CT-based deep foundation model for predicting immune checkpoint inhibitor-induced pneumonitis risk in lung cancer.

backgroundImmune checkpoint inhibitors (ICIs) have revolutionized cancer therapy, but can cause serious immune-related adverse events, with pneumonitis (ICI-P) being among the most severe. Early identification of high-risk patients before ICI initiation is critical for close monitoring, timely intervention, and optimizing outcomes. PURPOSE: To develop and validate a deep learning foundation model to predict ICI-P from baseline CT scans in patients with lung cancer.

methodsWe designed the Checkpoint-Inhibitor Pneumonitis Hazard EstimatoR (CIPHER), a deep learning-powered foundation model combining contrastive learning with a transformer-based masked autoencoder to predict ICI-P from baseline CT scans in patients with lung cancer. Using self-supervised learning, CIPHER was pretrained on 590,284 CT slices from 2,500 patients with non-small cell lung cancer (NSCLC) to learn representations of heterogeneous lung parenchyma. Following pretraining, CIPHER was adapted to the internal MD Anderson Cancer Center NSCLC immunotherapy cohort of 347 patients, of whom 33 developed adjudicated ICI-P. Fine-tuning was performed using 254 non-ICI-P patients only, and a held-out internal validation set of 93 patients, including 33 ICI-P cases and 60 non-ICI-P controls, was reserved for evaluation. CIPHER was benchmarked against clinical, radiomics, and ensemble comparator models and externally validated in an independent Johns Hopkins NSCLC cohort of 116 patients, including 20 ICI-P cases and 96 non-ICI-P controls.

resultsIn our internal immunotherapy cohort, CIPHER consistently distinguished patients at elevated risk of ICI-P from those without the event, with area under the curves (AUCs) ranging from 0.77 to 0.85. In head-to-head benchmarking, CIPHER achieved an AUC of 0.83, outperforming the clinical, radiomics and ensemble models. In the external validation cohort, CIPHER maintained high performance (AUC=0.83; balanced accuracy=81.7%), exceeding the radiomics model (DeLong p=0.0318) and demonstrating superior specificity without sacrificing sensitivity. By contrast, the radiomics model, despite high sensitivity (85.0%), showed markedly lower specificity (45.8%). Confusion matrix analyses confirmed CIPHER's robust classification, correctly identifying 80 of 96 non-ICI-P cases and 16 of 20 ICI-P cases.

conclusionsWe developed and externally validated CIPHER, a CT-based imaging biomarker for pretreatment ICI-P risk stratification in NSCLC. CIPHER shows promise as a non-invasive tool for ICI-P risk assessment but warrants prospective validation before clinical translation.