Related Experiment Video
Updated: May 8, 2026
Drug-Induced Sleep Endoscopy (DISE) with Target Controlled Infusion (TCI) and Bispectral Analysis in Obstructive Sleep Apnea
Published on: December 6, 2016
Interpretable two-stage deep learning for pediatric obstructive sleep apnea diagnosis using lateral cephalograms
Jiayi Zhang1,2,3,4, Jiao Tan1,2,3,4, Xuesha Tong1,2,3,4
1The Affiliated Stomatological Hospital of Chongqing Medical University, Chongqing, China.
Insights
An AI framework using lateral cephalograms (LCs) accurately detects pediatric obstructive sleep apnea-hypopnea syndrome (OSAHS). This tool aids early diagnosis and treatment by improving dental professionals' diagnostic accuracy.
Area of Science:
- Medical Imaging
- Artificial Intelligence in Healthcare
- Pediatric Sleep Medicine
Background:
- Obstructive sleep apnea-hypopnea syndrome (OSAHS) is a common pediatric disorder.
- Early detection is hindered by limited accessibility and efficiency of current diagnostic methods.
Purpose of the Study:
- Develop and validate an AI framework for automated, accurate, and interpretable risk evaluation of pediatric OSAHS.
- Utilize routine lateral cephalograms (LCs) for OSAHS diagnosis.
Main Methods:
- Retrospective enrollment of 188 children.
- Development of a two-stage interpretable AI framework using LCs for upper airway segmentation and OSAHS classification.
- Comparison of different input strategies and use of Grad-CAM for model interpretation.
- Evaluation of clinical utility via a reader study.
Main Results:
- High performance in upper airway segmentation (DSC: 0.931, IoU: 0.872).
- Superior OSAHS classification by the fusion model (AUC: 0.945) compared to LCs-only and ROI-based models.
- AI assistance significantly improved diagnostic accuracy for dentists of all experience levels.
Conclusions:
- The AI framework shows promise for automated pediatric OSAHS diagnosis using LCs in dental settings.
- The model enhances diagnostic accuracy and interpretability, supporting early detection and personalized management.
Objective:
Obstructive sleep apnea-hypopnea syndrome (OSAHS) is a prevalent sleep-breathing disorder in pediatrics, yet early and accurate detection remains challenging due to the limited accessibility and efficiency of conventional diagnostic methods. This study aims to develop and validate an artificial intelligence framework that utilizes routine lateral cephalograms (LCs) to provide automated, accurate, and interpretable risk evaluations for pediatric OSAHS.
Methods:
We retrospectively enrolled 188 children from two hospitals between January 2021 and October 2025. A total of 150 LCs were used for cross- validation, and 38 LCs were reserved as an independent test dataset. Using LCs, we proposed and developed an interpretable two-stage framework for pediatric OSAHS diagnosis. The first stage segmented the upper airway, and the second stage performed classification using a modified fusion model that integrate information from both craniofacial structures and the upper airway. We compared different input strategies and used Grad-CAM for model interpretation. Clinical utility was evaluated in a reader study comparing dentists' performance across different experience levels.
Results:
The upper airway segmentation achieved a mean DSC of 0.931 and an IoU of 0.872. For OSAHS classification, the fusion model achieved an AUC of 0.945 (95% CI: 0.863-0.994) and an F1 score of 0.933 (95% CI: 0.818-0.995), outperforming the LCs model (AUC 0.797, 95% CI: 0.585-0.968) and the mask-based ROI model (AUC 0.882, 95% CI: 0.748-0.983). Grad-CAM consistently highlighted anatomically plausible regions related to craniofacial structure and the upper airway. In the reader study, AI assistance increased diagnostic accuracy by 0.165 for junior dentists and 0.237 for senior dentists.
Conclusion:
Our model represents a promising tool for automated pediatric OSAHS diagnosis based on routinely acquired LCs in dental settings. By enhancing diagnostic accuracy and interpretability, it has the potential to support early detection and individualized management.
