CT radiomics combined with 3D deep learning interpretable model to predict the efficacy of PD-1 inhibitor plus platinum-based chemotherapy in advanced NSCLC: a three-center retrospective cohort study

Abstract Background Platinum‑based chemotherapy combined with PD‑1 inhibitors serves as standard systemic treatment for unresectable stage IIIB–IV non‑small‑cell lung cancer (NSCLC) in both first‑line and later‑line settings; however, primary resistance causes substantial treatment‑response heterogeneity. Tissue‑based biomarkers including PD‑L1 expression and tumor mutational burden (TMB) are limited by invasive biopsy requirements and high costs, whereas routine thin‑slice chest CT non‑invasively captures intratumoral spatial heterogeneity. We constructed a multimodal interpretable AI model integrating clinical variables, three‑dimensional (3D) radiomics and pre‑trained 3D‑DenseNet latent features to predict immunochemotherapy response and evaluate its cross‑center generalizability. Methods We retrospectively enrolled 378 advanced NSCLC patients receiving PD‑1 inhibitor plus platinum‑based chemotherapy from three tertiary hospitals. Three‑dimensional regions of interest (ROIs) of primary tumors were manually double‑blinded segmented. Following three‑step feature filtering, four models were built: clinical‑only, radiomics‑only, standalone 3D deep learning (DL), and multimodal fusion. Center 1 data were split into training (n = 66) and internal validation (n = 35) sets; Centers 2 and 3 formed two fully independent external test sets (n = 81 and n = 196). Model performance was evaluated via area under the receiver operating characteristic curve (AUC) and decision curve analysis (DCA). Ablation, multi‑classifier and subgroup analyses assessed robustness. SHAP and 3D Grad‑CAM enabled model interpretation; TCGA‑TCIA datasets supported immunological mechanism exploration, and an interactive prediction platform was developed. Results Only pathological subtype and treatment line exerted clinical predictive effects. The fusion model achieved competitive AUCs: training 0.909, internal validation 0.946, external‑test‑1 0.848, external‑test‑2 0.893. Although standalone 3D DL attained marginally higher AUC in training cohort, fusion outperformed clinical‑only and radiomics‑only models across all cohorts (all P < 0.05). 3D DL features represented the major predictive contributor. The model maintained AUC > 0.87 across most subgroups; calibration and DCA verified favorable consistency and clinical net benefit. SHAP‑Grad‑CAM visualized decision logic, and transcriptomic data associated CT‑defined high‑risk phenotypes with immunosuppressive microenvironments and reduced CD8⁺ T‑cell infiltration. Conclusions This low‑cost interpretable multimodal AI based on routine chest‑CT non‑invasively predicts immunochemotherapy resistance for advanced NSCLC across treatment lines. Validated by two external cohorts, it demonstrates complementary predictive value of radiomic and DL signatures and offers a standardized imaging tool for personalized immunotherapy stratification.