Superior prognosis prediction performance of deep learning for gastric cancer compared to Yonsei prognosis prediction model using Cox regression.

Abstract

164 Background: Staging systems for cancer are critical to predict the prognosis of patients. Current staging systems for gastric cancer have limitations to predict individualized and precise prediction of patient’s survival after treatment. We aimed to develop prediction model based on deep learning by estimating the survival probability of patients who underwent gastrectomy. Methods: To predict the survival probability, we used a deep neural network model which consisted of 5 layers: input layer, 3 fully connected layer, and output layer with 8 characteristics (age, sex, histology, depth of tumor, number of metastatic and examined lymph node, presence of distant metastasis, and resection extent) of patients which was previously published Yonsei prediction model using Cox regression. Each layer functioned as the nonlinear weighted sum of lower layer. Five-year overall survival was predicted using the deep learning method and it was compared to Yonsei prediction model. The average area under the curve (AUC) was compared between the models. For internal validation, 5-fold cross validations were carried out. We also performed external validation with a dataset from another hospital (n = 1549). . Results: Deep learning predicted 5-year overall survival of patients with an average accuracy of 83.5% in the test set. The average AUC of deep learning by integrating 8 characteristics was significantly higher than that of Yonsei prediction model (AUC: 0.844 vs. 0.831, P < 0.001) with the same variables. In the external validation the average accuracy of survival prediction was 84.1%. The AUC was also greater in a dataset from other hospital in Korea (AUC: 0.852 vs. 0.847, P = 0.023) Conclusions: Prognosis prediction with deep learning showed superior survival predictive power compared to prediction model using Cox regression. It can provide individualized and precise stratification based on the risk using characteristics of gastric cancer patients.