Researchers at The University of Texas MD Anderson Cancer Center have developed an artificial intelligence model that identifies lung cancer patients at increased risk of developing pneumonitis before immunotherapy begins. Published in the Journal for ImmunoTherapy of Cancer, the study demonstrates that routine chest CT scans contain measurable patterns associated with susceptibility to this potentially life-threatening lung inflammation.
How the CIPHER AI Model Predicts Immunotherapy Toxicity
The artificial intelligence model, named the Checkpoint-Inhibitor Pneumonitis Hazard EstimatoR (CIPHER), was trained on more than 590,000 CT image slices from 2,500 lung cancer patients. According to the study findings, the model learns subtle tissue patterns linked to future risk rather than analyzing confirmed pneumonitis cases directly. Researchers tested CIPHER using pretreatment CT scans from 347 non-small cell lung cancer patients treated at MD Anderson, subsequently validating the approach with an independent external dataset.
The model achieved an area under the curve of approximately 0.83 in both cohorts. This performance surpasses conventional clinical-factor models and standard radiomics approaches. According to Jia Wu, Ph.D., associate professor of Imaging Physics and Thoracic/Head and Neck Medical Oncology, the tool detects signals using routine imaging data already gathered during standard clinical care.
Clinical Implications and Patient Monitoring
Pneumonitis affects roughly 10% of lung cancer patients receiving immunotherapy and remains difficult to predict before symptoms manifest. The MD Anderson study indicates that patients flagged as high-risk by CIPHER tend to develop pneumonitis sooner after treatment initiation. These predictive signals remained statistically significant even after researchers adjusted for patient age, smoking history, tumor histology, and prior thoracic radiation exposure.
Co-senior author Ajay Sheshadri, M.D., associate professor of Pulmonary Medicine, and co-senior author Mehmet Altan, M.D., associate professor of Thoracic/Head and Neck Medical Oncology, contributed to the research alongside Wu. The team notes that future prospective studies involving larger, diverse patient groups are required before the tool can enter standard clinical workflows. Subsequent research will also assess whether similar artificial intelligence frameworks can forecast other immunotherapy-related toxicities across different cancer types.