Feasibility of Using Deep Learning to Detect Hypoxemia During Painless Gastrointestinal Endoscopy Based on Facial Images

Objective To develop and validate deep learning (DL) and machine learning (ML) models based on preoperative facial images and structured clinical data for predicting hypoxemia during adult painless gastrointestinal endoscopy, addressing the limitations of subjective and data‐intensive prediction methods. Methods A total of 424 patients aged 18–80 years from the East Campus of the First Affiliated Hospital of Zhengzhou University (training/validation set) and 60 patients from the West Campus (test set) between October 2022 and October 2023 were included. A ResNet‐50 DL architecture and five ML algorithms (LightGBM, XGBoost, Random Forest, Logistic Regression, were Naive Bayes) were trained. K‐fold cross‐validation, data augmentation, and class weighting were applied to address class imbalance. Model performance was assessed using AUC, precision–recall curves, calibration curves, confusion matrices, class activation maps, and accuracy. Results ResNet‐50 achieved the highest AUC of 0.8967 (95% CI: 0.8598–0.9402) and accuracy of 0.8500 (95% CI: 0.7333–0.9167), outperforming all ML models. LightGBM was the best ML performer (AUC: 0.8304; accuracy: 0.8176). Conclusions The proposed ResNet‐50 model enables accurate, objective, and noninvasive prediction of hypoxemia risk using only facial images, offering a rapid preoperative risk assessment tool. This approach may enhance patient safety and streamline anesthesia workflows in outpatient endoscopy. Trial Registration: Chinese Registry of Clinical Trials: ChiCTR2400082682