A CT-based AI model can identify lung cancer patients at higher risk of developing pneumonitis, suggest findings published September 18 in the Journal for ImmunTherapy of Cancer.
Researchers led by Amgad Muneer from the University of Texas MD Anderson Cancer Center in Houston developed and validated its AI model, which outperformed conventional clinical-factor models and radiomics approaches in predicting this side effect in lung cancer patients undergoing immunotherapy. The team named its model the Checkpoint-Inhibitor Pneumonitis Hazard EstimatoR (CIPHER).
“Although further validation in larger, multicenter datasets is essential, our pilot findings suggest that CIPHER could facilitate personalized monitoring strategies and preventive interventions for patients most vulnerable to [pneumonitis],” the Muneer team wrote.
Pneumonitis is a potentially life-threatening form of lung inflammation that occurs in about 10% of lung cancer patients receiving immunotherapy. Early identification of high-risk patients before starting immune checkpoint inhibitors can help with close monitoring, timely intervention, and improving outcomes.
CIPHER is a deep learning foundation model that aims to predict pneumonitis from baseline CT scans in patients with lung cancer. It combines contrastive learning with a transformer-based masked autoencoder.
Using self-supervised learning, the team pretrained CIPHER on over 590,000 CT slices from 2,500 patients with non-small cell lung cancer. It then adapted CIPHER to an internal cohort of 347 patients, of whom 33 developed adjudicated pneumonitis. From there, the team fine-tuned the model using 254 non-pneumonitis patients only, and a held-out internal validation set of 93 patients. The latter included 33 pneumonitis cases and 60 controls for evaluation.
CIPHER, an AI-powered CT foundation model, identifies patients with lung cancer at increased risk of immunotherapy-induced pneumonitis before treatment begins. The AI-generated output, shown here, highlights lung abnormalities to support early risk assessment and personalized patient care.University of Texas MD Anderson Cancer Center
Muneer and colleagues compared CIPHER’s performance to that of clinical, radiomics, and ensemble comparator models. They also performed external validated in an independent Johns Hopkins cohort of 116 patients with non-small cell lung cancer. This included 20 pneumonitis cases and 96 controls.
CIPHER achieved area under the curve (AUC) values ranging from 0.77 to 0.85, showing consistency in determining which patients were at higher risk. The model achieved an AUC of 0.83 in head-to-head benchmarking, which the researchers wrote outperformed the clinical, radiomics and ensemble models.
CIPHER also achieved an AUC of 0.83 in the external validation cohort and a balanced accuracy of 81.7%, which the researchers wrote outperformed the radiomics model (p = 0.03). While the radiomics model achieved 85% sensitivity (85.0%), it demonstrated low specificity at 45.8%.
Finally, confusion matrix analyses showed that CIPHER correctly identified 80 of 96 non-pneumonitis cases and 16 of 20 pneumonitis cases.
While prospective studies are needed before clinical applications, the study authors highlighted CIPHER as a promising non-invasive tool for pneumonitis risk assessment.
“CIPHER provides a basis for using baseline CT as a non-invasive pretreatment risk-stratification tool for [pneumonitis],” the authors wrote.
Read the full study here.