A new lightweight deep learning model optimized with pruning and dynamic quantization to detect freezing gait on...

Area of Science:

• Biomedical Engineering
• Neurology
• Artificial Intelligence

Background:

• Freezing of gait (FoG) is a major mobility issue in Parkinson's disease, increasing fall risk.
• Accurate, real-time FoG detection is needed for wearable devices to improve patient safety.
• Current deep learning (DL) models for FoG detection are often too large for resource-constrained wearables.

Purpose of the Study:

• To develop a lightweight deep learning (DL) model for accurate and efficient freezing of gait (FoG) detection.
• To optimize the DL model for deployment on wearable devices with limited computational resources.
• To enable continuous, real-time monitoring of FoG in Parkinson's disease patients.

Main Methods:

• Proposed a novel lightweight DL model combining convolutional neural networks (CNN) and gated recurrent units (GRU) with attention mechanisms.
• Incorporated model optimization techniques: pruning, dynamic quantization, and a feature selection method.
• Evaluated model performance against state-of-the-art supervised DL models for FoG detection.

Main Results:

• Achieved a high F1 score of 0.994 for FoG detection, outperforming existing models.
• The proposed model uses 29.9 times fewer parameters and has a maximum memory usage of only 420.91 KB.
• Pruning and quantization reduced model size by 7.84 times, to 44.04 KB, without accuracy loss.

Conclusions:

• The developed lightweight DL model provides accurate, real-time FoG detection suitable for wearable devices.
• Minimal memory footprint and parameter count make the model practical for resource-limited applications.
• This technology offers a viable solution for managing FoG and improving quality of life for Parkinson's patients.