Construction and validation of machine learning models combining clinical data and radiological characteristics for...

Deep neck infection (DNI) is a serious condition that spreads rapidly through cervical fascial planes, often leading to airway compromise and sepsis. Airway protection, antibiotic therapy, and surgical drainage are standard treatments, but some patients require reoperation when improvement is insufficient. Because no reliable tools exist to predict reoperation risk, this study developed a machine learning (ML) model integrating clinical and imaging data to anticipate reoperation in DNI.

We retrospectively analyzed 415 patients with surgically treated DNI. Reoperation was defined as an additional incision and drainage performed more than 48 h after the initial surgery. Clinical and computed tomography (CT)-derived features were incorporated into a Categorical Boosting (CatBoost)-based ML model using stratified five-fold cross-validation. Model performance was evaluated using Receiver Operating Characteristic-Area Under the Curve (ROC AUC), Logloss, confusion matrices, accuracy, precision, sensitivity, specificity, and F1-score. Predictive modeling was conducted using both the full feature set and a reduced set of nine significant features identified through feature importance analysis.

The population of study cohort had a mean age of 53.63 years, with a reoperation rate of 33.97%. During cross-validation, the full-feature model achieved an F1-score of 0.9041, accuracy of 0.9488, ROC AUC of 0.9678, precision of 0.9190, sensitivity of 0.9000, and specificity of 0.9670. When evaluated on the independent test set, it yielded an F1-score of 0.8980, accuracy of 0.9398, ROC AUC of 0.9891, precision of 0.8462, sensitivity of 0.9565, and specificity of 0.9333. The reduced nine-feature model attained an F1-score of 0.9025, accuracy of 0.9489, ROC AUC of 0.9503, precision of 0.9281, sensitivity of 0.8889, and specificity of 0.9713 in cross-validation. On the independent test set, it exhibited an F1-score of 0.9565, accuracy of 0.9759, ROC AUC of 0.9957, precision of 0.9565, sensitivity of 0.9565, and specificity of 0.9833. Logloss trajectories corroborated stable convergence with minimal overfitting.

This study demonstrates that an ML model can accurately predict reoperation in DNI by integrating clinical and CT-derived features. The simplified nine-feature model achieved superior performance while enhancing clinical acceptance and practical applicability. These findings highlight the potential of ML to support timely surgical planning and individualized reoperation risk assessment in patients with DNI.