Optimising the Apnoea Classification Performance of a Neural Network Classifier Processing ECG-Oximetry Signals

Area of Science:

• Biomedical Engineering
• Signal Processing
• Sleep Medicine

Background:

• Sleep apnea is a common disorder characterized by repeated breathing cessation during sleep.
• Accurate detection of sleep apnea events is crucial for diagnosis and treatment.
• Integrating electrocardiogram (ECG) and oximetry signals offers a multimodal approach to enhance detection accuracy.

Purpose of the Study:

• To optimize a neural network system for classifying sleep breathing events using principal components analysis (PCA).
• To develop an algorithm that processes combined ECG and oximetry data for improved sleep apnea detection.
• To evaluate the system's performance in identifying normal breathing, central apnoea (CA), and obstructive apnoea (OA) epochs.

Main Methods:

• Utilized principal components analysis (PCA) to optimize a neural network.
• Developed a pooled feature set incorporating oximetry-derived desaturations and ECG time/spectral features.
• Trained and tested the system on a dataset of 125 scored polysomnogram recordings with respiratory event annotations.

Main Results:

• The system achieved 91% specificity in classifying three epoch types (normal, CA, OA).
• Sensitivity for central apnoea (CA) was 28%, and for obstructive apnoea (OA) was 63%.
• When combining CA and OA into a single class, overall sensitivity reached 81%.

Conclusions:

• PCA-optimized neural networks can effectively process combined ECG and oximetry signals for sleep breathing event detection.
• The developed system demonstrates potential for automated sleep apnea diagnosis, particularly when grouping apnoea types.
• Further refinement may improve sensitivity for specific apnoea subtypes.