ESTRO 2020 Abstract book

S322 ESTRO 2020

label adjust by SMOTE approach. (2) feature dimension reduction by calculating correlation with tumour volume, feature pair-wise correlation, and Kolmogorov-Smirnov test. (3) LASSO was used to build feature pool. (4) recursive feature elimination was used to build signature pool using logistic regression. (5) optimal model selection. The AUC and Hosmer-Lemeshow test statistic were determined to assess the model discrimination and calibration, respectively.

Results At two year following RT, there was similar proportion of surviving to deceased patients on the in-train and in-test subsets). After implementing the proposed feature selection process for survival prediction, a signature was developed with 6 radiomic features. The signature achieved AUCs of 0.87 (95: CI:0.80-0.93) and 0.84 (95: CI:0.71-0.88) in the in-train and in-test subsets (Figure 2). The difference in AUCs between subsets were not statistically significant, as can be expected from the highly-overlapping confidence interval estimates. The Hosmer-Lemeshow test of the two-year survival prediction model yielded non-significant statistics (p=0.10 and p=0.22), indicating that deviation of model prediction from observed outcome was not statistically significant.

Conclusion We proposed a pooling-based feature selection method for quantitative imaging analysis. Furthermore, we show how it would be applied to the clinical question of predicting 2-year survival in patients treated for glioma by RT. Second, we demonstrate that CT images of glioma patients might have some relevant information to survival, as the developed CT radiomics-based signature was able to predict overall survival. There are two main limitations: