Modification of a Conventional Deep Learning Model to Classify Simulated Breathing Patterns: A Step toward Real-Time...

Area of Science:

• Biomedical Engineering
• Medical Informatics
• Physiological Monitoring

Background:

• The COVID-19 pandemic highlighted the need for remote patient monitoring solutions for respiratory diseases.
• Existing consumer-grade devices lack automated, continuous day-and-night monitoring capabilities.
• Wearable sensors offer a promising avenue for non-invasive physiological data collection.

Purpose of the Study:

• To develop and validate a real-time breathing pattern classification system.
• To utilize tissue hemodynamic responses measured by near-infrared spectroscopy (NIRS).
• To employ a deep convolutional neural network (CNN) for accurate classification.

Main Methods:

• Collected tissue hemodynamic data from 21 healthy volunteers using a wearable NIRS device.
• Developed three distinct one-dimensional CNN (1D-CNN) models based on the Pre-ResNet architecture.
• Modified the Pre-ResNet model for real-time classification of breathing patterns.

Main Results:

• Achieved an average classification accuracy of 88.79% without Stage 1 convolutional layer.
• Improved accuracy to 90.58% with a 1 × 3 Stage 1 layer.
• Reached a peak accuracy of 91.77% with a 1 × 5 Stage 1 layer.

Conclusions:

• The developed deep CNN-based method effectively classifies breathing patterns in real-time.
• NIRS combined with CNNs shows significant potential for continuous, automated respiratory monitoring.
• This technology could enhance remote care for patients with infectious respiratory diseases.