Impact of radiomic features on prediction of progression-free survival in patients with Non-Small Cell Lung Cancer (NSCLC) stage IV

Despite advancements in understanding NSCLC, survival rates for stage IV patients remain low, and predicting disease progression at treatment onset is challenging. Incorporating radiomic features may help predict Progression-Free Survival (PFS) rates. The goal of this study is to identify the key patient characteristics contributing to PFS prediction in stage IV NSCLC patients, comparing models with and without radiomic features.

 

This observational, retrospective study used real-world data sourced from the CIAN lung cohort, including patients with Stage IV NSCLC, between 2017 and 2023. PFS was defined as the time from the first treatment registered after stage IV diagnosis until disease progression or any censoring event. We collected 45 demographic and clinical covariates. The ROI corresponding to the main lesion was defined on baseline chest CT images using the lung nnU-Net model, and radiomic features were extracted with pyradiomics, followed by Combat harmonization. The dataset was split into 80-20% training-test sets. Cross-validation was used on training set for model adjustment. Feature selection was performed with Recursive Feature Elimination (RFE). Cox proportional hazards regression, Random Survival Forest (RSF), and Gradient-Boosting Cox models were tested for PFS prediction, adjusted with and without radiomic features. Performance was evaluated using Concordance Index (C-Index), Integrated Brier Score (IBS) and time-dependent AUC. Feature contribution was evaluated using Shapley Additive exPlanations.

 

216 eligible patients from one French hospital were included in the study (59.3±9.7 years of age, 59% male, 83% adenocarcinoma, 83% treated with different combinations of immunotherapy, platinum-based and single-agent chemotherapy). Preliminary results were based on data from 65 patients. For the clinical-only model, seven features were selected by RFE. RSF model showed the best results (C-index=0.74, IBS=0.12, AUC=0.82). For the mixed clinical-radiomics model, eight features were selected by RFE. RSF again performed best (C-index=0.74, IBS=0.18, AUC=0.82). Main clinical features contributing to the models were Aspartate Aminotransferase, Neutrophil and Lymphocyte count, and Serum Creatinine. Main radiomic features contributing to the mixed model included first-order descriptors, specifically log-sigma-5mm and wavelet-filtered median and skewness.

 

RSF model demonstrated the best performance. In preliminary analysis, adding radiomic features from baseline CT scans did not significantly improve PFS model performance but allowed for a reduction in the number of included clinical features, without compromising performance. Relevant clinical and radiomic contributors were identified. These insights could enhance individualized PFS prediction and patient management strategies.