An online tool for survival prediction of extrapulmonary small cell carcinoma with random forest

PURPOSE: Extrapulmonary small cell carcinoma (EPSCC) is rare, and its knowledge is mainly extrapolated from small cell lung carcinoma. Reliable survival prediction tools are lacking. METHODS: A total of 3,921 cases of EPSCC were collected from the Surveillance Epidemiology and End Results (SEER) database, which form the training and internal validation cohorts of the survival prediction model. The endpoint was an overall survival of 0.5–5 years. Internal validation performances of machine learning algorithms were compared, and the best model was selected. External validation (n = 68) was performed to evaluate the generalization ability of the selected model. RESULTS: Among machine learning algorithms, the random forest model performs best on internal validation, whose area under the curve (AUC) is 0.736–0.800. The net benefit is higher than the TNM classification in decision curve analysis. The AUC of this model on the external validation cohort is 0.739–0.811. This model was then deployed online as a free, publicly available prediction tool of EPSCC (http://42.192.80.13:4399/). CONCLUSION: This study provides an excellent online survival prediction tool for EPSCC with machine learning and large-scale data. Age, TNM stages, and surgery (including potential performance status information) are the most critical factors for the prediction model.