Abstract BACKGROUND Immunotherapy has been successful in a small subset of patients with brain cancer, glioblastoma. One of the reasons for failure are difficulty in identifying responders which has not been possible at the earliest stage of diagnosis and require surgery to obtain tumor tissue for analysis to determine immune biomarkers that can be utilized for patient selection for immunotherapy. To address it, we aim to identify immune related radiological biomarkers and design a radiogenomic machine learning model that can predict immune status within the tumor microenvironment of IDH-wildtype glioblastoma for patient stratification in immunotherapy trials at the time of diagnosis, prior to surgery. MATERIAL AND METHODS This was a retrospective multicenter machine learning based study utilizing curated open-access anonymized-matched radiogenomic datasets. Imaging data consisted of MRI based radiomic features extracted from deep learning-based auto-segmented glioblastoma tumors. Immune scores and immune cell-specific scores were extracted from the transcriptomic data using pan cancer and glioblastoma immune signature matrices. AI models were trained and validated with 10-fold stratified nested cross validation techniques with overall immune scores and cell-specific immune scores as a gaussian binary label. Inter-dataset leave-one-out cross validation was performed to assess the generalizability among different sources. The outlier flagging model was trained from the datasets. RESULTS One-hundred-and-seventy-six patients were included in the study. The immune-related radiomic signatures obtained after feature selection and employed for prediction models were first order and second order radiomic features. For the pan cancer signature matrix, stratified 10-fold cross validated ensemble models predicting overall immune scores, T cells, TAMs, DC and NK cell scores achieved mean accuracies of 0.70, 0.82, 0.81, 0.63 and 0.89, with high mean specificity and mean negative predictive values. The models demonstrated better accurate low score predictions. For a glioblastoma-specific signature matrix, ensemble models showed higher mean accuracies ranging from 0.74 to 0.84 for different immune cell signatures. LOOCV demonstrated that models were able to handle real-world hold out datasets with accurate predictions. Outlier identification modelling resulted in flagging of these anomalous cases along with the predictions. CONCLUSION The radiogenomic AI models non-invasively predicted immune status in IDH-wildtype glioblastoma. The models have the potential to stratify patients for immunotherapy within prospective glioblastoma clinical trials. We recommend that initially, the models should be used to prospectively predict response in immunotherapy clinical trials as an observer tool and then for the prediction to be unblinded after the trial is complete.
Ghimire et al. (Wed,) studied this question.