The Clinical Need Immune checkpoint inhibitors (ICIs) show substantial variability in clinical efficacy for unresectable NSCLC, with many patients not responding. Current biomarkers like PD-L1 and TMB rely on invasive biopsies and have limited predictive power, highlighting the need for non-invasive alternatives.
The Dual Approach This study developed a combined model integrating CT-based deep learning radiomic features from DenseNet121 with the systemic immune-inflammation-nutritional index (SIINI) - a novel blood-based composite biomarker reflecting inflammation, immune status, and nutritional state.
Multicenter Validation The study included 265 patients treated with ICIs across two independent medical centers (Centers A and B), split into training (70%), internal validation (30%), and external validation cohorts, providing a realistic assessment of generalization.
Performance The combined model achieved AUCs of 0.865 in internal validation and 0.823 in external validation, outperforming either CT features or SIINI alone, confirming synergy between imaging and blood-based biomarkers.
What SIINI Captures The Systemic Immune-Inflammation-Nutritional Index (SIINI) integrates six pre-treatment laboratory parameters: neutrophil count, lymphocyte count, platelet count, hemoglobin level, serum albumin level, and BMI. These reflect inflammation, immune competence, and nutritional reserves simultaneously.
Calculation Formula SIINI equals (neutrophil count x platelet count x hemoglobin level) divided by (lymphocyte count x BMI x serum albumin level), combining cellular immunity measures with nutritional markers into a single composite value.
Theoretical Basis Elevated neutrophils and platelets indicate systemic inflammation that can suppress anti-tumor immunity, while high lymphocyte counts, hemoglobin, albumin, and BMI reflect preserved immune function and nutritional status - key determinants of effective immunotherapy response.
Innovation Over Existing Indices Compared to single-dimensional indices like NLR (neutrophil-to-lymphocyte ratio), PLR (platelet-to-lymphocyte ratio), or SII (systemic immune-inflammation index), SIINI incorporates nutritional determinants, providing a more comprehensive evaluation of pre-therapeutic host status relevant to ICI outcomes.
DenseNet121 Architecture DenseNet121's densely connected layers enable reuse of features across layers, making it particularly effective at extracting subtle, high-dimensional radiomic features from CT images that reflect tumor heterogeneity, vascularization, and microenvironment characteristics.
Region of Interest Segmentation Before feature extraction, two cancer specialists jointly delineated the primary NSCLC lesions on 3D Slicer software. Any tumor mass at least 5 mm in diameter that was consistently identified across baseline and follow-up CT scans was included as a target lesion.
Gradient-Weighted Class Activation Mapping Grad-CAM visualization highlighted which specific CT image regions contributed most to the model's predictions, enhancing clinical interpretability and confirming that the model focused on biologically meaningful areas of the primary tumor rather than artifacts.
Response Classification Patients were classified as responders (complete remission, partial remission, or stable disease) versus non-responders (progressive disease) according to RECIST 1.1 criteria from follow-up CT assessments every 6-8 weeks after ICI initiation.
Internal Validation AUC The combined DenseNet121 plus SIINI model achieved AUC of 0.865 (95% CI: 0.771-0.960) in the internal validation cohort, demonstrating strong discriminative ability between responders and non-responders.
External Validation AUC In the independent external cohort from Center B, the combined model maintained AUC of 0.823 (95% CI: 0.663-0.983), confirming generalizability across different clinical settings and patient populations.
Synergistic Integration The combined model outperformed either CT features alone or SIINI alone in both validation cohorts, confirming that imaging features and blood-based inflammatory markers capture complementary, non-redundant aspects of immunotherapy response biology.
Practical Advantage Both components - CT scans and blood counts - are collected as routine standard of care before ICI initiation. The model therefore adds predictive value without requiring any additional tests, making it immediately implementable in current clinical workflows.
Inflammation and Immune Suppression Systemic inflammation - reflected by high neutrophil and platelet counts - promotes immunosuppressive signaling pathways within the tumor microenvironment. High systemic inflammation before ICI treatment is associated with resistance to checkpoint blockade across multiple cancer types.
Nutritional Status and T Cell Function Low albumin and BMI indicate malnutrition, which impairs T cell proliferation, cytokine production, and effector function. Even with ICIs disinhibiting T cells, nutritionally depleted T cells may lack the metabolic capacity to mount effective anti-tumor responses.
Tumor Microenvironment Heterogeneity CT-based deep radiomic features capture tumor heterogeneity, shape, and texture patterns that reflect the spatial organization of immune cells, vascular structures, and necrosis within the tumor - factors that directly influence ICI penetration and effectiveness.
Complementary Information Sources The SIINI reflects systemic host factors while CT features capture local tumor characteristics. Their combination covers both dimensions - what the patient's immune system can mount systemically and what the tumor environment presents locally - explaining the synergistic predictive performance.
Multi-Omics Integration Future work should incorporate additional biomarkers such as PD-L1 expression, tumor mutational burden, circulating tumor DNA, and gut microbiome profiles into the model to further improve predictive accuracy and gain mechanistic insight.
Larger Multicenter Validation While two centers provided meaningful diversity, larger prospective studies across diverse geographic regions, ethnic populations, and ICI regimens (monotherapy, combination, first vs. later lines) are needed to establish robust clinical thresholds.
Longitudinal Response Monitoring Beyond predicting early response, extending the model to serial CT timepoints could enable dynamic treatment monitoring - identifying emerging resistance patterns before they manifest as clear radiological progression by RECIST criteria.
Clinical Decision Support Tool Once validated, the model could be integrated into electronic health record systems as a real-time clinical decision support tool that automatically calculates predicted response probability when CT and laboratory results are uploaded, directly informing ICI therapy initiation decisions.