Document resource
Background ICI-related pneumonitis is a severe, potentially life-threatening toxicity, occurring in approximately 3.5%-19% of ICI therapy cases and accounting for 35% of ICI-related deaths in patients under anti-programmed cell death ligand-1(PD-L1) therapy. 1 Real-world studies have demonstrated higher incidence rates than initially reported in clinical trials, with some cohorts showing up to 19% pneumonitis rates.1 Recent advances in AI and radiomics have shown promise for predicting immunotherapy-induced adverse events using pretreatment imaging.2 Identifying patients at high risk before treatment initiation is a critical unmet need. In this study, we aimed to develop a prediction model for IRP in NSCLC patients by integrating baseline radiological features of the tumor and the surrounding lung parenchyma.Methods We retrospectively analyzed 102 patients with stage IIIB/IV NSCLC treated with ICIs. IRP was identified and distinguished from radiation-related pneumonitis via medical chart review. Radiomic features were extracted from primary tumors on baseline pre-ICI CT scans using computational image analysis. These were combined with clinical variables including demographics, PD-L1 expression, and genomic profiling data. Feature selection identified the most discriminative variables. An ensemble machine learning approach was developed combining multiple algorithms. Stratified 5-fold cross-validation was employed for the low event rate. Performance was evaluated using AUC and NPV. Risk stratification was performed using optimal cutpoint analysis.Results Among 102 patients (median 66 years, IQR 59.5-73.5), 13 (12.7%) developed IRP during median follow-up of 14 months (IQR 5-30). The ensemble radiomics model achieved mean cross-validation AUC of 0.72 (95% CI: 0.61-0.83, figure 1), with 53.8% sensitivity, 88.8% specificity, 41.2% PPV, and 92.9% NPV with the majority of low-risk patients remaining pneumonitis-free throughout follow-up. Deep learning-derived radiomic signatures, morphological intensity metrics, neutrophil-to-lymphocyte ratio, and baseline bone metastasis emerged as key predictors. Risk stratification revealed distinct groups: high-risk patients (n=26, 25.5%) experienced 26.9% IRP incidence versus 7.9% in low-risk patients (n=76), with hazard ratio 3.58 (95% CI: 1.20-10.66, p=0.014, figure 2).Conclusions Our radiomics model analyzing tumor and surrounding lung on baseline CT can identify NSCLC patients at elevated IRP risk. The model’s high NPV (92.9%) reliably identifies low-risk patients, while significant hazard separation enables risk-adapted monitoring strategies. This non-invasive screening tool could facilitate targeted surveillance and timely interventions, improving immunotherapy safety. A follow-up study with a larger cohort is ongoing for validation of these results.References Suresh K, Voong KR, Shankar B, Forde PM, Ettinger DS, Marrone KA, Kelly RJ, Hann CL, Levy B, Feliciano JL, Brahmer JR, Feller-Kopman D, Lerner AD, Lee H, Yarmus L, D’Alessio F, Hales RK, Lin CT, Psoter KJ, Danoff SK, Naidoo J. Pneumonitis in non-small cell lung cancer patients receiving immune checkpoint immunotherapy: incidence and risk factors. J Thorac Oncol. 2018;13:1930–1939.Braman N, Prasanna P, Whitney J, Singh S, Beig N, Etesami M, Bates DDB, Gallagher K, Bloch BN, Vulchi M, Turk P, Bera K, Abraham D, Sikov WM, Somlo G, Harris LN, Gilmore H, Plecha D, Varadan V, Madabhushi A. Radiomics and deep learning prediction of immunotherapy-induced pneumonitis from computed tomography. JCO Clin Cancer Inform. 2024;8:e2400198.Ethics Approval This study was approved by the Institutional Review Board Committee Northwestern University, Chicago, IL, USA (No. STU0020711; Approval date: 04/06/2018). Consent was waived for its retrospective manner and the was no identifiable risk for patients included in this analysis.Abstract 1044 Figure 1Pneumonitis prediction ROC curve. ROC curve for AI-assisted radiomics model predicting immune-related pneumonitis in NSCLC patients. AUC = 0.718 (95% CI: 0.611-0.826)Abstract 1044 Figure 2Pneumonitis risk stratification cumulative incidence function. Kaplan-Meier curves for immune-related pneumonitis by AI-predicted risk groups. High-risk patients (n=26, red) had significantly higher pneumonitis rates vs low-risk (n=76, green): HR 3.58, p=0.014