Prostate Cancer Prognosis Evaluation Assisted by Neural Networks

Neural networks (NN) are new promising tools that can assist the clinicians in the diagnosis process and in therapy decision making, because they can deal with a great number of parameters, learning from examples and assessing any nonlinear relationships between inputs and outputs. In this paper, the problem of prostate cancer evolution prediction is approached using NN. The original database contained 650 records of patients, which underwent radical prostatectomy for prostate cancer. The NN variables were the parameters with the highest prognostic value selected and pre-processed from the original database. Different NN architectures and NN parameters have been tested in order to obtain the best complexity/accuracy ratio. The input data were structured, according to the latest statistical and representation concepts used in the current medical practice, aiming to improve the global performance. Different experiments were done using the rough database and the structured database. The NN performances were compared with the most widely used prediction statistical method, the logistic regresion. All NN models performed better than the logistic regression. The best obtained global prediction of correct classification 96.94% is better than the results of similar experiments available in literature. The NN prediction performance might be improved, because, in our opinion, its limits are given by the relatively small number of cases and the methods of collecting data.