Development of technology for diagnosing and predicting severity of obstructive sleep apnea through deep learning CT analysis
Oct 06, 2024
|
A joint research team consisting of Professor Gong Hyun-joong of the Department of Convergence Medicine at Seoul National University Hospital, Professor Kim Hyun-jik of the Department of Otolaryngology at Seoul National University Bundang Hospital, Professor Kim Jung-hoon of the Department of Otolaryngology at Dongguk University Ilsan Hospital, Professor Park Seok-won and Professor Kim Jin-yeop of the Department of Electrical and Electronic Computer Engineering at DGIST (Professor Lee Kyung-soo at Chonbuk National University) developed a diagnosis and severity prediction method based on a deep learning model for a total of 1,018 patients with obstructive sleep apnea and published the results of.
Obstructive sleep apnea is a disease in which the upper airways are repeatedly narrowed or blocked during sleep, making breathing difficult, and about 6-38% of the world's population suffers from the disease. It causes various complications such as cardiovascular disease, diabetes, and depression, and seriously deteriorates sleep quality. However, conventional polysomnography is expensive and has limited access to medical care, so many patients have not been properly diagnosed.
To solve this problem, the research team developed a deep learning model called AirwaysNet-MM-H that can diagnose obstructive sleep apnea and predict severity using previously photographed cranial CT images including sinuses. The model significantly improved the accuracy of the prediction by combining 3D CT images with the patient's age, gender, and body mass index (BMI), and further improved the performance by applying a preprocessing algorithm that emphasizes the airway area.
The AirwaysNet-MM-H model combines a three-dimensional convolutional neural network (3D CNN), which analyzes the three-dimensional structure of CT images, and a multilayer perceptron (MLP), which processes patients' data, to diagnose obstructive sleep apnea, and to predict the severity of diagnosed patients. This model is classified as grade 4 (normal, mild, moderate, severe), or used as a grade 2 (moderate, mild/normal) classification method to predict moderate or higher obstructive sleep apnea.
In this study, we trained the model and validated its performance on 798 internal and 135 and 85 external datasets.
As a result, the AirwaysNet-MM-H model achieved 87.6% prediction accuracy with internal data in the 4th classification, and 84.0% and 86.3% high accuracy on external datasets, respectively.
In particular, the Grade II classification predicting moderate or higher obstructive sleep apnea recorded a predictive accuracy of 91.0% from internal data and high predictive accuracy from external data sets as well.
Additionally, we compare the diagnostic performance of the AirwaysNet-MM-H model with the existing deep learning model, which shows that the accuracy is up to 14.2% higher than that of the existing model on the internal dataset, and the AUROC value is 0.152 better. The accuracy was 11.9% higher on external datasets, and the AUROC value was 0.111 higher, outperforming six other state-of-the-art deep learning models.
The research team expects this deep learning model to be useful in clinical practice as it can also be used for risk assessment before and after surgery. It also emphasized that it is also cost-effective because it can be diagnosed using CT data that has already been taken without additional tests or costs.
Professor Hyun-Joong Gong (Seoul National University Hospital Convergence Medicine Department) said "This deep learning model can provide diagnosis quickly and accurately to patients who have difficulty accessing existing sleep polytests and is expected to significantly improve the quality of life of patients through early diagnosis and treatment."We plan to validate and improve performance with more external data in the future, and evaluate its applicability to different races and patient populations."
The findings were published in the latest issue of the American Journal of Respiratory and Critical Care Medicine (IF 19.3)'.
|
bellho@sportschosun.com
