Development and Validation of an Interpretable Preoperative Difficult Airway Risk Prediction Model Based on Multiple Machine Learning Algorithms

Difficult airway management remains a critical challenge in anesthesia, with significant implications for patient safety. Traditional preoperative risk assessment tools often lack sensitivity and interpretability. This study aimed to develop and validate an interpretable machine learning-based model for predicting difficult airway risk, integrating robust feature selection and model explanation techniques to enhance clinical utility and trust.

This single-center retrospective observational cohort study analyzed 400 adult patients who underwent surgery under general anesthesia with tracheal intubation. Candidate predictors were initially screened using univariate analysis, LASSO regression analysis was applied for feature selection to avoid overfitting, followed by multivariate logistic regression to identify independent risk factors. 7 machine learning algorithms were developed and compared on testing set to develop the prediction model. Model performance was assessed by AUC, calibration curves, and decision curve analysis. The final model was interpreted using SHapley Additive exPlanations (SHAP) to provide both global and local interpretability.

Univariate analysis revealed 23 variables in training set significantly associated with difficult airway (p<0.05). LASSO regression further selected 11 predictors, which were entered into multivariate logistic regression. The final prediction model included 5 independent risk factors: BMI, thyromental distance, sternomental distance, neck circumference, and glucose. Among the 7 algorithms, logistic regression demonstrated the best overall performance (AUC = 0.871), excellent calibration, with high net clinical benefit across clinically relevant thresholds. SHAP analysis revealed BMI as the most influential predictor, with clear directional effects. Local explanations for individual cases further enhanced model transparency.

We developed an interpretable prediction model for preoperative difficult airway risk. By combining feature selection with SHAP-based explanations, the model provides transparent individualized risk estimates. External and prospective validation is required before clinical implementation.