A deep learning foundation model called CIPHER successfully predicts immune checkpoint inhibitor-induced pneumonitis from baseline computed tomography scans in non-small cell lung cancer patients, according to findings published in the Journal for ImmunoTherapy of Cancer. The model detected subtle abnormalities in lung tissue associated with treatment toxicity before symptoms appeared.
Development and Training of the CIPHER Model
Researchers created the Checkpoint-Inhibitor Pneumonitis Hazard EstimatoR (CIPHER) using an architecture consisting of contrastive learning and a transformer-based masked autoencoder. Pretraining involved self-supervised learning on 590,284 slices of CT scans from 2,500 non-small cell lung cancer patients. This process allowed the model to learn representations of heterogeneous lung parenchyma. Funding for the research was provided by UT MD Anderson institutional sources, the Cancer Prevention and Research Institute of Texas (CPRIT), and the National Institutes of Health.
The research team applied the CIPHER model to an internal cohort of 347 non-small cell lung cancer patients who received immunotherapy. Within this group, 33 patients developed immune checkpoint inhibitor-induced pneumonitis. Fine-tuning followed using scans from 254 patients who did not develop the condition. An internal validation set of 93 patients included 33 who developed pneumonitis and 60 who did not.
CIPHER Model Achieves High Accuracy During Internal and External Validation
During internal validation, the CIPHER model achieved areas under the curve ranging from 0.77 to 0.85. The tool accurately detected 80 out of 96 non-immune checkpoint inhibitor-induced pneumonitis cases and 16 out of 20 immune checkpoint inhibitor-induced pneumonitis cases. When benchmarked against clinical, radiomics, and ensemble comparator models, CIPHER outperformed all of them with an AUC of 0.83.
In external validation on an independent cohort of 116 non-small cell lung cancer patients from Johns Hopkins—which included 20 patients who developed pneumonitis—CIPHER demonstrated an AUC of 0.83. The model achieved a balanced accuracy of 81.7%. This performance exceeded the radiomics model with a DeLong P value of .0318, showing greater specificity while maintaining high sensitivity at 85% compared to the radiomics model’s specificity of 45.8%.
Detection of Early Vulnerability Signals
Patients categorized by the artificial intelligence model as high risk tended to develop pneumonitis sooner after initiating immune checkpoint inhibition than other patients. According to 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, the model was not specifically designed to detect pneumonitis but instead identified subtle abnormalities linked to future risk by learning patterns within lung tissue. This performance remained strong across different cohorts featuring varied patient populations, imaging protocols, and CT scanners.
Frequently Asked Questions About CIPHER Risk Prediction
What patient cohorts were used to externally validate the CIPHER model?
The external validation cohort consisted of 116 patients with non-small cell lung cancer from Johns Hopkins, which included 20 patients who developed immune checkpoint inhibitor-induced pneumonitis.
How did CIPHER compare to traditional radiomics models in validation tests?
In external validation, CIPHER achieved an AUC of 0.83 and a balanced accuracy of 81.7%, exceeding the performance of the radiomics model with a DeLong P value of .0318 and demonstrating higher specificity while keeping sensitivity at 85%.
What funding sources supported the development of this deep learning model?
The research was supported by the National Institutes of Health, the Cancer Prevention and Research Institute of Texas (CPRIT), and institutional funding from UT MD Anderson.
Future Evaluation Plans for Other Toxicities
Study authors stated that more prospective studies are needed to determine whether CIPHER can be integrated into clinical workflows. Future plans include evaluating whether the model performs comparatively in other cancer types. Researchers also intend to test whether adding other biomarker information can improve risk prediction for pneumonitis and other immune-related toxicities, though it remains undetermined when these prospective trials will begin or which specific cancer types will be tested next.
Keep reading