OBJECTIVES: Predicting cancer survival is important for patient care management. The objective of the study was to evaluate the performance of common machine learning (ML) models in predicting 5-year overall survival of Head and Neck Squamous Cell Carcinoma (HNSCC) based on clinical and demographic prognostic factors.

METHODS: Patients diagnosed with malignant HNSCC between 2008 and 2019 of age≥20 years were identified from the Surveillance, Epidemiology, and End Results (SEER) 17 registries database. Patients having unknown/missing values, or multiple primary sites were excluded. The study evaluated the performance of three ML models - decision trees (DT), random forest (RF), and support vector machine (SVM) along with logistic regression (LR). The predictors included patient demographic and clinical characteristics. for 5-year survival. The classification performance of ML models was evaluated based on accuracy, the area under the receiver operating characteristic curve (AUROC), and the F1 score.

RESULTS: The study cohort included 54,263 patients with HNSCC. Most of the selected patients were White (84.3%), male (74.0%), aged ≥65 years (44.9%), and had oropharyngeal cancer (9.4%). The machine learning algorithms under study were able to predict the overall survival of HNSCC patients. DT outperformed RF, SVM, and LR in terms of AUROC score (DT:72%, RF:70%, SVM:66%, LR:67%). SVM and LR had the highest accuracy followed by DT and RF (SVM:70%, LR:70%, DT:69%, RF:67%) while having comparable F1 scores (SVM:56%, LR:57%, DT:55%, RF:56%).

CONCLUSIONS: Overall, the performance of all ML models varied with better performance with DT based on the AUROC. However, the F1 scores were similar for the three ML models. More work is needed to evaluate the ML models involving external validations for further considerations in healthcare.