AI Detecting Severe Sleep Apnea from Laterla Cepahalometric X-Rays

1. Title:

Machine learning for image-based detection of patients with obstructive sleep apnea: an exploratory study

2. Authors: Satoru Tsuiki; Takuya Nagaoka; Tatsuya Fukuda; Yuki Sakamoto; Fernanda R. Almeida; Hideaki Nakayama; Yuichi Inoue; Hiroki Enno

3. Summary:

This study investigated the use of a deep convolutional neural network (DCNN), specifically VGG-19, to detect severe obstructive sleep apnea (OSA) using 2-dimensional lateral cephalometric radiographs. The model was trained and tested on a dataset of 1,389 male patients (867 with severe OSA and 522 non-OSA). The researchers compared three image regions: the full image, a "main region" (focusing on the oropharynx/craniofacial structures), and a "head only" region. The "main region" model achieved the highest performance (Sensitivity 0.88, Specificity 0.75, AUC 0.92), identifying severe OSA with high accuracy and outperforming conventional manual cephalometric analysis performed by experts. The study concludes that DCNNs applied to standard radiographic images could be a useful tool for OSA triage

VISUAL SUMMARY

4. Key Words: Obstructive sleep apnea ; Oropharyngeal crowding ; Artificial intelligence; Machine learning

5. Extracted Data

5.1. Year: 2021

5.2. Modality: 2D Lateral Cephalometric Radiography

5.3. Dataset: 1389 total images

5.4. Dataset Split: 90% Training (1251 images) / 10% Testing (138 images)

5.5. Network Architecture: Deep Convolutional Neural Network (DCNN) - VGG-19

5.6. Metrics: AUC (ROC); Sensitivity; Specificity; PPV; NPV; Likelihood Ratios;

5.7. AP - Professional Qty: 2

5.8. AP - Supervisor Presence: Unavailable

5.9. AP - Experience Level: Unavailable

5.10. AP - Expertise Area: Radiologist; Orthodontist

5.11. AP - Tool or System: Ground truth via Polysomnography (PSG); Manual cephalometric analysis

5.12. ML Task: Binary Classification (Severe OSA vs Non-OSA)

5.13. Project Objective: To test whether a deep convolutional neural network could differentiate severe OSA from non-OSA using 2D cephalometric radiographs.

6. Clinical Relevance: This model demonstrates the potential of AI-based image analysis as a screening tool for severe OSA in dental and primary care settings, supporting earlier referral and reducing diagnostic subjectivity.