ABSTRACT
This chapter discusses general potentials and fundamental limits of prediction models for side effects in radiotherapy, in particular those models that result from data-driven analysis. The explanatory variables that are usually observed are items that are already recorded in the treatment process or variables that are known or expected to be potentially predictive. A particular property of dosimetric data is that usually dose distributions are very similar in comparable patients, resulting in a limited range of observed dose distributions and high correlations between dosimetric variables. Model fitting, which is at the heart of phenomenological modeling, is the act of adjusting model parameters such that the model predictions correspond as good as possible to the actually observed outcomes. One of the great advantages of phenomenological modeling is its potential to explore large datasets and to discover new, potential, or better predictors.
