Proportional Hazards Assumption of Radiomic Features in Survival Analysis

Abstract

Survival analysis is often used in radiomic studies for predicting prognosis and recurrence. A typical model for the survival analysis is Cox regression. An essential assumption when using the Cox regression model is proportional hazards. However, the proportional hazards assumption has not been verified in a lot of radiomic studies so far. In this study, we investigated the existence ratio of radiomic features that did not satisfy the proportional hazards in breast cancer and glioblastoma. Additionally, we proposed a hypothesis test of proportional hazards and an evaluation method using a scatter plot. The experimental results showed that radiomics features did not satisfy the proportional hazards assumption in 6.5% (24/369) of breast cancer and 21.4% (79/369) of glioblastoma. When the phenotype of tumor changed in a short period of time, it was found that radiomic features often did not satisfy the proportional hazards assumption. Therefore, in such radiomic studies, it is necessary to test the hypothesis of proportional hazards.