The effect of sample size and MLP architecture on Bayesian learning for cancer prognosis--a case study

In this paper we investigate the independent effects of training sample size and multilayer perceptron (MLP) architecture on Bayesian learning to build prognostic models for metastatic breast cancer. We trained two types of Bayesian neural networks on a data set of 1477 metastatic breast cancer patients followed at the Institut Curie using disjoint training sets of sizes k = 50, 100, 200, 300, and 450. The learning performance as measured by an expected loss appeared independent of the two architectures modelling the log hazard function under either proportional or non proportional hazard assumptions, thus indicating that no other sources of nonlinearity besides interactions are present. We found a performance breakdown at k = 50, and no sample size effect for k > or = 100.