Utilizing predictive modeling to identify patients at high risk of death from immune checkpoint inhibitor therapy-related pneumonitis

1552 Background: Immune checkpoint inhibitor therapy-related (ICI) pneumonitis poses a significant challenge in patients with cancer. There is a need for predictive modeling to identify those at high risk of developing serious adverse events of ICI pneumonitis. Methods: A newly created database of cancer patients diagnosed with ICI pneumonitis seen at Mayo Clinic from 2014-2022 was utilized for exploratory data analysis. Modeling was used to determine if three clinical outcomes could be predicted, namely risk of death from: pneumonitis prior to starting immunotherapy, pneumonitis at time of diagnosis of ICI pneumonitis, and any cause at time of ICI pneumonitis diagnosis. The data consisted of 170 patients with annotation for 71 clinical features. Data was divided into features available at time of cancer diagnosis and at time of ICI pneumonitis diagnosis. Exploratory modeling was performed using the gradient boosting technique Xgboost (Chen, 2015) and conducted using k-fold balanced cross validation best practices as well as a train/test/validate schema with 70%/20%/10% data proportions, respectively. Model reverse engineering was done with Shapley statistics (Lundberg, 2017) to determine which features had the largest contribution per model. Once identified, only those highly weighted features were used for logistic regression analysis providing more reproducible predictions by decreasing model variance. Results: Risk of death from pneumonitis prior to starting immunotherapy was modeled with an area under the curve of the receiver operator characteristic (AUC-ROC) of 0.79 with the most contributory features including lymphocyte count, oxygen dependence, PFT values, and PD-L1 expression. Logistic regression produced an AUC-ROC of 0.87 (95% CI 0.72-1.0, p < 0.0013). Risk of death from pneumonitis at the time of ICI pneumonitis diagnosis was modeled with an AUC-ROC of 0.85 and the most contributory features were similar to the list described in the previous endpoint. Logistic regression produced an AUC-ROC of 0.89 (95% CI 0.81-0.96, p < 0.0001). Risk of death from any cause at the time of ICI pneumonitis diagnosis until follow up concluded by 12/2022 produced a model with an AUC-ROC of 0.75 with the most contributory features including supplemental oxygen, PFT values, basic laboratory values, and PD-L1 expression, among others. Logistic regression produced an AUC-ROC of 0.85 (95% CI 0.76-0.93, p < 0.0001). Conclusions: We demonstrate that commonly available clinical data can be used to identify patients at high risk of death from ICI pneumonitis. This study identified clinical features that were predictive in each scenario from which further concerted effort could produce a new clinical model to provide clinician decision support when considering immunotherapy. Further studies should be done to further elucidate feature interdependence and generalizability.