A deep learning foundation model was able to predict pneumonitis risk from immune checkpoint inhibition therapy in patients with non–small cell lung cancer (NSCLC) based on baseline computed tomography (CT) scans, according to findings published in the Journal for ImmunoTherapy of Cancer.
“Pneumonitis remains one of the most challenging complications of immunotherapy because it can be difficult to predict before symptoms appear,” said co-lead investigator and senior study author Jia Wu, PhD, Associate Professor of Imaging Physics and Thoracic/Head and Neck Medical Oncology and an affiliate member of The University of Texas MD Anderson Cancer Center's Institute for Data Science in Oncology. “Our model was able to identify signals associated with future risk using information that already exists in routine CT scans.”
Study Methods
Researchers developed a deep learning–powered foundation model, which they called the Checkpoint-Inhibitor Pneumonitis Hazard EstimatoR (CIPHER), to predict immune checkpoint inhibitor–induced pneumonitis from CT scans in patients with lung cancer. The model's architecture consisted of contrastive learning and a transformer-based masked autoencoder. Pretraining was conducted with self-supervised learning on 590,284 slices of CT scans from 2,500 patients with NSCLC so that the model could learn representations of heterogeneous lung parenchyma.
Then, the CIPHER model was used on an internal cohort of 347 patients with NSCLC who had received immunotherapy, including 33 who developed immune checkpoint inhibitor–induced pneumonitis.
The model was then fine-tuned on scans from 254 patients who did not develop immune checkpoint inhibitor–induced pneumonitis. An internal validation set of 93 patients included 33 patients who developed pneumonitis and 60 who did not. The model was also externally validated on an independent cohort of 116 patients with NSCLC from Johns Hopkins, which included 20 patients who developed pneumonitis.
The CIPHER model was benchmarked against clinical, radiomics, and ensemble comparator models.
Key Findings
CIPHER identified patients at a higher risk of developing immune checkpoint inhibitor–induced pneumonitis. Areas under the curve (AUC) ranged from 0.77 to 0.85 in the internal cohorts. CIPHER correctly identified 16 of 20 cases of immune checkpoint inhibitor–induced pneumonitis and 80 of 96 cases of non–immune checkpoint inhibitor–induced pneumonitis.
When compared with clinical, radiomics, and ensemble models, CIPHER outperformed them all, with an AUC of 0.83.
In external validation, CIPHER demonstrated an AUC of 0.83 with balanced accuracy of 81.7%; this exceeded the performance of the radiomics model (DeLong P = .0318) and showed greater specificity for the CIPHER model than the radiomics model while sensitivity remained high as well, unlike with the radiomics model (sensitivity = 85%; specificity = 45.8%).
Strong performance was maintained across the different cohorts with various patient populations, imaging protocols, and CT scanners.
Patients who were characterized by the AI as high risk tended to develop pneumonitis sooner than others after initiating immune checkpoint inhibition. This suggested that the model could detect signals of lung vulnerability.
“What makes this approach particularly interesting is that it was not designed to look for pneumonitis itself,” Dr. Wu said. “Instead, the model learned patterns within lung tissue and identified subtle abnormalities associated with future risk. That suggests routine imaging may contain much more information about treatment toxicity than we previously recognized.”
The study authors believe that CIPHER could be a promising noninvasive tool for assessing risk of immune checkpoint inhibitor–induced pneumonitis, but that more prospective studies are needed to determine whether the method could be integrated into clinical workflows.
Going forward, the researchers plan to evaluate whether the model performs comparatively in other cancer types and whether adding other biomarker information could improve risk prediction even more for pneumonitis as well as for other immune-related toxicities.
DISCLOSURES: This research was supported by the National Institutes of Health, the Cancer Prevention and Research Institute of Texas (CPRIT) and UT MD Anderson institutional funding. For full disclosures of the study authors, visit jitc.bmj.com.
