When a patient complaining of chest pains is received by the emergency room, it is no simple matter to diagnose a heart attack. Merely examining the patient and performing an electrocardiogram (EKG) is not often enough. If a patient is suspected of having experienced a heart attack, several blood samples are drawn over the next 4 to 48 hours. Patients with heart tissue damage will have various cardiac enzymes appear in their blood. There is a characteristic pattern of the change in enzyme levels during the short period after a heart attack. Using all three techniques (EKG, exam, and blood analysis), a doctor can diagnose and treat heart attack patients. Neural network methods were found to correlate closely with expert human analysis, providing another opinion doctors can use to make a correct and timely diagnosis.
A physician at St. Joseph Mercy Hospital in Michigan designed a neural network that recognizes cases of acute myocardial infarction (AMI, commonly called heart attack) using the cardiac enzyme data from series of tests on patients.[1] The input consisted of two sequential cardiac enzyme tests and the number of hours between the tests. The output was "1" if the patient had a heart attack and "0" if the patient did not. The network was trained with 185 examples from 47 patients using blood tests that were not more than 48 hours apart. There were a total of 21 inputs and 1 output as shown below. The network was trained to a 10% error tolerance on all training data.
The neural network was then tested on 53 new sets of data. The data represented sets of serial cardiac enzyme data for ten patients with AMI and eight patients without AMI. Neural network outputs of less than 10% probability of AMI were classified as no-AMI cases. Outputs of at least 90% probability of AMI were classified as AMI cases. Outputs between 10% and 89% were interpreted as ambiguous or uncertain.
The neural network's diagnosis was then compared to three experts. One evaluated patients on the basis of ECHO/EKG changes. Another used the cardiac enzyme data. A third used autopsy reports. The network agreed with 100% of the AMI cases diagnosed by the cardiac enzyme expert, and 93% of the non-AMI cases. The 7% difference occurred where the network was uncertain. The network agreed with 86% of the AMI cases diagnosed by the EKG expert, and 33% of the non-AMI cases. In one case the EKG data was misleading due to multiple past heart attacks. In another case the network was uncertain. The network agreed with the autopsy results in 92% of the AMI cases, and 67% of the non-AMI cases. In one case the network was uncertain, and in another the heart had been damaged but by another cause.
INPUTS
1st test: total CK
1st test: CK type #1
1st test: CK type #2
1st test: CK type #3
1st test: total LDH
1st test: LDH type #1
1st test: LDH type #2
1st test: LDH type #3
1st test: LDH type #4
1st test: LDH type #5
2nd test: total CK
2nd test: CK type #1
2nd test: CK type #2
2nd test: CK type #3
2nd test: total LDH
2nd test: LDH type #1
2nd test: LDH type #2
2nd test: LDH type #3
2nd test: LDH type #4
2nd test: LDH type #5
hours between tests
OUTPUT
AMI probability
Reference:
J. Furlong, MD, M. Dupuy, MD and J. Heinsimer, MD, Neural Network Analysis of SerialCardiac Enzyme Data, AJ Clinical Pathology vol.96, no.1, July 1991 pp.134-141.