Identification of Comorbidities in Obstructive Sleep Apnea Using Diverse Data and a One-Dimensional Convolutional...

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Medicine
• Sleep Medicine Research

Background:

• Deep learning (DL) advances enable integrating diverse biomedical data for disease prediction.
• Obstructive sleep apnea (OSA) is frequently associated with major comorbidities.
• Accurate comorbidity identification is crucial for effective OSA patient management.

Purpose of the Study:

• To develop and evaluate a multimodal DL framework for multi-label classification (MLC) of OSA comorbidities.
• To utilize physiological time series signals (SpO2, airflow) and clinical data for comorbidity prediction.
• To assess the framework's robustness and generalization across patient demographics.

Main Methods:

• A retrospective observational study of 144 OSA patients.
• Development of a 1D Convolutional Neural Network (1D-CNN) to fuse SpO2, airflow, and clinical data.
• Model training with Optuna for hyperparameter optimization and weighted loss for class imbalance.

Main Results:

• The 1D-CNN model achieved superior performance (macro AUC-ROC=0.731, AUC-PR=0.750) compared to traditional ML classifiers.
• The model demonstrated consistent performance across different age, gender, and BMI groups, indicating minimal demographic bias.
• Physiological signals (SpO2, airflow) were found to encode comorbidity-specific patterns.

Conclusions:

• SpO2 and airflow signals can enable efficient and scalable screening of OSA-related comorbidities.
• The proposed DL framework provides a methodological basis for clinical decision support systems in sleep medicine.
• Future work should focus on multi-center datasets to enhance model interpretability and clinical adoption.