AI Model Uses Pretreatment CT Scans to Predict Immunotherapy-Related Pneumonitis Risk

Pneumonitis is a potentially life-threatening form of lung inflammation that affects about 10% of patients with lung cancer receiving immunotherapy. Because it can be difficult to predict before symptoms develop, clinicians may have limited ability to tailor monitoring to an individual patient’s risk. Current assessments based on clinical factors and imaging review do not fully capture this risk. Researchers have now developed an artificial intelligence model that identifies patients at increased risk before treatment begins.

Researchers at The University of Texas MD Anderson Cancer Center (Houston, TX, USA) developed the Checkpoint-Inhibitor Pneumonitis Hazard EstimatoR (CIPHER) to analyze routine pretreatment chest computed tomography (CT) scans. The model first learned to recognize patterns in lung tissue rather than learning directly from confirmed pneumonitis cases. Researchers then assessed whether those patterns were associated with pneumonitis after immunotherapy.

CIPHER was trained on more than 590,000 CT image slices from 2,500 patients with lung cancer. Researchers tested it using pretreatment scans from 347 patients with non-small cell lung cancer treated at UT MD Anderson and validated it in an independent external dataset. The model achieved an area under the curve of approximately 0.83 in both cohorts, outperforming conventional clinical-factor models and radiomics approaches. 

Performance remained strong across differences in patient populations, scanners and imaging protocols. Predictions remained significant after accounting for age, smoking history, tumor histology and prior thoracic radiation exposure. Patients classified as high-risk also tended to develop pneumonitis sooner after starting immunotherapy. 

Published in the Journal for ImmunoTherapy of Cancer on September 18, 2026, the findings require larger, more diverse prospective studies before clinical integration can be determined. Researchers also plan to examine performance in other cancer types and whether additional biomarkers improve prediction. 

“What makes this approach particularly interesting is that it was not designed to look for pneumonitis itself. 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,” said Jia Wu, Ph.D., associate professor of Imaging Physics and Thoracic/Head and Neck Medical Oncology at UT MD Anderson. 

Related Links
UT MD Anderson