Related Experiment Video
Updated: Jul 16, 2026
Drug-Induced Sleep Endoscopy (DISE) with Target Controlled Infusion (TCI) and Bispectral Analysis in Obstructive Sleep Apnea
Published on: December 6, 2016
Explainable Machine-Learning Model for Predicting Severe Obstructive Sleep Apnea in Patients Undergoing Metabolic
Wenhui Chen1, Lili Li2, Junsen Peng3
1Department of Bariatric Surgery Center, First Affiliated Hospital of Jinan University, Guangzhou, China. jnucwh2019@163.com.
Background:
Severe obstructive sleep apnea (SOSA) is associated with an increased risk of perioperative complications in patients undergoing metabolic bariatric surgery. However, polysomnography-based screening remains limited by poor accessibility and high cost. This study aimed to develop and validate an explainable machine learning (ML) model for predicting SOSA in patients with obesity undergoing metabolic bariatric surgery.
Methods:
A total of 1,690 participants from the Chinese Obesity and Metabolic Surgery Database were randomly divided into a training cohort (70%) and a validation cohort (30%). Feature selection was performed using correlation analysis, collinearity analysis, and five feature-selection algorithms and overlapping variables were retained. Eight ML algorithms were developed and evaluated using the area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, specificity, and F1 score. Model interpretability was assessed using SHapley Additive exPlanations (SHAP).
Results:
Among the 1,690 patients, 549 (32.5%) were diagnosed with SOSA. Of the eight ML algorithms, the random forest model demonstrated the best predictive performance using nine key variables: neck circumference, body mass index, age, observed apnea, hematocrit, gamma-glutamyl transferase, sex, creatinine, and type 2 diabetes mellitus. The model achieved an AUC of 0.931 in the training set and 0.869 in the validation set. The final model was implemented as an easy-to-use online tool.
Conclusions:
The explainable random forest model demonstrated excellent predictive performance and clinical applicability for identifying SOSA in patients undergoing metabolic bariatric surgery. This model may facilitate perioperative risk stratification, targeted monitoring, and individualized clinical decision-making.
Key Points:
• The prevalence of SOSA was 32.5% in this large Chinese multicenter bariatric cohort, highlighting the urgent need for efficient preoperative screening in this high-risk population. • The random forest model incorporating nine readily available clinical variables achieved excellent predictive performance with an AUC of 0.869 in the validation cohort. • The model is deployed as a free online calculator, enabling individualized SOSA risk assessment, targeted PSG referral, and cost-effective perioperative risk stratification.
