Connectionist models in medicine: an investigation of their potential

Abstract

The subjectiveness of patient data and the incompleteness of doctors’ causal models has led to a wide variety of techniques being used to build medical decision-aids. We briefly describe problems resulting from the use of conventional methods for building an advisor for chest pain, before exploring the way in which doctors’ decisions might be modelled using connectionist (neural network) techniques. We then examine the claims and the realities of current connectionist models. To determine if such techniques are promising for this domain, a series of proven 3-layer and a 4-layer backward- propagation networks were repeatedly trained on data from 174 chest pain cases. The best networks were then assessed on fresh data from a new set of 73 cases. The 3-layer network performed better, but still had a clinically unacceptable crude accuracy of 70% and false negative rate of 27%. We are concerned about the undisciplined use of connectionist systems, and about the problems of system validation. There is a need for further work on the provision of informative diagnostics and statistics for data exploration and training. Given this greater understanding, connectionist systems might have much to offer as components of medical decision-aids.