Prospects for clinical decision support in breast cancer based on neural network analysis of clinical survival data

The paper illustrates the potential contributions of neural networks to a clinical decision support framework for prognosis or for prediction of therapy response using real breast cancer data. Clinical data poses special problems for training of neural networks due to the inherently limited information available in biological systems. Moreover, in breast cancer and many other applications, the probability model for the neural network must be specially conceived to account properly for censoring of follow-up data. The network presented was trained to learn the risk of relapse based on classical factors that are available in most hospitals. Follow-up data from 745 primacy breast cancer patients was used to train the neural model and to estimate a linear proportional hazards model. Although the scores due to the two models are highly correlated, the neural net score performed better in separating high-risk and low-risk groups than the proportional hazards score. This enhanced performance could be due to factor interactions detected by the network.