Energy-Efficient Respiratory Anomaly Detection in Premature Newborn Infants

Area of Science:

• Biomedical Engineering
• Wearable Technology
• Neonatal Monitoring

Background:

• Accurate respiratory rate monitoring is crucial for premature infants.
• Existing wired systems are invasive and impede infant movement.
• Wireless, non-invasive monitoring is needed for neonatal care.

Purpose of the Study:

• To develop a deep-learning-enabled wearable system for non-invasive respiratory monitoring in premature infants.
• To predict respiratory cessation using wireless signals from a wearable sensor.
• To explore energy-efficient solutions for wearable neonatal monitoring.

Main Methods:

• A five-stage design pipeline including data collection, feature scaling, model selection (1DCNN), hyperparameter tuning, training, validation, testing, and deployment.
• Evaluation of quantization techniques for energy reduction in wearable devices.
• Development and conversion of a 1DCNN to a Spiking Neural Network (SNN) for neuromorphic hardware.

Main Results:

• The baseline 1DCNN achieved 97.15% classification accuracy.
• Quantization techniques reduced energy but significantly degraded accuracy.
• The proposed SNN solution achieved 93.33% accuracy with 18x lower energy than the 1DCNN, outperforming quantized models in energy efficiency.

Conclusions:

• Deep learning with a 1DCNN offers high accuracy for neonatal respiratory monitoring.
• Model compression techniques like quantization are suboptimal for balancing accuracy and energy in this application.
• Spiking Neural Networks (SNNs) provide a promising, energy-efficient alternative for wearable neonatal monitoring on neuromorphic hardware.