The prognosis of esophageal squamous cell carcinoma (ESCC) is very poor and the overall survival rate is very low, which is a major threat to human health. Using preoperative pathological information to predict prognostic risk is of great significance for improving prognosis evaluation. Aiming at the problem that the characteristic variables extracted by different modeling methods are different and the stability is not strong, this paper proposes a scheme that fuses the idea of ensemble learning with feature selection, establishes the final survival risk prediction model for ESCC. Five classification techniques are integrated to rank the importance of risk factors, including stochastic configuration network (SCN), Logistic regression (LR), support vector machine (SVM), decision tree (DT) and random forest (RF). The selected important risk factors as predictors of SCN, LR, SVM, DT and RF to construct ESCC prognostic risk assessment models. The results showed that INR, APTT and albumin are important factors affecting the prognosis of ESCC. The classification accuracy of SCN, LR, SVM, DT and RF models established by this scheme are all higher than that without using this scheme. Among them, the RF model established has the best performance (accuracy: 80.19%).