Evaluation of Machine Learning and Statistical Models for Predicting Long Term Gastrostomy Tube Dependency in Patients Undergoing Major Oral Cavity Cancer Surgery With Free Flap Reconstruction.

Abstract

Swallowing dysfunction and long-term gastrostomy tube dependence are common morbidities after oral cavity cancer resection with free flap reconstruction, highlighting the need for preoperative risk prediction. We performed a multicenter retrospective pooled cohort study of patients with non-metastatic oral cavity cancer undergoing free flap reconstruction. Firth's penalized logistic regression, LASSO, and random forest models were trained using 10-fold cross-validation and predictive performance was assessed via area under the receiver operator curve (AUC). Among 500 patients (mean age 61.7 years), 78 (16%) were gastrostomy tube dependent at 12 months. The random forest model was most accurate in predicting 12-month gastrostomy tube dependency (AUC; 0.77), with comparable performance from LASSO (AUC; 0.71) and Firth's penalized logistic regression (AUC; 0.70). Key predictors included TNM stage IV, T4 disease, unilateral neck dissection, reconstruction site, and no tongue resection. The three models predicted 12-month gastrostomy tube dependency after oral cavity free-flap surgery with fair accuracy and novel predictors; larger multicenter datasets may further improve predictive performance.