enhanced performance.
• The system shows significant potential for protecting elderly individuals by enabling timely fall detection in daily life.