A Multi-Head Attention Transformer Model for Wearable in Situ Fall Detection

Area of Science:

• Biomedical Engineering
• Artificial Intelligence
• Wearable Technology

Background:

• Falls are a major health risk for seniors, causing injuries and reduced quality of life.
• Current fall detection systems using traditional machine learning struggle with complex sensor data.
• There's a need for accurate, real-time fall detection systems integrated into wearable devices.

Purpose of the Study:

• To develop a novel multi-head attention transformer architecture for wearable, in situ fall detection.
• To improve the accuracy and sensitivity of fall detection systems, especially in real-world scenarios.
• To demonstrate the feasibility of deploying the fall detection model on an ultra-low-power FPGA for efficient wearable applications.

Main Methods:

• A novel multi-head attention transformer architecture was designed for processing inertial measurement unit (IMU) sensor data.
• Key innovations include a position-aware embedding layer, multi-head self-attention blocks, and a hybrid feature fusion module.
• The model was trained on a balanced dataset using a specialized Focal Loss function to address data imbalance.

Main Results:

• The proposed transformer model achieved 90.5% fall sensitivity and 93.1% overall accuracy on the test set.
• Addressing dataset imbalance was critical, as a baseline model on imbalanced data had only 48.0% sensitivity.
• The system was co-designed with a wearable PCB and deployed on an FPGA, consuming only 22 mA.

Conclusions:

• The novel transformer architecture effectively models temporal dependencies and spatial features in IMU data for accurate fall detection.
• Addressing dataset imbalance is crucial for developing safe and reliable fall detection systems.
• The efficient hardware/software co-design enables practical, low-power deployment of advanced fall detection on edge devices.