Abstract
Radiomic features derived from longitudinal imaging offer a non-invasive approach to quantify tumor heterogeneity, but their integration with clinical variables in small cohorts remains methodologically challenging. We analyzed 23 patients with advanced non-small cell lung cancer (NSCLC) enrolled in a Phase I study of pembrolizumab and vorinostat. Radiomic features were extracted from baseline and two-month follow-up computed tomography scans, aggregated to the patient level, and transformed into delta features representing early changes over time. A unified modeling framework was implemented using strict leave-one-out cross-validation (LOOCV), incorporating both radiomic and clinical variables. Within each training fold, outcome-guided feature screening was performed using elastic net penalized Cox regression, followed by category-aware principal component analysis and ridge-penalized Cox modeling. Model performance was evaluated using out-of-fold concordance indices for overall survival (OS) and progression-free survival (PFS). Radiomics-only models demonstrated moderate discrimination for OS (C-index 0.609) and favorable discrimination for PFS (C-index 0.770), whereas clinical-only models showed weaker performance (OS 0.605; PFS 0.571). The combined radiomics and clinical model demonstrated numerically higher discrimination without statistically distinguishable differences for OS (C-index 0.648) while maintaining favorable discrimination for PFS (C-index 0.718). Stable features were predominantly texture-based and included Laws filter and co-occurrence-derived metrics associated with spatial heterogeneity. In this exploratory analysis, early changes in radiomic features reflecting tumor heterogeneity suggest potential associations with survival outcomes in advanced NSCLC. Integration of radiomic and clinical variables showed trends toward improved performance for overall survival while maintaining favorable discrimination for progression-free survival. These findings support further investigation of longitudinal radiomic features as candidate imaging biomarkers in larger, independently validated cohorts.