Breast cancer survival prediction is presently based on several traditional prognostic factors such as tumor size, lymph node involvement, estrogen and progesterone receptor status and histologic grade. New prognostic factors such as proliferative index, cathepsin D, genetic mutations, and expression of tumor associated antigens are under study in predicting outcome for node-negative patients. Neural networks have been highly successful in analogous multi-dimensional pattern classification problems in many engineering applications and may be useful in breast cancer relapse and survival prediction and related clinical decision making. Furthermore, as new prognostic factors are easily proven, they could be easily incorporated within the neural network methodology. Exploring these ideas as the basis of a new project, a back error propagation neural network to predict patient disease-free survival using the classical, well-established parameters listed above was developed. Training and testing the network was performed with clinical data derived from 170 cases of breast cancers in the Torino, Italy area. We are grateful to Drs. Cappa, Liscia and Gaglia of the s. Giovanni Hospital in Torino for collecting and sharing their data. Several concerns arose when transporting neural network technology to clinical medicine classification problems such as breast cancer relapse and survival prediction. These concerns include: sparseness of data usually encountered in medical applications and the need for confidence measures and explanations in medical decisions-- information neural networks do not usually provide. Our experience has led us to several conclusions: (1) input parameters should be scaled, (2) weights should be computed at each significant time point, (3) improved weight convergence methods are needed, (4) single patient survival versus time plots are desirable outputs, (5) confidence intervals are needed and can be computed with bootstrapping methods, (6) larger data bases are needed to establish meaningful confidence intervals, (7) higher performance computing is needed to process larger data sets and confidence intervals, and (8) post-censor survival probabilities in training data can be completed using Kaplan-Meier survival analysis methods.
