Learning target-based preferences through additive models: An application in radiotherapy treatment planning

Abstract

This article presents a new Multi-Criteria Decision Aiding preference disaggregation method based on an asymmetric target-based model. The decision maker's preferences are elicited considering the choices made given a set of comparisons among pairs of solutions (the stimuli). It is assumed that the decision maker has a reference value (target) for the stimulus. Solutions that do not comply with this reference value for some of the criteria dimensions considered will be penalized, and an inferred weight is associated with each dimension to calculate a penalty score for each solution. One of the differentiating features of the proposed model when compared with other existing models is the fact that only solutions that do not meet the target are penalized. The target is not interpreted as an ideal solution, but as a set of threshold values that should be taken into account when choosing a solution. The proposed approach was applied to the problem of choosing radiotherapy treatment plans, using a set of retrospective cancer cases treated at the Portuguese Oncology Institute of Coimbra. Using paired comparison choices made by one radiation oncologist, the preference model was built and was tested with in-sample and out-of-sample data. It is possible to conclude that the preference model is capable of representing the radiation oncologist's preferences, presenting small mean errors and leading, most of the time, to the same treatment plan chosen by the radiation oncologist.