Novel radiogenomics approach to predict and characterize pneumonitis in stage III NSCLC

Abstract

Unresectable stage III NSCLC is now treated with chemoradiation (CRT) followed by immune checkpoint inhibitors (ICI). Pneumonitis, a common CRT complication, has heightened risk with ICI, potentially causing severe outcomes. Currently, there are no biomarkers to predict pneumonitis risk or differentiate between radiation-induced pneumonitis (RTP) and ICI-induced pneumonitis (IIP). This study analyzed 293 patients from two institutions, with 140 experiencing pneumonitis (RTP: 84, IIP: 56). Two models were developed: M1 predicted pneumonitis risk using seven radiomic features, achieving high accuracy across internal and external datasets (AUCs: 0.76 and 0.85). M2 differentiated RTP from IIP, with strong performance (AUCs: 0.86 and 0.81). Gene set enrichment analysis linked high pneumonitis risk to pathways such as ECM-receptor interaction and T-cell signaling, while high IIP risk correlated with MAPK and JAK-STAT pathways. Radiomic models show promise in early pneumonitis risk stratification and distinguishing pneumonitis types, potentially guiding personalized NSCLC treatment.

Fig. 1. UMAP visualizations of feature separation in pneumonitis and subtypes.

a UMAP visualizations for illustrating the separation of features between pneumonitis occurrence and non-occurrence, and b RTP and IIP. The UMAP plots reveal relatively distinct clusters for differentiating pneumonitis occurrence from non-occurrence and RTP from IIP.

Fig. 2. AUROC curves and nomogram for predicting pneumonitis in stage III NSCLC.

a AUCROC curve for predicting the development of pneumonitis in the internal training set (D₁), b the internal validation set (D₂), and c the external validation set (D₃). d A nomogram that quantifies the probability of developing pneumonitis in unresectable stage III NSCLC patients treated with chemoradiation followed by consolidative durvalumab. e Calibration curve for predictive model. The x-axis shows the predicted probability, while the y-axis shows the actual observed occurrence of pneumonitis. The dotted blue line represents an ideal agreement between the actual and predicted probabilities of pneumonitis. The calibrated.orig curve represents the original, unadjusted calibration curve of the model, showing its initial performance. The calibrated.corrected curve demonstrates the effectiveness of the calibration process (Platt scaling) in adjusting the predicted probabilities.

Fig. 3. AUROC Curves, radiomic heatmaps, and spiderweb plots for analyzing pneumonitis risk and subtypes.

a AUROC curves for distinguishing between RTP and IIP in the training set (D₄), and b the validation set (D₅). c Radiomic heatmap of Law L*E feature shows differences between tumor regions for patients with high risk of developing pneumonitis as compared to patients who will not develop pneumonitis. d Radiomic heatmap of Law L*E feature from multiple inflammatory lesions of patients who have experienced IIP as compared to patients who have experienced RTP. c Spiderweb plot for the specific image phenotypes associated with IIP or RTP and d the same concept in those who do and do not develop pneumonitis. e spiderweb plot of specific radiomic feature phenotypes linked to IIP or RTP, and f spiderweb plot of specific radiomic feature phenotypes linked to the presence or absence of pneumonitis. Each spoke on the spider webplot (labeled f1 through f7) corresponds to one of the seven most predictive features selected through LASSO for constructing the PRS or DRS models. The lengths of each spoke reflect the magnitude of each radiomic feature’s contribution, with higher values indicating stronger expression of specific textures, such as edge detection or speckled patterns, in the CT scans.

Fig. 4. AUROC curves and radiomic heatmaps for distinguishing between RTP and IIP using “delta radiomic” change between baseline and post-treatment CT scans.

a AUCROC curve in the training set (D₄), and b the validation set (D₅). c Segmented tumor region and heatmap of Law feature on the baseline CT and inflammatory region on the post-treatment CT scans of two patients who will develop IIP. d Segmented tumor region and heatmap of Lawfeature on the baseline CT and inflammatory region on the post-treatment CT scans of two patients who will develop RTP.

Fig. 5. UMAP embeddings highlighting radiomic feature separation by pneumonitis grades.

a UMAP plot for distinguishing pneumonitis grades from baseline CT scans, reflecting a complex and heterogeneous imaging phenotype with overlapping clusters that do not clearly differentiate pneumonitis risk. b UMAP plot for distinguishing pneumonitis grades from post-treatment CT scans, where pneumonitis has developed, showing distinct separation among clusters that capture the severity and texture changes associated with different pneumonitis grades. This indicates that post-treatment scans are more effective in identifying pneumonitis-related alterations based on radiomic signatures.

Fig. 6. Molecular characteristics associated with PRS and DRS.

Gene set enrichment analysis identified a PRS-positively correlated (pneumonitis induced) biological pathways, b PRS-negatively correlated (no evidence of pneumonitis) biological pathways, c Volcano plotshowing the differential expressed genes between pneumonitis induced and no pneumonitis occurrence component. The top positively correlated genes with the PRS for pneumonitis include IL6, TNF, and CCL2, which are known to play critical roles in inflammatory responses. Conversely, negatively correlated genes such as TP53 and VEGF are associated with tumor suppression and angiogenesisregulation. d DRS-positively correlated (IO induced) biological pathways, and e DRS-negatively correlated (RT induced) biological pathways from KEGG pathway database (FDR < 0.05).

Fig. 7

The consort diagram illustrates the enrollment process, eligibility criteria, and exclusion details of the dataset.