A Bayesian approach to evaluate the impact of change in IGRT protocol using real world data

Abstract

Purpose or Objective Radiotherapy has a rich history of technological innovation. Image-guided radiotherapy (IGRT) was implemented to increase the precision of patient positioning during radiotherapy. Previous work found that patients with residual set-up errors following IGRT that shifted the radiotherapy dose towards the heart had worse survival than those that shifted the dose away from it (Johnson-Hart et al. IJROBP 2018). In that study the IGRT action threshold was 5mm. In November 2016 this threshold was reduced to 2mm. In this work, we investigate whether this change in IGRT protocol affected patient survival using a Bayesian approach. Materials and Methods Two datasets of patients with NSCLC treated with IGRT were evaluated. The first dataset were treated pre-protocol change (N=780, 5mm action threshold) and the second dataset were treated post (N=411, 2mm action threshold). Patients were followed up for 24 months. Residual shifts were recorded and averaged over all fractions. Weibull survival models were fit to the post-protocol change data to evaluate the effect of residual shift, adjusting for age, sex, performance status, stage, and GTV. Two Bayesian models were fit, one with an uninformative prior (i.e. model fit driven by the post-protocol change data only) and one with a sceptical prior using the data pre-protocol change (i.e. model assumes an effect of shift on survival and updates this belief with the new data). Posterior distributions were calculated using Markov Chain Monte Carlo sampling. Results The hazard ratio (HR) between the survival of patients with residual shifts towards versus away from the heart was reduced post-protocol change, for both priors (Figure 1). Pre-protocol change, the HR was 1.34 (95% credible interval 1.11, 1.60). Post-protocol change, this HR is reduced to 0.909 (0.792, 1.16) (uninformative prior), and 1.14 (0.980, 1.31) (sceptical prior). There is a high probability that the HR has reduced post-protocol change (P>0.9 for both priors). The probability the effect is reduced by 5% is 0.828-0.986, and by 10% is 0.685-0.966. Median survival, calculated from a Weibull fit, for patients with a shift towards the heart is increased from 15.9 (14.2, 17.7) months pre-protocol change to 17.2 (15.7, 18.8) months (sceptical) and 19.6 (16.9, 22.8) months (uninformative) post-protocol change. The probability median survival is increased post-protocol change for these patients is 0.867-0.988; and increased by 1 Conclusion The Bayesian approach allowed us to calculate probabilities that our outcome had improved by specified amounts. The effect of having a residual shift towards the heart during a course of radiotherapy was reduced after the IGRT action threshold was reduced from 5mm to 2mm, likely by at least 5%. Median survival for patients with a shift towards the heart is increased, likely by 1 month. The conclusions are affected by the choice of prior but even with the sceptical prior the probabilities suggest patient benefit.