0507 A Novel Machine Learning Model to Predict Obstructive Sleep Apnea Using Craniofacial Photography with Questionnaire

Le résumé fourni par la source

Abstract Introduction Obstructive sleep apnea (OSA) is a complex and heterogeneous sleep-related breathing disorder, associated with systemic consequences such as hypertension, stroke, and cardiovascular diseases. Although initial screening tools such as STOP-BANG or the Berlin questionnaires have been developed, it is still challenging to capture the wide spectrum of the condition alone. Besides the questionnaires including clinical presentations, craniofacial abnormalities are also recognized as an important risk factor for detecting OSA. Herein we aim to develop an efficient approach to predict the risk of having moderate to severe OSA using machine learning techniques by incorporating anatomical information from 2D photographs. Methods This retrospective analysis included 348 patients, who completed answering the STOP-BANG questionnaires and took facial images at Dankook University Medical Center in South Korea between 2012 and 2022. A 1:1 Random under-sampling (RUS) method was applied to solve the imbalance problem between moderate to severe OSA cases (Apnea-Hypopnea Index ≥15 events/hour) and control. Balanced data were randomly divided into two data sets (training and validation: 80%; testing: 20%). We performed leave-one-out cross-validation (LOOCV) to seek the optimal parameters to improve the model performance and generalization using four machine learning models (logistic regression, random forest, support vector machine, XGBoost) in the training set. Thereafter, we adopted the model with the highest area under the receiver operating characteristic (AUROC) in the testing set. Finally, we evaluated the importance of each input feature in assessing OSA risk by calculating the Shapley additive explanations (SHAP) values. Results The logistic regression model achieved the best AUROC of 94.1% for predicting moderate to severe risk OSA. The value of craniofacial images built from deep learning algorithms was found to be a predominant contribution in the risk screening models of OSA with the aforementioned levels of severity. Conclusion The results of the study suggest that the proposed model improved the diagnosis of patients with moderate to severe OSA based on the combination of sleep-related questionnaires and 2D facial images by identifying the risk factors more efficiently. Support (if any)

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.