Deep-Learning Model Based on Convolutional Neural Networks to Classify Apnea-Hypopnea Events from the Oximetry...

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Medicine
• Sleep Medicine

Background:

• Nocturnal oximetry (SpO2) analysis aids obstructive sleep apnea (OSA) diagnosis by detecting respiratory events.
• Previous SpO2 studies focused on general respiratory events, not differentiating between apneas and hypopneas.
• Apnea and hypopnea episodes exhibit distinct oxygen desaturation patterns.

Purpose of the Study:

• To develop and evaluate a deep learning model for automatic identification and classification of apnea and hypopnea events using SpO2 signals.
• To assess the diagnostic performance of the model in classifying normal respiration, apnea, and hypopnea.
• To compare the model-derived apnea-hypopnea index (AHI) with standard polysomnography (PSG) measurements.

Main Methods:

• A convolutional neural network (CNN) deep learning architecture was employed.
• The CNN was trained on 30-second epochs of SpO2 signals from 398 adult OSA patients.
• The model classified signals into three categories: normal respiration, apnea, and hypopnea.

Main Results:

• The CNN achieved 80.3% 3-class accuracy and a Cohen's kappa of 0.539 on an independent test set.
• High agreement was observed between model-derived indices (AI_CNN, HI_CNN, AHI_CNN) and standard PSG values (ICCs: 0.8023, 0.6774, 0.8466).
• The model demonstrated promising diagnostic performance for automated OSA diagnosis.

Conclusions:

• CNN-based analysis of SpO2 recordings can effectively differentiate between apnea and hypopnea events.
• This approach offers a potential method for automated diagnosis of OSA using at-home oximetry.
• The findings support the use of deep learning for simplifying OSA diagnosis in clinical practice.