Lightweight Deep Learning Approaches on Edge Devices for Fetal Movement Monitoring.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Healthcare
• Wearable Technology

Background:

• Fetal movement monitoring (FMM) is vital for fetal well-being assessment.
• Current FMM methods (clinical assessment, maternal perception) have limitations.
• Need for efficient, real-time FMM solutions, especially for edge devices.

Purpose of the Study:

• Develop a lightweight deep learning framework for real-time FMM on edge devices.
• Optimize the framework for resource-constrained microcontrollers.
• Improve accuracy and efficiency of FMM for enhanced prenatal care.

Main Methods:

• Collected accelerometer and gyroscope data from 120 participants using wearable IMUs.
• Implemented a two-stage labeling protocol with maternal perception and ultrasound validation.
• Utilized virtual-rotation augmentation and adaptive clustering for class imbalance.
• Transformed data to spectrograms for deep learning input.
• Employed knowledge distillation and post-training integer quantization for model optimization.

Main Results:

• Achieved high performance with sensitivity (90.05%), precision (87.29%), and F1-score (88.64%).
• Reduced model memory footprint by 74.8% through quantization.
• Demonstrated continuous operation for ~25 hours on a single battery charge.
• Framework is suitable for resource-constrained microcontrollers.

Conclusions:

• The novel lightweight deep learning framework enables efficient, real-time FMM on edge devices.
• Optimized model performance and reduced memory footprint significantly.
• Offers a practical solution for improved prenatal care, especially in resource-limited settings.
• Framework is adaptable for other edge-based medical monitoring tasks.