Construction and validation of a machine learning-based prediction model for postoperative complications in patients
Jinlei Wang1, Fuxun Zhang2, Lijing Liu1
1Department of Eye, Ear, Nose, Throat and Oral Ward, No. 971 Hospital of People's Liberation Army Navy, Qingdao, China.
Objective:
A machine learning prediction model was developed and validated using Clinical Multi-omics (CMO) indicators to assess the risk of postoperative complications in patients with Chronic Otitis Media (COM). This model is intended to aid in perioperative management and individualized intervention.
Methods:
Patients were randomly allocated into a training set (n = 237) and a validation set (n = 101) in a 7:3 ratio. A total of 21 CMO indicators, including demographic, clinical, laboratory, and imaging data, were collected. In the training set, univariate analysis, Least Absolute Shrinkage and Selection Operator (LASSO) regression, and multivariate logistic regression were used to identify core predictive variables. Four machine learning models-Random Forest, Logistic Regression, K-Nearest Neighbors, and Gradient Boosting Machine-were constructed using these variables. Model performance was assessed using the area under the curve (AUC), calibration curves, and decision curve analysis (DCA). Interpretability was analyzed with SHapley Additive exPlanations (SHAP).
Results:
Multivariate analysis identified a history of diabetes, previous ear surgery, otorrhea, middle ear mucosal status, cholesteatoma presence, Eustachian tube function score, and preoperative C-reactive protein level as independent risk factors. Among the constructed models, the Random Forest model demonstrated superior overall performance, and the model was 0.885 in the training set and 0.853 in the validation set. The model also showed good calibration. DCA indicated significant clinical net benefit across a wide threshold probability range. SHAP analysis confirmed that a history of previous ear surgery and cholesteatoma presence were the most influential predictors.
Conclusion:
A machine learning-based prediction model for complications after COM surgery was developed and validated. The Random Forest model performed optimally, effectively predicting complication risk with favorable performance and considerable potential for clinical translation. It can serve as a promising tool for preoperative risk assessment and targeted postoperative monitoring.
