Radiomics to predict checkpoint inhibitor pneumonitis based on chest CT scans in patients with advanced cancer.

person Wim Vos Radiomics, Liege, Belgium info_outline Wim Vos, François Cousin, Mathieu Wendling, Pierre Freres, Julien Guiot, Mariaelena Occhipinti, Roland Hustinx

Authors person Wim Vos Radiomics, Liege, Belgium info_outline Wim Vos, François Cousin, Mathieu Wendling, Pierre Freres, Julien Guiot, Mariaelena Occhipinti, Roland Hustinx Organizations Radiomics, Liege, Belgium, University Hospital (CHU) of Liège, Liege, Belgium, CHU de Liege, Liege, Belgium, Department of Pneumology, University Hospital of Liège, Liege, Belgium, Department of Nuclear Medicine and Oncological Imaging, University Hospital of Liège, Liege, Belgium Abstract Disclosures Research Funding No funding received None. Background: Checkpoint inhibitor pneumonitis (CIP) is a concerning immune-related adverse event responsible for treatment discontinuation, hospitalization, with acknowledged possible fatal cases. The aim of our study was to assess the potential role of machine learning models, combining pre-treatment CT lung radiomic features with clinical data, to predict the risk of developing CIP. Methods: A cohort of 116 patients with different type advanced malignancies (81% lung cancer, 14% melanoma, 8% other solid tumors) treated with immune checkpoint inhibitors (ICIs) were retrospectively enrolled in this study and divided into a CIP (n = 35) and a control group (n = 81). Each patients had a chest CT scan at baseline and associated clinical covariates. Lungs and lobes were segmented semi-automatically on pre-treatment CT scans. We extracted radiomics features and selected them by the Minimum Redundancy Maximum Relevance (mRMR) method. Clinical-only, radiomics-only, and clinical-radiomics models (both lung- and lobe-based) were built using generalized linear (GLM) and random forest (RF) models to predict the occurrence of CIP. A nested 5-fold cross-validation scheme was used in the training, validation and test sets. Model performances were assessed in terms of Area under the Curve (AUC). Results: The most accurate clinical (RF) and the best radiomic (lobe-based GLM) models achieved AUCs of 0.72 (95%CI:0.51-0.92) and 0.71 (95%CI:0.58-0.84) on the test set, respectively. No statistically significant differences have been found between lung- and lobe-based radiomic models. The RF model combining lung-based radiomic and clinical features showed the best performances overall with an AUC of 0.75 (95%CI:0.62-0.88) on the test set. Conclusions: CT-based lung radiomic, and clinical models alone show moderate to good performances at identifying patients at risk of developing CIP. Nevertheless, we demonstrated the added value of machine learning models built combining clinical and radiomic features to better stratify the risk of developing CIP in patients with advanced malignancies treated with ICIs.