Machine learning-based prediction of intraoperative desaturation risk in pediatric dental sedation using preoperative data

Dental sedation in children is widely used to reduce anxiety and facilitate successful dental procedures. However, intraoperative desaturation remains a significant safety concern. This study aimed to evaluate whether machine learning models using only readily available preoperative indicators could predict intraoperative desaturation risk. This study included 206 pediatric patients aged 2–8 years who underwent deep sedation for dental procedures at the outpatient dental clinics of Chang Gung Memorial Hospital, Taiwan. Ten machine learning algorithms were developed to predict desaturation risk. Model performance was evaluated using accuracy, precision, recall, F1-score, and area under the curve (AUC). To enhance interpretability and identify key preoperative risk factors associated with desaturation, SHapley Additive exPlanations (SHAP) was applied to quantify feature contributions. Among 206 participants, 47 (22.8%) experienced intraoperative desaturation (SpO₂ < 90%). Among the evaluated models, XGBoost demonstrated the best overall performance, achieving an accuracy of 0.7039, a precision of 0.6879, a recall of 0.7039, and an F1-score of 0.6952. The SHAP analysis identified the top five preoperative risk factors associated with intraoperative desaturation, including body weight, age, recent upper respiratory infection (URI), sex, and active URI status. Machine learning models using only readily available preoperative indicators demonstrated the feasibility of predicting desaturation risk in pediatric dental sedation. The XGBoost model achieved the best overall performance, while SHAP analysis enhanced model interpretability by identifying clinically relevant preoperative risk factors associated with desaturation. These findings suggest that machine learning–based approaches could assist clinicians in early risk stratification in pediatric sedation practice. Further prospective studies are needed before the proposed models can be considered for clinical application.