On the Evaluation of Probabilistic Networks

摘要

As more and more probabilistic networks are being developed for medical applications, the question arises as to their value for clinical practice. Often the clinical value of a network is expressed as the percentage correct of predicted overall outcome, based upon an evaluation study using real-life patient data. In this paper, we propose another method of evaluation that focuses on intermediate outcomes of interest. We illustrate this method for a real-life probabilistic network for the staging of oesophageal cancer and show that it can provide valuable information in addition to a percentage correct.