Deep learning approach for survival prediction for patients with synovial sarcoma

Abstract

Synovial sarcoma is a rare disease with diverse progression characteristics. We developed a novel deep-learning-based prediction algorithm for survival rates of synovial sarcoma patients. The purpose of this study is to evaluate the performance of the proposed prediction model and demonstrate its clinical usage. The study involved 242 patients who were diagnosed with synovial sarcoma in three institutions between March 2001 and February 2013. The patients were randomly divided into a training set (80%) and a testing set (20%). Fivefold cross validation was performed utilizing the training set. The test set was retained for the final testing. A Cox proportional hazard model, simple neural network, and the proposed survival neural network were all trained utilizing the same training set, and fivefold cross validation was performed. The final testing was performed utilizing the isolated test data to determine the best prediction model. The multivariate Cox proportional hazard regression analysis revealed that size, initial metastasis, and margin were independent prognostic factors. In fivefold cross validation, the median value of the receiver-operating characteristic curve (area under the curve) was 0.87 in the survival neural network, which is significantly higher compared to the area under the curve of 0.792 for the simple neural network (p = 0.043). In the final test, survival neural network model showed the better performance (area under the curve: 0.814) compared to the Cox proportional hazard model (area under the curve: 0.629; p = 0.0001). The survival neural network model predicted survival of synovial sarcoma patients more accurately compared to Cox proportional hazard model.