Purpose: We propose and simulate a model-based methodology to incorporate heterogeneous treatment benefit of proton therapy (PrT) versus photon therapy into randomized trial designs. We use radiation-induced pneumonitis (RP) as an exemplar. The aim is to obtain an unbiased estimate of how predicted difference in normal tissue complications probability (ΔNTCP) converts into clinical outcome on the patient level.
Materials and Methods: ΔNTCP data from in silico treatment plans for photon therapy and PrT for patients with locally advanced lung cancer as well as randomly sampled clinical risk factors were included in simulations of trial outcomes. The model used at point of analysis of the trials was an iQUANTEC model. Trial outcomes were examined with Cox proportional hazards models, both in case of a correctly specified model and in a scenario where there is discrepancy between the dose metric used for ΔNTCP and the dose metric associated with the "true" clinical outcome, that is, when the model is misspecified. We investigated how outcomes from such a randomized trial may feed into a model-based estimate of the patient-level benefit from PrT, by creating patient-specific predicted benefit probability distributions.
Results: Simulated trials showed benefit in accordance with that expected when the NTCP model was equal to the model for simulating outcome. When the model was misspecified, the benefit changed and we observed a reversal when the driver of outcome was high-dose dependent while the NTCP model was mean-dose dependent. By converting trial results into probability distributions, we demonstrated large heterogeneity in predicted benefit, and provided a randomized measure of the precision of individual benefit estimates.
Conclusions: The design allows for quantifying the benefit of PrT referral, based on the combination of NTCP models, clinical risk factors, and traditional randomization. A misspecified model can be detected through a lower-than-expected hazard ratio per predicted ΔNTCP.