Bayesian personalized dose constraints in selecting patients with non-small cell lung cancer for cardiac...

To develop a personalized approach for selecting patients with non-small cell lung cancer (NSCLC) for cardiac risk-adaptive treatment by creating a normal-tissue complication probability (NTCP) model that accounts for heterogeneous radiation dose effects and deriving personalized dose constraints that incorporate model uncertainty.

We analyzed a training cohort of 160 patients from a completed prospective trial and a validation cohort of 91 patients from an ongoing trial. The endpoint was high-sensitivity cardiac troponin T (hs-cTnT) elevation > 5 ng/L during radiotherapy. A Bayesian hierarchical NTCP model based on risk stratification by decision tree was developed to predict the risk of hs-cTnT elevation, treating mean heart dose (MHD) as a group-specific effect. To address model uncertainty in deriving personalized dose constraints, the probability cut-off parameter was optimized to maximize sensitivity and specificity based on posterior distributions. The patient selection accuracy of the uncertainty-incorporated dose constraints was compared against the conventional point-estimate-based ones in internal validation, same-institution external validation, and prospective implementation testing.

Patients were stratified into 3 risk subgroups based on tumor location and age. The MHD strongly affected patients aged > 64 years with left/mediastinal tumors (odds ratio = 2.16 [95 % credible interval = 1.07-4.14]). The uncertainty-incorporated dose constraints outperformed point-estimate-based dose constraints in specificity (0.57-0.68 vs 0.28-0.51) and accuracy (0.60-0.69 vs 0.43-0.59) across validations.

Based on our Bayesian hierarchical NTCP model, we proposed using uncertainty-incorporated personalized MHD constraints to select patients with NSCLC for cardiac risk-adaptive treatment. This framework represents an important step toward personalized radiotherapy to reduce cardiac toxicity.