A 51-year-old man walks into an emergency department

with mild left anterior chest pain. He denies any

significant personal or family history of heart disease and
he has no other cardiac risk factors. Physical examination
reveals minimal bibasilar rates and is otherwise normal.
An electrocardiogram (ECG) reveals slight T-wave
flattening. The doctor then enters all the pertinent
information accrued from the patient into the on-line
emergency department patient medical record and
management system. That system includes a neural
network trained to identify acute myocardial infarction.
The network analyses this information and relays the fact,
via an on-screen window, that this patient has sustained a
myocardial infarction. The physician combines this
information with his own impression and admits the
patient to the coronary-care unit. 4 h later cardiac enzyme
studies confirm acute myocardial infarction.
This scenario is based on a real network, representing
one of the first applications of artificial neural networks to
clinical medicine. Preliminary studies have revealed that
this network is more accurate than physicians in
identifying acute myocardial infarction in patients
presenting to the emergency department with anterior
chest pain.
Many biological and pathophysiological processes
manifest "chaotic" behaviour. Chaotic processes are not
best analysed by classical linear methods. The first article
in this series, by Cross and colleagues, explained how
non-linear processing, as afforded by an artificial neural
network, might improve upon the predictive power of the
other approaches. The hope has been that the networks
ability to identify multidimensional relationships in
clinical data not apparent to other forms of analyses
would allow the network to improve diagnostic accuracy.
Non-linear statistical methods have been tried before but
they were complex and computationally intensive and
never became widely accepted. The advent of the artificial
neural network and the much greater speed of todays
computers have changed this.
Clinical diagnosis became one of the first areas to
which the artificial neural network was applied. Acute
myocardial infarction was one of the earliest applications
but the range is wide, from appendicitis to the
examination of biopsy specimens (panel 1). Starting with
myocardial infarction this review will cover some of those
clinical applications. The use of neural networks in the
laboratory will be covered by Dybowski and Gant in the
last article in this Lancet series.2

Myocardial

diagnostic sensitivity

as

high

as

possible, leading

to

reduction in specificity and the unnecessary admission to

hospital of significant numbers of patients who have not
had a myocardial infarction. A large study published i
1988 put physicians sensitivity at 88% and specificity a
71%."
The first application of the artificial neural network to
chest pain appeared in 1989.12 This work trained a
multilayer network on 174 patients presenting with
anterior chest pain and it put patients into one of three
diagnostic groups-high risk cardiac, low risk cardiac,
and non-cardiac. However, these areas were not defined
by any standard criteria. Another application was based
on a retrospective study of a set of 356 patients admitted
to a cardiac intensive care unit. 120 had had a myocardial
infarction. The network was trained, using backpropagation, on half of the patients with and without
myocardial infarction, and it was tested on the remaining
patients, to whom the network had not been exposed.

infarction

is especially challenging because it

disease of low incidence yet a very high price has to be
paid for its misdiagnosis. As a result, physicians have
tended to err on the side of safety: they push their

Myocardial infarction
is

Lancet 1995; 346: 1135-38

Clinical pharmacology

Predicting: tumour sensitivity to drugs, patients

58-60

response to warfarin, and central-nervous-

Department of Emergency Medicine, University of Pennsylvania

Medical Center, Philadelphia, PA 19104-4283, USA (W G Baxt MD)

system activity of alfentanil

1135