Abstract
Objectives: Combining domain knowledge and machine learning, we developed a screening tool using the smallest number of bio-features and self-reported symptoms to predict obstructive sleep apnea.
Methods: This study analyzed data on 30 self-reported symptoms plus 24 clinical features collected from 3,447 patients who received polysomnography in one regional hospital and one medical center during 2011-2018. We compared the area under the Receiver Operating Characteristic curve among three prediction modules: multivariate logistic regression model, support vector machine, and neural network method. In addition, the odds ratio was also evaluated by gender and age group.
Results: We found AUROC consistently increased by 0.01-0.10 after adding the item of self-reported snoring. Besides, the performance of 4-items (gender, age, body mass index, snoring) was similar to those of adding two more items (neck and waist circumference) for the prediction of moderate to severe OSA (Apnea Hypopnea Index ≥15) in all three prediction models, indicating the significance of domain knowledge. The AUROC of the 4-item test set were 0.83, 0.84, and 0.83 for MLR, SVM, and NN, respectively. Further analysis indicated participants with BMI ≥25 and frequent snoring were at higher risk of moderate to severe OSA.
Conclusions: The module of OSA developed in this study indicated overweight and frequent snoring as suitable criteria for early detection of OSA by primary care physicians.
Footnotes
Cite this article as ERJ Open Research 2021; 7: Suppl. 7, 43.
This is an ERS Lung Science Conference abstract. No full-text version is available. Further material to accompany this abstract may be available at www.ers-education.org (ERS member access only).
- Copyright ©the authors 2021
