Artificial neural networks: laboratory aid or sorcerer's apprentice?

Artificial neural networks (ANNs) are being explored increasingly in medicine as decision-making aids. In the last 2 years, >630 articles dealing with the use of ANNs have been listed in Medline, including several recent reviews with application to the clinical laboratory [1, 2]. Is the interest and enthusiasm warranted , or will this introduction mirror that of other diagnostic tests in medicine: initially promising performance followed by disappointment as the tool emerges from the research laboratory into practical use? In this issue of Clinical Chemistry, J#{248}orgensen et al. [3] and Pedersen et at. [4] present two studies on ANNs. The first serves as a tutorial on ANNs, while exploring practical methodological issues surrounding the use and efficient implementation of an ANN. The second studies the use of ANNs in the diagnosis of acute myocardial infarction. Both articles compare ANN performance with that of the more common statistical technique of discriminant analysis. These two articles illustrate well the strengths, the pitfalls, and the issues surrounding the use of ANNs, and should be of interest to those wishing to learn more about ANNs. Widely available ANN software and the decreasing cost of computing power have resulted in blossoming interest in the use of ANNs. Although research in ANNs began in the 1950s, the level of interest they received has varied over the years. In the late 1960s, Minsky and Papert [5] demonstrated the limitations of the original approach, showing it to have little advantage over traditional statistical methods. In the I 980s, ANN methodology was changed to accommodate these criticisms and interest was renewed [6]. The general procedure today is similar to the original in " training " the ANN with a large number of example cases. In medical applications, one typically feeds it diagnostic parameters (e.g., presence of chest pain, serum creatine kinase concentration, etc.) and the best diagnosis available (e.g., whether myocardial infarction was present or not.) The ANN then iteratively changes the weighting of each input variable to produce a combined score, designed to be related to the diagnosis. In the later modification, the ANN " learns " by modifying a new set of internal mathematical variables derived from the original parameters until it is able to correctly diagnose as many cases as possible. These internal variables, called " hidden nodes, " allow ANNs to circumvent Minsky and Pap-ert's criticisms. They also theoretically provide ANNs superior performance over traditional …