PET/CT radiomics-based prognostic model with exploratory deep learning features in locally advanced non-small cell lung cancer

Locally advanced non-small cell lung cancer (LA-NSCLC) is characterized by substantial heterogeneity, and the traditional TNM staging system has limited capacity for individualized prognostic assessment. This study aimed to develop and evaluate a comprehensive prognostic model integrating clinical characteristics, tumor burden, and PET/CT-based radiomics with exploratory deep learning features, and to assess model interpretability using machine-learning techniques. This single-center retrospective study included patients with LA-NSCLC who underwent pretreatment 18F-FDG PET/CT and were divided into training and internal test cohorts. Radiomic features were extracted separately from PET and CT images, and endpoint-specific radiomics risk scores were constructed using elastic-net Cox regression. A dual-branch 3D CNN-Transformer framework generated deep-learning risk scores. Clinical variables, radiomics, and deep-learning features were integrated to establish multimodal prognostic models for overall survival (OS) and progression-free survival (PFS). All feature selection, model fitting, hyperparameter tuning, and cutoff definition were restricted to the training cohort. Model performance was evaluated in the locked test cohort using the concordance index (C-index), time-dependent receiver operating characteristic curves, calibration, decision curve analysis, and Kaplan-Meier analysis. SHAP analysis was used to describe contributions to model predictions. A total of 221 patients were included (training cohort, n = 176; internal test cohort, n = 45). Radiomics and deep-learning risk scores were associated with survival outcomes. The integrated model yielded OS C-indices of 0.731 and 0.743 in the training and test cohorts, respectively; the gain over the radiomics model was modest. In the test cohort, AUCs for predicting 1-, 2-, and 3-year OS were 0.658, 0.820, and 0.874, respectively. For PFS, the deep-learning model had a higher test C-index than the integrated model. Exploratory subgroup results were limited by small sample sizes. SHAP identified the radiomics risk score as the largest contributor to the primary model predictions. The proposed multimodal PET/CT-based model demonstrated promising prognostic performance for LA-NSCLC.