The Treatment Landscape for Inoperable Stage III NSCLC Inoperable stage III non-small cell lung cancer represents a major clinical challenge - tumors are locally advanced but cannot be surgically removed. The current standard of care combines concurrent chemoradiotherapy (CCRT) with a PD-1 immune checkpoint inhibitor, an approach validated by landmark trials showing improved progression-free and overall survival compared to chemoradiation alone.
The Problem of Treatment Heterogeneity Despite the proven benefit of CCRT plus PD-1 inhibitors as a group, individual patient responses vary enormously. Some patients achieve complete or partial remission, while others progress during or shortly after treatment. Identifying non-responders before treatment begins would allow clinicians to offer alternative strategies rather than exposing patients to the toxicities of an ineffective regimen.
CT Radiomics as a Non-Invasive Predictor Radiomics extracts hundreds to thousands of quantitative features from routine CT scans that capture tumor texture, shape, and heterogeneity patterns invisible to the naked eye. Since CT scans are performed as part of standard staging for all stage III NSCLC patients, radiomics can potentially identify treatment-predictive signatures without additional procedures or costs.
Study Design and Innovation This study built and validated a pretreatment CT-based radiomics machine learning model using 88 patients from multiple hospitals in Guizhou, China, across three independent cohorts: a retrospective training set, a prospective validation set, and an additional test set. Including a prospective validation cohort - where patients were enrolled after the model was built - provides a stronger test of real-world predictive utility than retrospective validation alone.
Three-Cohort Study Design The study enrolled 88 patients from hospitals in Guizhou Province, China. The retrospective training cohort included 38 patients, the prospective validation cohort included 17 patients enrolled after model development, and the test cohort included 33 patients. This three-cohort structure enables assessment of both internal and prospective performance, with the prospective cohort providing the most rigorous validation of real-world applicability.
Treatment Protocol All patients received intensity-modulated radiation therapy (IMRT) delivering 60-66 Gray to the primary tumor, concurrent with platinum-based doublet chemotherapy. After completing CCRT, patients received sequential PD-1 inhibitor therapy - either tislelizumab or sintilimab - as maintenance immunotherapy. This combined CCRT-plus-immunotherapy approach follows current treatment guidelines for inoperable stage III disease.
Response Classification Treatment response was assessed using RECIST criteria at restaging imaging after completing the combined treatment. Patients were classified as treatment-sensitive (complete response or partial response, 48.86% of total) or treatment-resistant (stable disease or progressive disease, 51.14%). This binary classification served as the outcome variable for all predictive models, reflecting clinically meaningful differences in treatment benefit.
CT Imaging and ROI Delineation Pretreatment CT scans were acquired for all patients using standard institutional protocols before any treatment. Regions of interest (ROIs) encompassing the primary lung tumor were manually delineated using ITK-SNAP software by experienced radiologists. ROIs were carefully limited to the primary tumor mass, excluding adjacent mediastinal lymph nodes and atelectasis, to ensure radiomic features reflected intrinsic tumor biology rather than surrounding anatomical structures.
High-Dimensional Feature Space Using PyRadiomics 3.0, a widely validated open-source radiomics toolkit, 1,834 quantitative features were extracted from each patient's tumor ROI. These features span multiple categories: first-order statistics (intensity distribution), shape features (tumor morphology), and texture features from multiple methods including GLCM (gray-level co-occurrence matrix), GLSZM (gray-level size zone matrix), NGTDM (neighboring gray-tone difference matrix), and local binary pattern (LBP) transformations across multiple wavelet filter banks.
Feature Preprocessing and Normalization All 1,834 features underwent Z-score normalization to ensure mean 0 and standard deviation 1, preventing features with larger absolute values from dominating model training. Variance-based filtering removed features with near-zero variance. Intraclass correlation coefficients (ICC) were computed to select only reproducible features with ICC greater than 0.75, reflecting stable measurements across different delineation attempts.
LASSO Regression for Dimensionality Reduction Least Absolute Shrinkage and Selection Operator (LASSO) regression was applied to the filtered feature set using cross-validation to select the optimal regularization parameter lambda. LASSO shrinks coefficients of redundant and low-predictive-value features to exactly zero, automatically selecting a parsimonious set of non-redundant predictors. This approach simultaneously prevents overfitting and identifies the most informative radiomic features.
Six Selected Features LASSO selected 6 radiomic features from the initial 1,834: exponential_firstorder_InterquartileRange (tumor intensity spread), exponential_ngtdm_Complexity (local texture complexity), lbp.3D.k_glszm_SmallAreaLowGrayLevelEmphasis (small homogeneous low-density regions), lbp.3D.m2_glcm_Correlation (local texture correlation), wavelet.HHH_glszm_SizeZoneNonUniformityNormalized (regional heterogeneity after high-frequency filtering), and wavelet.LHH_firstorder_Skewness (asymmetry in intensity distribution). All six features showed statistically significant differences between treatment-sensitive and treatment-resistant groups.
Six Algorithms Tested Six machine learning classifiers were trained on the 6 LASSO-selected features: Logistic Regression (LR), Discriminant Analysis (DA), Neural Network (NN), Random Forest (RF), Support Vector Machine (SVM), and K-Nearest Neighbors (KNN). Each algorithm was evaluated using the same training, prospective validation, and test sets to enable fair comparison. Multiple evaluation metrics were assessed including AUC, sensitivity, specificity, accuracy, and calibration.
