The Core Problem Non-small cell lung cancer (NSCLC) accounts for 85% of all lung cancer cases, and while immune checkpoint inhibitors (ICIs) have transformed treatment, responses vary widely between patients. Some benefit greatly while others experience immune-related adverse events (irAEs) - harmful side effects that can be life-threatening.
What Habitat Radiomics Offers Standard radiomics analyzes the whole tumor as a single region. Habitat radiomics instead divides the tumor into distinct sub-regions (habitats) based on imaging patterns, capturing the internal heterogeneity of tumors - the fact that different areas of a tumor have different biology. This mirrors real tumor behavior more accurately.
Study Design This multicenter retrospective study enrolled 343 NSCLC patients treated with ICIs across two centers: 248 for model development and 95 for validation (75 internal, 20 external). The goal was to build a model that predicts both progression-free survival (PFS) and the likelihood of developing irAEs before treatment begins.
Why This Matters Clinically Predicting which patients will benefit from ICIs - and which will suffer serious side effects - allows oncologists to tailor therapy. Patients at high irAE risk might need prophylactic management or alternative regimens, while those predicted to respond can be prioritized for immunotherapy.
Creating Tumor Habitats CT images were analyzed at multiple phases (arterial, venous, plain). K-means clustering was applied to divide each tumor into three distinct sub-regions or habitats based on local imaging intensity patterns. Each habitat represents a biologically distinct zone within the tumor with different vascular, cellular, or metabolic characteristics.
Feature Extraction From each habitat sub-region, hundreds of radiomic features were extracted including first-order statistical features (intensity distribution), shape features (geometry), texture features (GLCM, GLRLM patterns), and high-order wavelet features. This generated a rich feature set capturing intratumor heterogeneity at the sub-regional level.
Feature Selection with Machine Learning Four machine learning algorithms were applied for feature selection: LASSO regression, Random Forest, Support Vector Machine, and a gradient boosting approach. Only features consistently selected across methods were retained, reducing overfitting risk and identifying the most predictive radiomic signatures.
Four Prediction Models Built The team constructed four models: a Clinical Signature (using demographic and clinical data alone), a Whole-Tumor Radiomics Signature, a Habitat Radiomic Signature, and a Radiomics Nomogram combining habitat features with clinicopathologic factors. All were tested on training, internal validation, and external validation cohorts.
Nomogram Outperforms All Other Models The Radiomics Nomogram achieved AUCs of 0.923, 0.817, and 0.899 in the training, internal validation, and external validation cohorts respectively. This significantly outperformed the Whole-Tumor Radiomics Signature (AUCs: 0.870, 0.736, 0.626) and the Habitat Signature alone (AUCs: 0.900, 0.804, 0.808).
Habitat Beats Whole-Tumor Analysis A key finding was that habitat-based models consistently outperformed whole-tumor radiomics models, especially in external validation. This demonstrates that capturing intratumor heterogeneity via sub-regional analysis adds meaningful predictive value beyond analyzing the tumor as a single entity.
irAE Prediction Validated Beyond PFS prediction, the habitat nomogram also showed strong performance in predicting immune-related adverse events. This dual capability - predicting both efficacy and toxicity - makes the model uniquely useful for clinical decision-making before starting immunotherapy.
Decision Curve Analysis Confirms Clinical Utility Decision Curve Analysis (DCA) showed that using the nomogram provided net clinical benefit across a wide range of threshold probabilities, confirming that the model is clinically useful rather than just statistically accurate. Calibration curves showed good agreement between predicted and observed outcomes.
Non-Invasive Stratification The habitat radiomics approach uses standard pre-treatment CT scans - no additional biopsies or invasive procedures are required. This makes it practical to implement in routine oncology workflows and accessible for patients who may not be suitable for repeat biopsies.
Dual Risk Assessment The model's ability to predict both progression-free survival and irAE risk simultaneously gives clinicians a more complete picture. A patient predicted to have low PFS benefit and high irAE risk could be redirected to alternative therapies, while high-responders with low irAE risk could confidently proceed with ICI treatment.
Addressing Unmet Clinical Need Current biomarkers like PD-L1 expression and TMB have limitations in predicting ICI response at the individual patient level. Habitat radiomics offers a complementary, image-based approach that captures spatial tumor heterogeneity not reflected in single-point biopsies.
Integration with Existing Workflow Since the model relies on pre-existing CT imaging routinely performed for staging, adoption requires no changes to imaging protocols - only the addition of computational analysis. This lowers barriers to clinical translation significantly.
Sample Size Constraints The external validation cohort included only 20 patients, which limits the strength of conclusions about generalizability. Larger multicenter prospective studies are needed to validate these findings across diverse patient populations, institutions, and CT scanner types.
Retrospective Design As a retrospective study, the analysis is subject to selection bias and unmeasured confounders. Prospective validation incorporating the nomogram into treatment decision workflows would be the critical next step to establish clinical utility.
Manual Segmentation Dependency Tumor ROI delineation currently requires manual input from radiologists, which is time-consuming and subject to inter-observer variability. Automated or semi-automated segmentation tools integrated with AI pipelines would be needed for real-world scalability.
Future Directions Future work should explore integrating habitat radiomics with genomic biomarkers (PD-L1, TMB, gene mutations) for multimodal predictive models, examine whether the approach generalizes to other ICI-treated cancers, and develop prospective clinical decision support tools.