An evolutionary approach to constructing prognostic models.

Abstract

A prognostic model is sought to determine whether or not patients suffering from an uncommon form of cancer will survive. Given a set of case histories, we attempt to find the relative weightings of the different variables that are used to describe the cases. Our first innovation is to use a diffusion genetic algorithm (DGA) to find weightings which will give optimal survival predictions. The DGA enables a number of criteria to be satisfied simultaneously, making it particularly suitable for model building. A further innovation is a method of representing synergies between interacting factors. The evolved model correctly predicts 90% of the survivors and 87% of deaths, an improvement over the current model. More significantly, the method enables a simple model to be evolved, one that produces well-balanced predictions, and one that is relatively easy for clinicians to use. The method was validated by running it on a training set made up of 90% of the original database and then studying the performance of the generated models on a test set consisting of the remaining 10% of the cases.