Application of artificial neural networks to clinical
medicine
1135
A 51-year-old man walks into an emergency departmentwith mild
left anterior chest pain. He denies anysignificant personal or
family history of heart disease andhe has no other cardiac risk
factors. Physical examinationreveals minimal bibasilar rates and is
otherwise normal.An electrocardiogram (ECG) reveals slight
T-waveflattening. The doctor then enters all the
pertinentinformation accrued from the patient into the
on-lineemergency department patient medical record andmanagement
system. That system includes a neuralnetwork 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 amyocardial infarction. The physician combines
thisinformation with his own impression and admits thepatient to
the coronary-care unit. 4 h later cardiac enzymestudies confirm
acute myocardial infarction.This scenario is based on a real
network, representing
one of the first applications of artificial neural networks
toclinical medicine. Preliminary studies have revealed thatthis
network is more accurate than physicians inidentifying acute
myocardial infarction in patientspresenting to the emergency
department with anteriorchest pain.Many biological and
pathophysiological processes
manifest "chaotic" behaviour. Chaotic processes are notbest
analysed by classical linear methods. The first articlein this
series, by Cross and colleagues, explained hownon-linear
processing, as afforded by an artificial neuralnetwork, might
improve upon the predictive power of theother approaches. The hope
has been that the networksability to identify multidimensional
relationships inclinical data not apparent to other forms of
analyseswould allow the network to improve diagnostic
accuracy.Non-linear statistical methods have been tried before
butthey were complex and computationally intensive andnever became
widely accepted. The advent of the artificialneural network and the
much greater speed of todayscomputers have changed this.
Clinical diagnosis became one of the first areas towhich the
artificial neural network was applied. Acutemyocardial infarction
was one of the earliest applicationsbut the range is wide, from
appendicitis to theexamination of biopsy specimens (panel 1).
Starting withmyocardial infarction this review will cover some of
thoseclinical applications. The use of neural networks in
thelaboratory will be covered by Dybowski and Gant in thelast
article in this Lancet series.2
Myocardial infarctionMyocardial infarction is especially
challenging because itis a disease of low incidence yet a very high
price has to bepaid for its misdiagnosis. As a result, physicians
havetended to err on the side of safety: they push theirLancet
1995; 346: 1135-38
Department of Emergency Medicine, University of
PennsylvaniaMedical Center, Philadelphia, PA 19104-4283, USA (W G
Baxt MD)
diagnostic sensitivity as high as possible, leading to
areduction in specificity and the unnecessary admission tohospital
of significant numbers of patients who have nothad a myocardial
infarction. A large study published in1988 put physicians
sensitivity at 88% and specificity at71%."The first application of
the artificial neural network to
chest pain appeared in 1989.12 This work trained amultilayer
network on 174 patients presenting withanterior chest pain and it
put patients into one of threediagnostic groups-high risk cardiac,
low risk cardiac,and non-cardiac. However, these areas were not
definedby any standard criteria. Another application was basedon a
retrospective study of a set of 356 patients admittedto a cardiac
intensive care unit. 120 had had a myocardialinfarction. The
network was trained, using back-propagation, on half of the
patients with and withoutmyocardial infarction, and it was tested
on the remainingpatients, to whom the network had not been
exposed.
Clinical pharmacologyPredicting: tumour sensitivity to drugs,
patients 58-60response to warfarin, and central-nervous-system
activity of alfentanil
