AI-based radiomics model for predicting immune checkpoint inhibitor–related pneumonitis (CIP) in patients with advanced NSCLC: An external validation study.

12136 Background: The growing implementation of immunotherapy in advanced non-small cell lung cancer (NSCLC) management has led to an increase in adverse events, notably immune checkpoint inhibitor-related pneumonitis (CIP). CIP, a significant and potentially fatal complication, often necessitates the discontinuation of immunotherapy, thereby impacting patient outcomes severely. The absence of biomarkers for early detection and management of CIP represents an urgent and unmet clinical need. This expanded study explores the potential of Artificial Intelligence (AI) algorithms in predicting CIP in NSCLC patients undergoing immunotherapy from pre-treatment CT scans. Methods: A cohort of 220 stage III-IV NSCLC patients receiving immunotherapy was considered. The patients were divided into a training set (D1, n=105, Institution A), an internal validation set (D2, n=45, Institution A), and an external validation set (D3, N=70, Institutions B and C). Manual delineation of the tumor on baseline CT was performed by three physicians working in consensus. The Picture Health Px platform was employed for AI-powered deep phenotyping of the tumor and its surrounding habitat, including segmentation of the tumor associated vasculature and featurization. Quantitative features relating to the twistedness of the tumor-associated vasculature and tumoral heterogeneity patterns were extracted. These features were used to train a neural network classifier for CIP prediction. Weighting techniques during training were used to compensate for the rarity of CIP. A threshold for CIP classification was set within D1 and applied to the testing sets. Results: 43.6% of patients received an immunotherapy-only regimen and 56.4% received combined immunochemotherapy. Of these, 18.2.% experienced pneumonitis events, with 50.0% being CIP. The CIP subgroup had 50.0% grade 1 CIP, 30.0% grade 2, and 20.0% grade 3. The cross-validated AUC on D1 was 0.76 (95% CI: 0.62-0.90). The AUC was 0.61, 0.62, and 0.64, respectively, on D2, D3, and all validation data combined (D2+D3). The model correctly identified 71.4% of grade 1/2 CIP events and 33.3% of grade 3 CIP, with a corresponding false positive rate of 33.0%. Severe/grade 3 IO-related pneumonitis prediction was likely limited by the small number of severe cases from the training institution (n=1). Conclusions: We demonstrated a radiomic AI signature could identify patients at risk of CIP prior to treatment across multiple external institutions. This approach could improve the identification and management of CIP in NSCLC patients on immunotherapy, thereby enhancing patient outcomes. Expanded training and validation to account for rare, but possibly fatal, high grade CIPs is needed to expand the clinical benefit of radiologic predictors of pneumonitis to the most severe cases.