Predicting Obstructive Sleep Apnea Based on Computed Tomography Scans Using Deep Learning Models.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Sleep Medicine

Background:

• Clinically undiagnosed obstructive sleep apnea (OSA) is prevalent due to limited polysomnography access.
• Craniofacial computed tomography (CT) scans can potentially predict OSA and its severity.

Purpose of the Study:

• To predict OSA and its severity using paranasal CT scans with a novel 3D deep learning algorithm.
• To develop and validate an AI model for non-invasive OSA diagnosis.

Main Methods:

• A multimodal deep learning model (AirwayNet-MM-H) was developed using a 3D convolutional neural network and multilayer perceptron.
• The model integrated CT images with structured data (age, sex, BMI) from internal (N=798) and external (N=135, N=85) datasets.
• The model was trained to classify OSA severity (normal, mild, moderate, severe) and significant OSA (moderate to severe).

Main Results:

• The AirwayNet-MM-H model achieved 87.6% accuracy for four-class OSA severity prediction in the internal dataset.
• For predicting significant OSA, the model demonstrated 91.0% accuracy and an AUC of 0.910 in the internal dataset.
• The model outperformed six other state-of-the-art deep learning models in accuracy and AUC (P < 0.001).

Conclusions:

• A novel multimodal deep learning model utilizing CT scans can accurately diagnose OSA and its severity.
• The developed AI model, AirwayNet-MM-H, offers a precise and potentially accessible tool for OSA diagnosis.
• This approach leverages existing CT scans, improving diagnostic capabilities for obstructive sleep apnea.