1246 Prediction of pneumonitis in immunotherapy patients from prior thorax CT

<h3>Background</h3> Pneumonitis, or inflammation of the lung tissue, is one of the most significant potential side effects of immune checkpoint inhibitor (ICI) therapy. Predicting pneumonitis could improve both the safety and potential efficacy of ICI, since treatments could be administered for longer and discontinued before intolerable toxicity. <h3>Methods</h3> We curated 2700 pre-treatment CT lung volumes for 671 patients (509 negative and 162 positive for pneumonitis) who received prior ICI therapy at Vanderbilt University Medical Center.¹ We manually curated pneumonitis occurrence and ICI treatment dates for each patient. Patients who experienced treatment-impacting pneumonitis during ICI therapy, regardless of etiology, were labeled positive. Three different imaging-only models were built for predicting pneumonitis using a single, pre-treatment CT. The first model used radiomics alone; the second used deep learning (DL); and the final model combined the two (figure 1). The radiomics model used 109 radiomics features² plus the patient age and CT age (time between CT acquisition and first ICI treatment) from the DICOM metadata. A random forest classifier on the 111 features performed the final, patient-level pneumonitis prediction. The DL architecture was based on EfficientNet-B0.³ Each volume was examined slicewise, and the distribution of predictions for each slice was aggregated using the mean, median, minimum, maximum, standard deviation, skewness, and count and then combined with patient age and CT age to input nine features to a patient-level random forest classifier. The third model combined the 109 radiomics features plus the 7 DL features with patient age and CT age. These features were input into a patient-level random forest classifier that was trained with identical parameters to the two separate models. Ten-fold cross validation was used with identical fold assignments across models, and 95% confidence intervals for area under the receiver operating characteristic curve (AUC) were computed from fold AUC distributions. <h3>Results</h3> The first radiomics only model achieved an AUC of 0.747 (0.705–0.789 CI). The CNN-only model was a significant improvement to an AUC of 0.819 (0.781–0.857 CI). The combined CNN and radiomics model achieved a slightly improved AUC of 0.829 (0.797–0.862 CI). <h3>Conclusions</h3> Our results demonstrated the performance of deep learning for pre-treatment prediction of pneumonitis in ICI-treated patients. We found that deep learning outperformed radiomics by a significant margin and captured virtually all of the information in the radiomics features. Pre-treatment imaging models could promise better management for patients receiving ICI and the ability to better stratify patients in cancer immunotherapy drug trials. <h3>Acknowledgements</h3> We thank the data curation team for their work in curating data, including (alphabetical) Ryan Ahmed, Madeline Ball, Mary Banasiewicz, Kimberly Fields, Adara Holland, Marilyn E. Holt, Steven Houtschilt, Alexandria Manis, Christina Noel Maxwell, Lilyanna McArthur, Lori Michalowski, Margaret Mitchell, Ndidiamaka Odinkemelu, Rhonda Potter, Joyce Raisan, Andrew Sayce, Timothy Schurz, Alysse Sephel, Marina Shannon, Morgan Shannon, Jarrod Smith, Susan Sommers, Samuel Trump, Stephanie Winchell, and Annie Zwaschka. We also thank (alphabetical) Lara Franck, Lucy Wang, and the Vanderbilt Coordinating Center for their contributions. <h3>References</h3> Lippenszky L,<i> et al</i>. Prediction of effectiveness and toxicities of immune checkpoint inhibitors using real-world patient data.<i> JCO Clin Cancer Inform</i>. 2024;<b>8</b>:e2300207. Van Griethuysen JJ,<i> et al</i>. Computational radiomics system to decode the radiographic phenotype.<i> Cancer Research</i>. 2017;<b>77</b>:e104–e107. Tan M, Le QV. EfficientNet: rethinking model scaling for convolutional neural networks. arXiv:2020;1905.11946. <h3>Ethics Approval</h3> Vanderbilt University Medical Center Institutional Review Board Study Approval #211814.