Fine-Grained and Lightweight OSA Detection: A CRNN-Based Model for Precise Temporal Localization of Respiratory...

Area of Science:

• Medical Technology
• Artificial Intelligence
• Respiratory Medicine

Background:

• Obstructive Sleep Apnea (OSA) is a prevalent condition often underdiagnosed due to limited Polysomnography (PSG) resources.
• Audio-based screening offers scalability but struggles with precise respiratory event localization and Apnea-Hypopnea Index (AHI) estimation.

Purpose of the Study:

• To develop a fine-grained, lightweight detection framework for OSA screening using non-contact audio signals.
• To enable precise respiratory event localization and accurate AHI estimation for improved OSA diagnosis.

Main Methods:

• A Dual-Stream Convolutional Recurrent Neural Network (CRNN) integrating Log Mel-spectrograms and energy profiles with BiLSTM was proposed.
• The model was trained on the PSG-Audio dataset and validated on an independent external cohort, evaluating frame-level, event-level, and patient-level performance.

Main Results:

• The CRNN achieved accurate event localization (IoU of 0.82) and strong correlation between audio-derived and PSG-derived AHI (r = 0.96).
• The model demonstrated high sensitivity and specificity across various AHI thresholds in external validation, indicating robust screening capabilities.

Conclusions:

• The developed Dual-Stream CRNN offers a robust and clinically interpretable solution for non-contact OSA screening.
• The framework shows potential as a reliable tool for accessible home-based OSA screening, particularly effective for mild and severe cases.