ESTRO 2021 Abstract Book

S678

ESTRO 2021

New York City, USA; 4 University Medical Center Utrecht, Pulmonology, Utrecht, The Netherlands; 5 University Medical Center Utrecht, Biostatistics and Research Support, Julius Center for Health Sciences and Primary Care, Utrecht, The Netherlands; 6 New York University, Center for Data Science, New York City, USA Purpose or Objective Randomized Trials (RCT) are the gold standard for inferring treatment effects but suffer from selection bias, as patients in trials are often younger and in better overall health than real-world cancer patients. Observational studies in real-world populations suffer from confounding by indication. The objective of this study is to establish a methodology for estimating individual treatment effects in observational studies on real-world cancer populations in the presence of unobserved confounders. We applied this method to compare concurrent chemoradiation (CCRT) versus sequential chemoradiation (SCRT) in stage III Non-Small Cell Lung Cancer (NSCLC). Results from RCTs indicate that CCRT has a higher overall survival than SCRT (hazard ratio, 0.84; 95% confidence interval, 0.74 to 0.95), though not every patient is fit enough to endure CCRT with its higher acute toxicity. We applied our new method and developed an individual treatment effect model and estimated the average treatment effect in the real-world population. Materials and Methods We conducted a multi-center retrospective observational cohort study in stage III NSCLC patients receiving either CCRT or SCRT. Literature review and expert opinion indicated the presence of an unobserved confounder: overall fitness of the patient. Direct adjustment for observed covariates will then not completely eliminate the confounding bias. We developed a method for estimating treatment effects when only proxies of the true confounder are available. The estimated treatment effect was compared with results from RCTs and conventional confounding adjustment methods.

Fig 1. Causal Directed Acyclic Graph

Results We included 504 patients from 9 different hospitals, treated with either CCRT or SCRT as primary treatment for stage III NSCLC. We observed 141 deaths in 224 patients who underwent CCRT (632 patient years) and 214 deaths in 280 patients who underwent SCRT (603 patient years). Standard methods for treatment effect estimation based on direct adjustment for confounders estimated more extreme treatment effects when compared to those published in randomized trials (hazard ratio, 0.81; 95% credible interval, 0.60 to 1.09) which is likely due to residual confounding. Our method found that on average, the treatments were equivalent (hazard ratio, 1.01; 95% credible interval, 0.68 to 1.53). CCRT was less advantageous for patients with adenocarcinoma, weight loss or clinical substage IIIB, but unaffected by ECOG performance score and age.

Fig 2. Treatment effects estimates

Conclusion We present an individual treatment effect estimation method for real-world cancer populations with unobserved confounding, which we applied to stage III NSCLC patients. Our results indicate that the treatment effect of concurrent chemoradiation from RCTs may be an overestimation of the treatment effect in the real- world population. Our method may be widely applicable to other cancer settings and can help bridge the gap between RCTs and real-world observational studies .