Mattias Ohlsson and Lars Edenbrandt
Case-based sensitivity analysis for artificial neural networks with applications in medicine
In Proceedings of the Neural Networks and Expert Systems in Medicine and Healthcare Conference, 102-106 (2001)
It is often important to be able to explain the reasoning behind a machine learning algorithm, such as an artificial neural network, especially in the medical domain. In this paper we develop a case-based sensitivity analysis method for neural networks. The method uses a trained neural network committee in order to find both important and unimportant input variables for individual cases. The sensitivity analysis is formulated as combinatorial optimization problem, where the mean field annealing method is used as a tool for finding good solutions. The approach is tested on a problem from the medical domain; namely the problem of identifying patients suffering from acute myocardial infarction in the presence of left bundle branch block. We feel that the case-based sensitivity analysis developed here can be used to understand the complex functioning of a neural network and that the method can be applied on other problems from the medical domain.
LU TP 01-19