A clinically interpretable model for predicting pharyngocutaneous fistula after total laryngectomy.

backgroundPharyngocutaneous fistula (PCF) is a frequent complication following total laryngectomy. While various risk models exist, their practical utility is often limited by poor calibration and a lack of guidance on modifiable risk factors. This study aimed to develop a calibrated risk assessment tool to support preoperative counseling and clinical decision-making.

methodsWe analyzed a retrospective cohort of 328 patients who underwent total laryngectomy. Model development was restricted to preoperative variables. Intraoperative and postoperative variables were recorded for descriptive reporting only and were not used for model training to avoid information leakage. We compared a conventional logistic regression model against machine learning algorithms, including Random Forest and XGBoost. Model performance was evaluated using discrimination metrics, calibration plots, and decision curve analysis. To explore lifestyle modification, we used model-based counterfactual prediction to estimate the change in predicted risk under hypothetical alcohol cessation. A nomogram was constructed to visualize the model for clinical use.

resultsAmong the evaluated algorithms, the Logistic Regression model yielded the most robust performance. It showed more stable generalization and calibration than tree-based machine learning methods. It achieved an AUC of 0.63 (95% CI 0.56-0.70) with satisfactory calibration. Decision curve analysis confirmed a positive net clinical benefit within a threshold probability range of 10% to 30%. Key predictors identified included COP-NLR score, heavy alcohol consumption, tumor stage, and hemoglobin levels. Counterfactual prediction suggested a lower predicted risk when heavy alcohol consumption was hypothetically set to 0. This estimate is theoretical and does not imply causality.

conclusionsGiven the inherent challenges in predicting PCF using exclusively preoperative variables, this study establishes a calibrated nomogram that prioritizes clinical utility and reliability over raw discrimination. The tool provides accurate risk estimates in the critical 10% to 30% decision window. It highlights heavy alcohol consumption as a potentially modifiable risk marker for preoperative counseling. This finding is hypothesis-generating. The tool supports shared decision-making and perioperative optimization.