Automatically detecting OSAHS patients based on transfer learning and model fusion

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Medicine
• Sleep Medicine

Background:

• Snoring is a primary symptom of obstructive sleep apnea hypopnea syndrome (OSAHS).
• Developing non-invasive, automated methods for OSAHS detection is crucial for patient management.
• Current diagnostic methods can be invasive or require specialized facilities.

Purpose of the Study:

• To develop and evaluate a non-invasive, automated approach for detecting OSAHS patients based on snoring sounds.
• To investigate the efficacy of a fused model combining transfer learning techniques for OSAHS classification.

Main Methods:

• A model fusion approach was employed, integrating three basic models: Visual Geometry Group-16 (VGG16), pretrained audio neural networks (PANN), and Mel-frequency cepstral coefficient (MFCC).
• XGBoost was utilized for feature selection based on feature importance.
• A majority voting strategy was used for model fusion, with leave-one-subject-out cross-validation for evaluation.

Main Results:

• The fused model, incorporating top-5 VGG16 features, top-5 PANN features, and MFCC features, achieved 100% accuracy in identifying OSAHS patients (Apnea-Hypopnea Index > 5).
• The model demonstrated high classification performance.

Conclusions:

• The proposed fused model offers a robust and computationally efficient method for OSAHS detection.
• This non-invasive technique holds significant potential for enabling remote and at-home diagnosis of OSAHS.