A Random Forest Model for Post-Treatment Survival Prediction in Patients with Non-Squamous Cell Carcinoma of the Head and Neck

Background: Compared to squamous cell carcinoma, head and neck non-squamous cell carcinoma (HNnSCC) is rarer. Integrated survival prediction tools are lacking. Methods: 4458 patients of HNnSCC were collected from the SEER database. The endpoints were overall survivals (OSs) and disease-specific survivals (DSSs) of 3 and 5 years. Cases were stratified–randomly divided into the train & validation (70%) and test cohorts (30%). Tenfold cross validation was used in establishment of the model. The performance was evaluated with the test cohort by the receiver operating characteristic, calibration, and decision curves. Results: The prognostic factors found with multivariate analyses were used to establish the prediction model. The area under the curve (AUC) is 0.866 (95%CI: 0.844–0.888) for 3-year OS, 0.862 (95%CI: 0.842–0.882) for 5-year OS, 0.902 (95%CI: 0.888–0.916) for 3-year DSS, and 0.903 (95%CI: 0.881–0.925) for 5-year DSS. The net benefit of this model is greater than that of the traditional prediction methods. Among predictors, pathology, involved cervical nodes level, and tumor size are found contributing the most variance to the prediction. The model was then deployed online for easy use. Conclusions: The present study incorporated the clinical, pathological, and therapeutic features comprehensively and established a clinically effective survival prediction model for post-treatment HNnSCC patients.