Sequential decision-making using feedforward neural networks: applications in surgery

Abstract

In clinical decision-making, sequential decisions are very common. To base these decisions on clinical scores or test results makes them accessible for the use of formal statistical methods. We investigate the problem of constructing a sequential decision procedure allowing for a classification into two (risk) groups at an average minimum cost or at the earliest possible time, and on the condition that the number and sequence of test scores is fixed in advance. The quality of the procedure is maintained by specified upper bounds for the conditional errors of the entire procedure. Focus is mainly on the case that a fixed number and sequence of K continuous test scores is given. The proposed decision theoretic model is based on, but not restricted to, K different feedforward neural networks (multilayer perceptron and radial basis networks) which form continuous scores for each decision step. The advantages of the method are demonstrated by two prospective longitudinal studies from the Medical School Hanover.