Logistic Regression Selected as Best Model Logistic Regression achieved the best balance of performance and generalizability across all three cohorts. Training set AUC was 0.895; prospective validation set AUC was 0.900 (95% CI 0.75 to 1.00); test set AUC was 0.850 (95% CI 0.72 to 0.98). The fact that prospective validation AUC (0.900) was essentially the same as training AUC (0.895) confirms that the model generalizes without overfitting, a key criterion for clinical reliability.
Performance Consistency Across Cohorts The near-identical AUC values across training (0.895), prospective validation (0.900), and test (0.850) sets indicate robust and consistent predictive performance. The minimal drop from training to validation demonstrates that the 6-feature LASSO-selected model captured generalizable biology rather than overfitting to training dataset characteristics. Calibration curves showed good alignment between predicted probabilities and actual response rates.
Clinical Decision Thresholds The logistic regression model's predicted probabilities were used to classify patients into high-risk (predicted treatment-resistant) and low-risk (predicted treatment-sensitive) groups. The optimal cutoff was selected to maximize sensitivity and specificity, providing clinically actionable binary predictions that could guide pre-treatment decision-making and patient counseling.
Significant PFS Differences Between Risk Groups Kaplan-Meier survival analysis showed that patients classified as low-risk (model-predicted treatment-sensitive) had significantly longer progression-free survival (PFS) than high-risk patients in all three cohorts. Log-rank p-values were p=0.001 in the training set, p=0.010 in the prospective validation set, and p=0.014 in the test set - all statistically significant and consistent across independent cohorts.
Overall Survival Separation Similarly, model-predicted risk groups showed significantly different overall survival (OS). Log-rank p-values were p=0.019 in the training set, p=0.028 in the prospective validation set, and p=0.041 in the test set. The progressive p-values from training to validation to test sets remain statistically significant despite decreasing sample sizes, reinforcing the clinical relevance of the risk stratification.
The Radiomic Signature Captures Tumor Biology The 6 LASSO-selected features collectively represent multiple aspects of tumor heterogeneity. High radiomic complexity and intensity heterogeneity likely reflect aggressive, hypoxic, or genomically unstable tumor phenotypes less likely to respond to CCRT or PD-1 inhibition. Conversely, more homogeneous texture patterns may indicate tumors with characteristics associated with PD-L1-mediated immune sensitivity and better DNA damage response.
Radiomic Model Outperforms Clinical Variables Alone While traditional clinical factors such as tumor stage, histology, and performance status have limited individual predictive value for CCRT plus immunotherapy response, the radiomic signature integrates global tumor heterogeneity information that clinical variables cannot capture. The model's AUC of 0.900 substantially exceeds typical AUC values of 0.6 to 0.7 for clinical-variable-only prediction models in similar patient populations.
Pre-Treatment Decision Support The model's key clinical value is that it uses only the standard pretreatment CT scan - obtained before any treatment is initiated - to predict which patients will respond to CCRT plus PD-1 inhibitor therapy. Patients predicted to be treatment-resistant could be offered alternative regimens, enrollment in clinical trials testing novel combinations, or intensive early response monitoring with planned treatment modification.
Avoiding Ineffective Treatment Toxicity CCRT is physically demanding, causing esophagitis, pneumonitis, bone marrow suppression, and fatigue in many patients. Patients who will not respond to this aggressive regimen derive no survival benefit while bearing full treatment toxicity. Pre-identification of likely non-responders enables clinicians to discuss the evidence-based risk-benefit ratio and potentially preserve quality of life through less toxic alternatives.
Guiding PD-1 Inhibitor Selection The study used two different PD-1 inhibitors - tislelizumab and sintilimab - in the cohorts. Future development of the model could potentially distinguish which specific immunotherapy agent a patient is most likely to respond to, given that anti-tumor immune responses are shaped by the tumor immune microenvironment features that radiomics partially captures.
Integration with Liquid Biopsy and Pathological Biomarkers The radiomic model could be combined with tissue-based predictive biomarkers including PD-L1 expression, tumor mutational burden (TMB), and circulating tumor DNA to create multi-modal predictive signatures. Each modality captures different aspects of treatment sensitivity - radiomics provides spatial tumor heterogeneity information that cannot be obtained from small tissue biopsies or blood samples alone.
Small Sample Size With only 88 patients total across three cohorts, the study is limited by statistical power, particularly for the 17-patient prospective validation cohort. While the 0.900 AUC in prospective validation is promising, the wide 95% confidence intervals (0.75 to 1.00) reflect this small sample size. A larger prospective trial with 200 or more patients would provide narrower confidence intervals and more reliable performance estimates.
Single-Center Training Data The retrospective training cohort came from a single center in Guizhou Province, China. CT scanner models, imaging protocols, and tumor delineation practices vary between institutions, and radiomic features can be sensitive to these technical variables. External validation at institutions using different CT scanners and imaging protocols would confirm the model's generalizability beyond its training environment.
Manual ROI Delineation Variability Although experienced radiologists delineated tumor ROIs and ICC-based filtering was applied, manual segmentation introduces interobserver variability that could affect radiomic feature values. Semi-automated or deep learning-based segmentation approaches that consistently define tumor boundaries could improve reproducibility and reduce the expert time required for model deployment in clinical workflows.
Integration of Functional Imaging and Molecular Data Future studies should incorporate additional pretreatment imaging modalities including PET-CT for metabolic activity and diffusion-weighted MRI for cellularity, combined with molecular biomarkers such as PD-L1 expression and genomic mutation profiles. Multimodal integration would likely improve predictive accuracy beyond what CT radiomics alone can achieve, particularly for identifying the immune-responsive phenotypes that best benefit from PD-1 inhibition.