Clinical Problem Checkpoint inhibitor pneumonitis (CIP) is a potentially life-threatening immune-related adverse event that can occur in patients receiving PD-1 or PD-L1 inhibitors. It occurs unpredictably and can force treatment discontinuation, making early prediction critical for patient safety and management.
Study Objective This study investigated whether radiomic features extracted from pre-treatment CT scans of the lung - combined with clinical variables - could predict which patients would develop CIP. The hypothesis is that baseline lung architecture and texture encode vulnerability to immune-mediated inflammation.
Patient Population The study included 116 patients: 35 who developed CIP and 81 controls who received checkpoint inhibitors without developing pneumonitis. This imbalanced dataset reflects the real-world incidence of CIP and required careful handling during model development.
Feature Set A total of 171 radiomic features were extracted from CT images, capturing shape, texture, and intensity characteristics of the lung parenchyma. These were combined with clinical variables including smoking history and PD-L1 expression level for model training.
CT Image Segmentation Pre-treatment CT scans were processed to segment the lung parenchyma. Radiomic features were then extracted from these regions, capturing quantitative properties of lung tissue texture, density distribution, and structural heterogeneity that are invisible to the naked eye.
Feature Categories The 171 features spanned first-order statistics (intensity histograms), gray-level co-occurrence matrices (texture patterns), wavelet transforms, and shape-based descriptors. Together, these characterize the microscopic and macroscopic organization of lung tissue prior to treatment.
Machine Learning Algorithms Multiple classifiers were evaluated, including random forests (RF) and generalized linear models (GLM), each tested on three input configurations: radiomic features alone, clinical variables alone, and a combined feature set. This factorial design allowed assessment of both individual and synergistic predictive contributions.
Nested Cross-Validation Model selection and performance estimation used nested 5-fold cross-validation to prevent overfitting and provide unbiased performance estimates on held-out data, accounting for the small and imbalanced dataset.
Best Model Performance The random forest model using combined radiomic and clinical features achieved the highest AUC of 0.75 on the test set. This outperformed both the radiomic-only GLM (AUC 0.71) and the clinical-only random forest (AUC 0.72), confirming that combining both data types provides complementary predictive value.
Clinical Risk Factors Among clinical variables, smoking history and high PD-L1 expression were more prevalent in the CIP group compared to controls. These findings align with prior literature suggesting that inflammatory baseline states and heightened immune activation may predispose patients to immune-related lung toxicity.
Radiomic Contributions Even without clinical variables, radiomic features alone achieved reasonable discrimination (AUC 0.71), suggesting that baseline lung tissue characteristics captured by CT encode genuine biological differences relevant to CIP susceptibility. Texture features reflecting parenchymal heterogeneity were particularly informative.
Combined Model Advantage The incremental gain from combining radiomic and clinical features (0.75 vs. 0.72 or 0.71 alone) suggests that each input type captures distinct aspects of CIP risk - imaging reflecting tissue vulnerability and clinical variables reflecting systemic immune context.
Lung Tissue as a Reflection of Immune Vulnerability Radiomic features quantify subtle baseline abnormalities in lung texture and density that may reflect pre-existing subclinical inflammation, fibrosis, or immune cell infiltration. These features may identify lungs already primed for an exaggerated immune response when PD-1 checkpoints are blocked.
PD-L1 and Immune Activation High PD-L1 expression in the CIP group may indicate an already hyperactivated immune environment. When checkpoint blockade is applied in this context, the resulting immune activation may preferentially damage normal lung tissue, consistent with CIP pathophysiology.
Smoking History as a Confounder and Risk Factor Smokers have chronic airway inflammation, which may lower the threshold for immune-mediated lung injury. The enrichment of smoking history in the CIP group underscores the importance of including this variable in predictive models and patient counseling.
Small Sample Size With only 35 CIP cases, statistical power is limited, and the AUC values may have wide confidence intervals. A larger multicenter dataset would be needed to confirm these findings and improve model robustness before clinical deployment.
Class Imbalance The 35:81 case-to-control ratio introduces class imbalance that may affect model calibration. While nested cross-validation helps mitigate overfitting, dedicated imbalance-handling strategies such as SMOTE or cost-sensitive learning may further improve CIP recall.
Generalizability Models trained on a single institutional dataset may not generalize across different CT scanners, imaging protocols, or patient populations. Prospective external validation is essential before these models could be used in clinical decision-making.
Future Integration Incorporating blood biomarkers such as inflammatory cytokines, prior lung function tests, and CT features at multiple time points could further improve CIP prediction. Longitudinal radiomic changes early in treatment may be particularly valuable early warning signals.