Explainable Machine-Learning Model for Predicting Severe Obstructive Sleep Apnea in Patients Undergoing Metabolic...

Area of Science:

• Medical Informatics
• Sleep Medicine
• Bariatric Surgery

Background:

• Severe obstructive sleep apnea (SOSA) increases perioperative risks in bariatric surgery patients.
• Current polysomnography screening is costly and inaccessible.
• Need for accessible screening methods for SOSA in bariatric surgery candidates.

Purpose of the Study:

• Develop and validate an explainable machine learning (ML) model for SOSA prediction.
• Identify key clinical variables for SOSA risk assessment.
• Create a tool for perioperative risk stratification in bariatric surgery.

Main Methods:

• Utilized data from 1,690 patients in the Chinese Obesity and Metabolic Surgery Database.
• Employed feature selection techniques and evaluated eight ML algorithms.
• Assessed model interpretability using SHapley Additive exPlanations (SHAP).

Main Results:

• Identified a random forest model with nine key predictors (e.g., neck circumference, BMI, age).
• Achieved an AUC of 0.931 (training) and 0.869 (validation) for SOSA prediction.
• The model was deployed as an accessible online tool.

Conclusions:

• An explainable random forest model effectively predicts SOSA in bariatric surgery patients.
• The model offers clinical applicability for risk stratification and decision-making.
• The free online tool facilitates targeted screening and cost-effective care.