Machine learning to predict complications after salvage surgery in head and neck cancers

The aim of this study was to develop a machine learning classification model that can forecast salvage surgery complications. This was a retrospective analysis of all patients who underwent salvage surgery for head and neck cancer at Amrita Institute of Medical Sciences, Kochi, India. Clinical and treatment characteristics were used as predictors. Machine learning models, logistic regression, regularized logistic regression, Random Forest, and Support Vector Machine (SVM) were employed. Overall, 277 patients who underwent salvage surgery were identified, of whom 125 had complications. Random Forest gave the best test sensitivity and specificity of 80.0% (95% confidence interval 59.3-93.2%) and 71.0% (95% confidence interval 52.0-85.8%), respectively. The important features identified were reconstruction status (none), disease-free interval, type of surgery (when primary and neck surgery were combined), alcohol history, and recurrent node positivity. This study explored multiple machine-learning approaches in the predictive analysis of the morbidity of salvage surgery related to head and neck cancers. Random Forest emerged as the best classification model to predict the complications.