Decoding spatial tumor heterogeneity via multi-view radiomics for long-term survival prediction in advanced NSCLC immunotherapy.

Abstract

e20643 Background: Current radiomics models for NSCLC immunotherapy primarily focus on short-term outcomes (e.g., 6–12 months' durable clinical response), neglecting the dynamic evolution of tumor ecosystems. We address this gap by innovatively integrating subregional heterogeneity parameters with multi-view subspace learning, enabling quantitative characterization of spatial intratumoral heterogeneity and extended survival prediction (12–30 months). Methods: A multicenter cohort of 414 stage IIIB-IV NSCLC patients receiving chemoimmunotherapy was analyzed. Baseline CT-derived tumor subregions (10 partitions via SLIC) provided 110 radiomic features/subregion. Four feature views—tumor core, peritumoral zone, subregional textures, and intratumoral heterogeneity (ITH)—were fused using subspace learning with adaptive constraints. Mutual information (MI) and random survival forests (RSF) optimized feature selection for PFS and OS prediction, respectively. An accelerated failure time model validated performance across 12-/18-/24-/30-month intervals. Results: The model demonstrated superior long-term predictive capacity, achieving 24-month OS AUC = 0.742(95%CI:0.698–0.786) and 30-month PFS C-index = 0.607 in external cohorts. Subregional heterogeneity parameters (original_gldm_GrayLevelVariance) outperformed conventional whole-tumor features, reducing cross-center variability by 15.3% (ITH vs. tumor-view models). Multi-view fusion captured synergistic spatial patterns: elevated tumor core heterogeneity (original_glcm_SumSquares) with disordered peritumoral textures (original_glszm_ZoneEntropy) identified high-risk patients (median OS = 14.2 vs. 28.6 months, p < 0.001). Conclusions: This study establishes the first radiomics framework leveraging subregional spatial heterogeneity for dynamic survival prediction beyond 2 years. Its clinical applicability is enhanced by stable performance across heterogeneous treatment protocols and scanner platforms, offering a practical tool to address immunotherapy resistance.