Artificial Intelligence-Based Risk Prediction Models for Complications After Tongue Cancer Surgery

Glossectomy and reconstruction for tongue tumors carries substantial risk of postoperative morbidity, yet current tools offer limited individualized risk stratification.

To develop and validate machine learning (ML) models to predict major 30-day postoperative complications after glossectomy and to compare performance with the American College of Surgeons National Surgical Quality Improvement Program (ACS-NSQIP) risk calculator.

This retrospective cohort study used 2008-2024 data from the ACS-NSQIP database, comprising more than 700 hospitals predominantly in the US. Thirty-one preoperatively available variables were used to train a logistic regression model and 5 ML models (neural network, support vector classifier, LightGBM, XGBoost, and stacked generalization). Among adult patients with tongue cancer, data for 2008 to 2023 were split 85:15 into training and validation sets, while data for 2024 were used as a test set.

Glossectomy of malignant tongue tumors.

Six 30-day outcomes: surgical site infection, bleeding requiring transfusion, pneumonia, unplanned reoperation, any complication, and serious complications. Model prediction performance was primarily assessed on the test set using risk stratification (lift). Discrimination (area under the receiver operating characteristic and area under the precision-recall curve) and calibration (Brier score) were compared with the ACS-NSQIP risk calculator.

Data from 8266 adults (median [IQR] age, 63 [54-71] years; 4895 [59.2%] male) undergoing tongue resection for cancer were used to develop PRO-TONGUE (Postresection and Reconstruction Outcome Prediction Tool for Tongue Cancer Surgery), a publicly available, outcome-specific risk prediction tool that provides individualized, interpretable risk estimates to support preoperative planning. Procedures included partial glossectomy among 6132 patients (74.2%), composite/extended glossectomy among 784 (9.5%), total glossectomy among 341 (4.1%), and minor excisions among 1274 (15.4%). Unplanned reoperation occurred among 708 patients (8.6%), surgical site infection among 510 (6.2%), bleeding requiring transfusion among 546 (6.6%), pneumonia among 262 (3.2%), serious complications among 1197 (14.5%), and any complications among 1428 (17.3%). Outcome-specific model selection based on risk stratification of the test set identified XGBoost as optimal for bleeding, serious complications, and any complications, and LightGBM for reoperation, surgical site infection, and pneumonia. ML and logistic regression models demonstrated discriminative performance of the test set comparable to the ACS-NSQIP risk calculator across outcomes. Prediction of bleeding requiring transfusion, which was not available in the ACS-NSQIP risk calculator, showed the highest area under the receiver operating characteristic (range, 0.88-0.90) and area under the precision-recall curve (range, 0.41-0.52) across all outcomes.

In this cohort study of a large, nationally representative sample, ML demonstrated excellent performance for predicting major 30-day complications after glossectomy in patients with tongue cancer. These models informed development of a tongue cancer surgery risk calculator, warranting further external validation studies.