Comparison of multivariable logistic regression and machine learning models for predicting pharyngocutaneous fistula after surgery for laryngeal and hypopharyngeal carcinoma.

Purpose: The aim of this study was to comprehensively compare multifactorial logistic regression and various machine learning models in predicting pharyngocutaneous fistula (PCF) following laryngectomy in patients with laryngeal carcinoma and hypopharyngeal carcinoma. Methods: Utilizing a significant dataset from West China Hospital, Sichuan University, we retrospectively analyzed the medical records of 2,863 patients diagnosed with laryngeal or hypopharyngeal cancer who underwent surgical treatment from 17 March 2008 to 9 May 2022 to identify critical risk factors for postoperative PCF. Our approach encompassed traditional statistical methods and advanced machine learning techniques, including Random Forest, Decision Tree, XGBoost, and Support Vector Classification. Results: Of the 2,863 patients undergoing laryngectomy, 263 (9.18%) developed postoperative PCF. In the validation set, the XGBoost model achieved the highest AUC (0.759), while the multivariable logistic regression model achieved an AUC of 0.753; however, the difference was not statistically significant. Logistic regression showed favorable calibration and clinical net benefit and was selected as the final model. Skin flap reconstruction and advanced tumor stage, particularly T3/T4 and N2/N3 disease, were important predictors of PCF. Conclusion: Machine learning models showed predictive performance comparable to multivariable logistic regression but did not significantly improve discrimination. Considering its interpretability, calibration, clinical net benefit, and ease of implementation, multivariable logistic regression may be a practical model for predicting postoperative PCF after laryngectomy. Further prospective studies with external validation are warranted.