Wearable Fall Detector Using Recurrent Neural Networks

Area of Science:

• Biomedical Engineering
• Computer Science
• Gerontology

Background:

• Falls are a significant public health concern for the elderly, leading to severe health consequences.
• Wearable fall detection devices offer continuous monitoring but face challenges with energy consumption and real-time processing on low-power hardware.
• Recurrent Neural Networks (RNNs) excel at analyzing sequential data but require optimization for resource-constrained embedded systems.

Purpose of the Study:

• To investigate the feasibility of using RNN-based deep learning models for real-time fall and fall risk detection.
• To evaluate the performance and energy efficiency of different RNN architectures on embedded systems.
• To assess the potential of these models for autonomous wearable fall detection systems.

Main Methods:

• Analyzed four RNN architectures using the SisFall dataset at various frequencies.
• Integrated optimized RNN models into two distinct embedded systems to measure execution times and effectiveness.
• Conducted power consumption studies on the integrated systems.

Main Results:

• Achieved high accuracy with 88.2% sensitivity and 96.4% specificity.
• Demonstrated real-time detection capabilities with inference times under 34 ms for simpler models.
• Confirmed high energy efficiency suitable for low-power microcontrollers.

Conclusions:

• RNN models offer an effective and energy-efficient solution for real-time fall detection.
• The developed models can be implemented on low-power microcontrollers for autonomous wearable systems.
• This approach holds significant promise for improving elderly safety through advanced fall detection technology.