Probabilistic Networks as Probabilistic Forecasters

Abstract

To establish its clinical value, a probabilistic network is typically subjected to an evaluation study using real patient data from the field of application. The results of such a study are often summarised in the percentage of correctly predicted outcomes. In this paper, we propose the use of a forecasting score as an alternative way of expressing the clinical value of a network. Such a score takes not just the predicted outcome into consideration but also the associated distribution of uncertainty. We illustrate the use and interpretation of the Brier forecasting score for a real-life probabilistic network in oncology.

This research is (partly) supported by the Netherlands Organisation for Scientific Research (NWO).