Automatic decision using dirty databases: Application to prostate cancer diagnosis

Abstract

Currently, the best way to reduce the mortality of cancer is to detect and treat it in its early stages. Automatic decision support systems, such as automatic diagnosis systems, are very helpful in this task but their performance is constrained by the integrity of the clinical input data. This could be a problem since clinical databases, in which these systems are based on, are commonly built up containing dirty data (empty fields, non-standard or normalized values, etc). This article presents a study of the performance of a clinical decision support system, based on an artificial neural networks, using sets of clean and dirty prostate cancer data. The study shows that is possible to obtain an implementation that allow us to avoid the problems associated to the database's lack of integrity and reach a similar performance using either clean or dirty data.