Objective  To explore the value of feature selection and subgroup analysis based on radiomics in predicting prognosis of patients with lung adenocarcinoma.

  Methods  A total of 293 lung adenocarcinoma patients with radiotherapy were selected, and 107 radiological features (14 shape features, 18 first-order statistical features and 75 texture features) were extracted from chest CT images. The effects of three different feature selection (FS) methodsretest and multiple segmentation (FS1), Pearson correlation analysis (FS2) and the combination of FS1 and FS2 (FS3)on survival prediction performance were analyzed. Subgroup analysis was performed for each T stage, and the prognostic performance was evaluated by consistency index (C-index) and Kaplan-Meier method. Quintuple cross validation was used in subgroup analysis to ensure the reliability of the model.

  Results  In the training and test data sets of radiology model, the C-index of FS2 was the highest among all the selection methods (values were 0.64 and 0.61 respectively). Similarly, FS2 showed the highest C-index (values were 0.65 and 0.63, respectively) among all the selection methods in the training and test data sets of the combined model. Subgroup analysis showed that the C-index of the prediction model based on T stage was higher than that based on the full data.

  Conclusion  Feature selection method improves the performance of survival prediction to a certain extent, and the subgroup prediction model based on T stage can improve the prediction performance.