Immune-related adverse event prediction and influential factor identification.

Abstract

e14650 Background: Immune checkpoint inhibitors (ICIs), have demonstrated efficacy in treating patients with non-small cell lung cancer (NSCLC) and advanced melanoma. However, ICIs can lead to immune-related adverse events (irAE), which will cause discontinuance of the ICI treatment. The objective of this study is to identify factors that predict which patients with NSCLC or melanoma will develop irAE(s) from ICIs. Methods: We curated the patient cohort from the Immune-Oncology data repository developed at the Georgetown-Lombardi Comprehensive Cancer Center. This cohort encompassed 466 patients (245 individuals with NSCLC and 221 with advanced melanoma) who received single-agent anti-PD-(L)1 or combination of anti-PD-1 and anti-CTLA4 therapy. Parameters selected for prediction comprises baseline demographics, laboratory results, treatment information, and cancer-related mutations. The classification of irAE were identified by CTCAE V4.03 guidelines. Overall, 219 patients experienced one or more irAE, while the remaining 247 patients were irAE-free. We employed machine learning models for prediction, which were trained using 80% of the dataset with a five-fold cross-validation, and their performance was evaluated using the area under the receiver-operating curve (AUROC) and F1 score. Furthermore, we identified the key factors that significantly impacted the model by SHAP values. Results: Nine machine learning models were deployed, including logistic regression, support vector machine, gradient boosting (GB), etc. The GB model demonstrated the highest performance, achieving an AUROC score of 75% and an F1 score of 77%. According to the SHAP value obtained from the GB model, NSCLC patients exhibited a higher propensity for irAE development compared to melanoma patients. The top three features associated with a heightened likelihood of irAE development are the treatment of the combination of anti PD-1 and anti CTLA-4, a higher albumin/globulin ratio, and a pre-treatment ECOG Performance Status score of 0. The top three features associated with a lower likelihood of irAE development are an elevated red cell distribution width, the treatment with single agent anti-PD-1, and a higher globulin level. Conclusions: Our research harnesses a comprehensive set of variables to forecast the development of irAEs in NSCLC and melanoma patients undergoing ICI treatment. The results underscore the predictive capabilities of ML models. Investigating the connections between pivotal biomarkers, such as globulin and red cell distribution width, has the potential to offer insights that enable patients to mitigate irAE. These findings emphasize the promising potential of ML techniques not only in predicting the likelihood of irAE but also in the grade of irAE. To mitigate the limitation of a relatively small sample size, we are actively working to expand the data registry by collaborating with additional institutions.