Clinical Parameters Combined With Radiomics Features to Predict the Efficacy of Immunotherapy for Advanced Non-Small Cell Lung Cancer.

PURPOSE/OBJECTIVE(S) Many tools have been developed to predict the efficacy of immunotherapy, such as LIPI, EPSILoN and mLIPI scores. The aim of this study was to determine the ability of to predict outcomes in Chinese aNSCLC patients treated with ICIs. With the development of imaging technology, radiomics was increasingly used in the differentiation of benign and malignant lung nodules, the pathological typing of lung cancer, and the prediction of lung cancer prognosis. We use 3D-slicer to deline at the region of interest in the patient's CT image, extract a large number of features, and further screen out the radiomics features that have predictive value for the outcome. MATERIALS/METHODS We retrospectively enrolled 317 patients with histologically proven aNSCLC (IIIB-IV) treated with ICIs. The discriminative ability of the predictive models was evaluated by AUC in the ROC analysis. Patients were randomly divided into training and validation cohorts using a 2:1 ratio. We used the semi-automatic segmentation method of the 3D-slicer platform to delineate the ROI of the tumor's lesion area and extract 854 image features for each patient. In the training set of patients, the LASSO algorithm was used to screen the radiomic features. Hosmer-Lemeshow tests (H-L tests) were conducted to determine the fit of the prediction models. PFS and OS curves were generated using the Kaplan-Meier method and differences were assessed using log-rank tests. Univariate and multivariate analyses were performed using Cox proportional-hazards regression models. The glmnet R package was used for the LASSO regression method. The rms, Hmisc R packages used for C-index. RESULTS Among the 317 patients included in the study, the median OS and PFS were 14.2 months and 5.6 months, respectively and the ORR was 23.1%. The AUC values of LIPI, mLIPI, and EPSILoN scores for predicting PFS were 0.649 (95% CI: 0.588-0.709), 0.765 (95% CI: 0.713-0.818]), and 0.637 (95% CI: 0.567-0.698), respectively (P < 0.001 for all models). The AUC value of mLIPI scores was significantly higher than that of LIPI and EPSILoN scores (P < 0.05). The C-index of mLIPI was 0.645 (95% CI: 0.617-0.673). In this study, 5 radiomics features with predictive value were selected from radiomics features. The C-indexes of the radscore were 0.643 (95% CI: 0.602-0.684) and 0.632 (95% CI: 0.571-0.693). Then we combined mLIPI and radscore to obtain a mixed model. The C-index of the combination mLIPI scores with radscore for predicting PFS was 0.810 (95% CI: 0.770-0.849) and 0.706 (CI: 0.633-0.778) in the training and validation cohorts, respectively. CONCLUSION By externally validating LIPI, mLIPI, and EPSILoN scores, we found that all three of these predictive models could identify different prognostic subsets of patients treated with ICIs to statistically significant degrees. We also found that mLIPI had the highest accuracy among the three models. With the addition of radiomics features, the prediction performance of the mixed model has been further improved.