Two-Stream Convolutional Neural Networks for Breathing Pattern Classification: Real-Time Monitoring of Respiratory...

Area of Science:

• Biomedical Engineering
• Artificial Intelligence
• Respiratory Medicine

Background:

• Continuous monitoring of breathing patterns is crucial for managing infectious respiratory diseases.
• Existing methods may lack accuracy or efficiency for real-time patient surveillance.
• Wearable near-infrared spectroscopy (NIRS) offers a non-invasive approach to capture physiological data.

Purpose of the Study:

• To develop and evaluate a two-stream convolutional neural network (TCNN) for accurate breathing pattern classification.
• To compare the TCNN's performance against single-stream CNNs (SCNNs) and other state-of-the-art models.
• To assess the TCNN's effectiveness in classifying normal, slow, rapid, and breath-holding patterns.

Main Methods:

• A TCNN architecture was designed, integrating a CNN-based autoencoder for feature extraction and a classifier.
• The TCNN processes hemodynamic response data from chest tissue, measured via wearable NIRS.
• Performance was evaluated by classifying four breathing patterns in 14 healthy adults and compared to SCNN and random forest models.

Main Results:

• The TCNN achieved the highest classification accuracy at 94.63%, outperforming random forest (88.49%) and SCNN (92.03%).
• The TCNN demonstrated a 2.6% improvement in accuracy over the SCNN.
• The TCNN effectively addressed the diminishing learning performance associated with increased network depth, unlike the SCNN.

Conclusions:

• The TCNN is a robust and accurate model for breathing pattern classification.
• The TCNN offers an efficient solution for continuous patient monitoring, requiring fewer parameters and computations.
• This AI-driven approach shows promise for improved management of respiratory conditions.