Clinical-Pathological Classification of Glioblastomas Investigated by a Non-Supervised Neural Network

Abstract

Using a variant of unsupervised neural networks (Self-Organizing Maps, SOM), it is intended to examine the ability to reproduce a clinical-pathological classification of patients with glioblastomas. This SOM provides a powerful means to visualize and analyze complex data sets without prior statistical knowledge. 58 consecutive cases were selected for evaluation. The single criterion for the selection of the cases was the survival of the patients. The SOM is realized by a two-dimensional hexagonal grid (a map) with 2047 nodes or neurons. The nodes on the grid with 6 clinical-pathological variables of glioblastomas gradually adapt themselves to the intrinsic shape of the data distribution. After the operation 25 patients were treated with conventional radiotherapy and 17 patients with radiotherapy and chemotherapy, 16 Patients received neither radiation nor chemotherapy. The minimal survival was < 1 month, the maximal survival 26 months. Viscovery SOM displays colored map regions, in our study patients with glioblastomas, grouped in clusters. Two clusters show the maximal significance. In a small cluster the glioblastoma patients with the long survival are grouped. These patients have been treated with radiotherapy and chemotherapy. The second cluster contains all other patients. The clusters can be compared to clusters of sex, age, survival and histological malignancy. The SOM allows a specific visual evaluation of new treatments and a more effective comparison with established tumor management.