Abstract 3668: AI-based models integrating serum proteomics and clinicopathologic features to predict survival outcomes in NSCLC patients undergoing immunotherapy

Abstract Background: Responses to immune checkpoint inhibitors (ICIs) vary among patients with NSCLC. Combining serum proteomics with clinical data offers a novel approach to individualized outcome prediction. Methods: This study analyzed 148 consecutive NSCLC patients treated with ICI monotherapy or ICI-chemotherapy combinations. Serum proteomics quantified via mass spectrometry were integrated with clinicopathological data (age, gender, smoking history, cancer stage, performance status, tumor mutational burden (TMB), PD-L1 IHC, and treatment regimen). Machine learning models(XGBoost, logistic regression, and multilayer perceptron) were used with single and multimodal approaches and were evaluated using 5-fold cross-validation. Primary outcomes were progression-free survival (PFS) and overall survival (OS) at 90, 180, 270, and 360 days. Performance was assessed using the area under the receiver operating characteristic curve (AUC). Results: The cohort's median age was 71 years, with 52.0% being female. 14.2% never smoked, 71.0% former smokers, and 14.2% current smokers. Histologically, 68.2% had adenocarcinoma, 27.7% squamous cell carcinoma, and 4.1% other types. Performance status was 0 in 20.3%, 1 in 54.7%, 2 in 17.6%, and 3 in 6.1%. Most patients had stage IV disease (82.1%). Median TMB was 5.4 mut/Mb; PD-L1 >1% in 56.8%, <1% in 29.7%. Treatment regimens included mono-ICI (35.8%), dual ICI (14.2%), and ICI + chemotherapy (50.0%). Median OS and PFS were 19.5 and 7.1 months, respectively. Proteomics-based models outperformed other approaches for PFS prediction, with XGBoost achieving a peak AUC of 0.67 at 3 months and maintaining strong performance through 6 months (LogReg, AUC=0.65) and 9 months (XGBoost, AUC=0.64). For OS prediction, model performance improved over time, reaching the highest AUC of 0.61 with XGBoost using proteomics at 360 days. Risk groups were stratified at 9 months OS using XGBoost regression, leveraging 4, 000 highly variable proteomic features identified by mass spectrometry (m/z). Significant differences in PFS between high- and low-risk groups were confirmed by a log-rank test (p < 0.01) and Cox proportional hazards analysis (HR = 1.72, 95% CI: 1.18-2.51, p < 0.01), with median PFS of 166 days for the high-risk group and 297 days for the low-risk group. Conclusions: The AI-based models demonstrated moderate to strong discrimination, with AUCs ranging from 0.70 to 0.77 for the best-performing models. These results highlight the potential of multimodal machine learning approaches for personalized prediction of outcomes with immunotherapy in NSCLC. Citation Format: Junho Song, Leeseul Kim, Donghoon Shin, Jongwoo Kim, Taegyu Um, Allen Cho, Jaeyoung Bae, Liam IL-YOUNG Chung, Jeeyeon Lee, Young Kwang Chae. AI-based models integrating serum proteomics and clinicopathologic features to predict survival outcomes 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 3668.