Abstract 3667: AI-based models integrating serum proteomics and clinical features to predict immune-related adverse events in NSCLC patients undergoing immunotherapy

Abstract Background: Immune checkpoint inhibitors (ICIs) improved outcomes in NSCLC treatment, but causing irAEs. Serum proteomic biomarkers combined with clinical data may offer a better approach for toxicity risk prediction. Methods: To predict five irAE types (pneumonitis, rash, arthritis, thyroiditis, and other adverse events) at multiple time points, pre-treatment serum samples and clinical data (age, gender, smoking history, cancer stage, performance status, and treatment regimen) were analyzed from consecutive 173 NSCLC patients treated with ICI monotherapy or ICI-chemotherapy combination. Serum proteomics was performed via mass spectrometry, and three machine learning models (XGBoost, HistGradientBoosting, and logistic regression) were evaluated using 5-fold cross-validation. Primary outcomes were five toxicity types evaluated at 1, 3, 6 months, and 1 year. Model performance was assessed using the area under the receiver operating characteristic curve (AUC). Results: The median age was 70.7 years, with 50.9% female and 66.5% adenocarcinoma. 16.2% had never smoked, 68.2% former smokers, and 15.0% current smokers. Performance status (PS) was 0 in 17.9%, 1 in 55.5%, 2 in 20.2%, and 3 in 5.2%. The median tumor mutational burden (TMB) was 5.4 mut/Mb, with PD-L1 negative in 30.1%, positive in 58.4%. Treatment regimens included mono-ICI (35.3%), dual ICI (12.7%), and ICI + chemotherapy (52.0%). 91.3% received palliative treatment. Median overall survival (OS) was 20.9 months, progression-free survival (PFS) was 7.1 months, and toxicity-free survival was 35.2 months. The multimodal analysis demonstrated distinct predictive patterns across toxicity types and timepoints. At 30 days, combined modality approaches performed best: Logistic Regression for pneumonitis (AUC=0.60), HistGradientBoosting for rash (AUC=0.60), and XGBoost for arthritis (AUC=0.60). Feature importance analysis identified KRAS mutation status and smoking history as primary predictors for pneumonitis, ECOG status and histology type for rash, disease stage and treatment regimen for arthritis, and PD-L1 expression for other adverse events. Proteomics-based models performed best at 3 months for pneumonitis (AUC=0.60) and thyroiditis (AUC=0.57). Notably, the HistGradientBoosting model achieved superior performance (AUC=0.75) for overall irAE prediction at 9 months using both proteomics and clinical features, demonstrating the value of integrated multimodal analysis. Conclusions: The models showed moderate discrimination, with AUCs ranging from 0.556 to 0.75, particularly excelling in early toxicity prediction. These findings support multimodal machine learning for personalized toxicity risk assessment and proactive management. Citation Format: Junho Song, Leeseul Kim, Donghoon Shin, Jongwoo Kim, Taegyu Um, Allen Cho, Jieun Park, Liam IL-YOUNG Chung, Jeeyeon Lee, Young Kwang Chae. AI-based models integrating serum proteomics and clinical features to predict immune-related adverse events in NSCLC patients undergoing immunotherapy [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2025; Part 1 (Regular Abstracts); 2025 Apr 25-30; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2025;85(8_Suppl_1):Abstract nr 3667.