The rapid evolution of state-of-the-art treatments for thoracic cancers leads to a systematic
underperformance of published prediction models of toxicity, i.e., normal tissue complication
probability (NTCP) models, in recently treated patients. This implies a suboptimal NTCP model-based
planning of photon-based radiotherapy (XT) and errors in NTCP model-based patient selection for the
potentially beneficial but more expensive proton therapy (PT). Our hypothesis is that prospectively
learning from every patient could detect early the treatment-related changes relevant for the
outcome, leading to the necessary NTCP model revisions. We will therefore build a framework for
continuous prediction model updating and will evaluate its functioning using literature-based
simulations and existing retrospective patient cohorts. The framework will then be used prospectively
during the introduction of PT in esophageal cancer (European PROTECT phase III trial randomizing
between XT and PT, currently running in the Particle Therapy Interuniversity Center Leuven) and lung
cancer (XT and PT data from international collaborations) in order to detect XT/PT-related changes in
toxicity outcome. Innovative NTCP modeling will be developed to fit the new PT era data by focusing
on linear energy transfer maps. Weekly repeated 4DCT scans and adapted treatment plans will
inform the optimization of novel intra-treatment NTCP models for the adaptation era. Finally, the
benefit the framework would have had in a setting of model-based patient selection for PT will be
quantified. For this purpose a health economic assessment will be associated to the NTCP model
framework, taking into account cost parameters to weigh the cost savings related to the predicted
decreased risk of side effects with PT and its higher operational costs compared to XT.