Clinical Chemistry 43: 1919-1925, 1997;
(Clinical Chemistry. 1997;43:1919-1925.)
© 1997 American Association for Clinical Chemistry, Inc.
Early assessment of patients with suspected acute myocardial infarction by biochemical monitoring and neural network analysis
Johan Elleniusa,
Torgny Groth,
Bertil Lindahl1 and
Lars Wallentin1
Department of Biomedical Informatics and Systems Analysis, and
1
Department of Cardiology, University of Uppsala, Uppsala, Sweden.
2
The term "validation set" is used to denote a set of example cases used to tune the parameters of a classifier.
a Address correspondence to this author at: Department of Biomedical Informatics and Systems Analysis, University hospital, S-751 85 Uppsala, Sweden. Fax +46 18–531202; e-mail Johan.Ellenius{at}BMSA.uu.se
 |
Abstract
|
|---|
Neural network analysis was applied for early diagnosis/exclusion of
acute myocardial infarction (AMI), prediction of infarct size, and
estimation of "time from onset of infarction." Eighty-eight
patients admitted within 8 h after onset of chest pain were
included. Blood samples for measurement of myoglobin, creatine kinase
isoform MB, and troponin T were obtained every 30 min during the first
3 h and then after successively longer intervals. Data from 50
patients were used to train a set of neural network components of a
decision support system. The performance of the system was evaluated
and compared with experienced clinicians for the remaining 38 patients.
The computer system detected myocardial infarction and predicted
infarct size earlier than the clinicians, but did not differ
significantly in terms of diagnostic sensitivity, specificity, and
predictive values when disregarding time for diagnosis. With a
cross-validation procedure the cumulated sensitivities of the computer
system for the first five measurements were estimated to be (mean
± 2SEM, n = 100): 0.77 ± 0.03, 0.89 ± 0.02, 0.94
± 0.02, 0.97 ± 0.01, and 0.99 ± 0.01, respectively, with
corresponding cumulated specificities between 0.93 ± 0.01 and
0.91 ± 0.01. We concluded that neural network analysis of serial
measurements of biochemical markers might provide useful support for
the early assessment of patients with suspected AMI.
|
Introduction
|
|---|
An early diagnosis within the first hours after onset of symptoms
is essential for the appropriate handling and optimal treatment of
patients admitted with chest pain of suspected cardiac origin
(1)(2). The 12-lead electrocardiogram (ECG) is
usually immediately available, and in cases with typical ST-segment
elevations the diagnosis of acute myocardial infarction (AMI) is
straightforward.2
However, in the vast majority of patients
with chest pain the 12-lead ECG is nondiagnostic on admission. In these
patients ruling in and ruling out the diagnosis of AMI must be on the
basis of repeated measurements of biochemical markers, currently a
time-consuming procedure. Therefore, in recent years much interest has
been focused on new biochemical markers for early rule-in and rule-out
of AMI (3)(4). Although the diagnostic
sensitivity and specificity for AMI are generally high for these
markers (5)(6)(7)(8), they are not sensitive enough for ruling
out AMI immediately upon admission (9). In the
heterogeneous group of patients with chest pain, considerable economic
resources might be saved by early identification of those patients
(~60–70%) who are at sufficiently low risk of AMI and its
complications to be transferred to a general ward outside the coronary
care unit (CCU) (3).
Several different methods have been proposed to achieve some of these
goals, including diagnostic algorithms based on clinical data
(10)(11) or biochemical markers
(12)(13). The use of artificial neural
networks is an alternative technique that has been applied for AMI
diagnosis based on clinical data including ECG
(14)(15) and biochemical markers
(16)(17). Artificial neural networks can be
used to classify input patterns of measured variables as belonging to
one of several predefined categories by pattern recognition. This is
achieved by "training" of the network with a number of
representative example cases and their corresponding correct
classifications as defined by a gold standard. The network will
correctly classify a major portion of the patterns in this training set
by a procedure in which the parameters of the network (the connection
weights and biases) are adjusted to minimize a cost function measuring
the difference between the network's classification and the gold
standard (so-called "supervised learning"). When exposed to a new,
not yet classified case, the network uses its previous training and its
capability of generalization for assignment to a specific diagnostic
category. For further reading see ref. 18.
The aim of the present investigation was to apply artificial neural
network methods, based on frequent sampling and measurement of selected
markers of AMI, for (a) early diagnosis or exclusion of AMI,
(b) early prediction of infarct size, (c)
estimation of time from onset of infarction, and (d) to
compare the performance of the neural networks with experienced
physicians.
|
Materials and Methods
|
|---|
patients
The present study was part of a Swedish multicenter study
(BIOMACS) (13). Included in the present analysis were all
patients admitted to the participating CCUs with chest pain, and with
onset within the previous 8 h and, in case of AMI, not given
thrombolytic treatment; in total, 88 patients. The study was approved
by the local ethics committees, and all enrolled patients provided
informed consent.
laboratory investigations
Standard 12-lead ECGs were recorded on admission and after 24 and
48 h. Blood samples were drawn from an indwelling forearm venous
catheter, the first sample on admission and thereafter every 30 min
during the first 3 h and then at 4, 6, 9, 12, 18, 24, and 48
h after admission. The mass concentration of creatine kinase isoenzyme
MB (CK-MB) in plasma was measured with the NovoClone system (Novo
BioLabs) (19); the upper reference limit for healthy
individuals is 7 µg/L and the total analytical CV was estimated to be
8.4% at a concentration of 25 µg/L (n = 154). The plasma
myoglobin mass concentration was determined with a modified RIA method
developed by Roxin et al. (20). The upper reference limits
are 57 µg/L for women and 90 µg/L for men. The total analytical CV
for this modified method was 6.1% at a concentration of 61 µg/L
(n = 49). The mass concentration of troponin T (Tn-T) in plasma
was determined with the Enzymun-Test system (Boehringer Mannheim)
(21). The total analytical CV was estimated to 6.3% at a
concentration of 3.8 µg/L (n = 56). All plasma samples and ECGs
were centrally analyzed and interpreted, without knowledge of the
patients' diagnoses or outcomes.
gold standard
AMI was considered present if at least two of the three WHO
criteria were fulfilled (22). The infarct size was
arbitrarily labeled "major" if the peak plasma CK-MB mass
concentration was >80 µg/L and "minor" if peak CK-MB mass
concentration was 
80 µg/L.
computational methodology
The system was designed to select neural network architectures of
minimal complexity. A balanced version of the cross-validation
method (cf. (23)(24)) was used to
select an adequate neural network for each classification problem: Data
from all 88 patients were randomly partitioned into a training set
(67%) and a cross-validation set (33%) with realistic prevalences of
major AMI, minor AMI, and non-AMI in the two sets. The procedure was
repeated 20–100 times for each architecture and the mean and standard
error (SEM) of diagnostic sensitivity, specificity, and positive and
negative predictive values were calculated. Each candidate neural
network was trained and evaluated on these sets.
The measured values of the biochemical markers were normalized by
division with the respective upper reference limits to facilitate
transferability of the method. The values were further fuzzified with
piecewise linear membership functions selected individually for the
input variables to achieve as good univariate separation as possible
between the diagnostic classes in the training set. The membership
functions were characterized by two parameters,
cmin and cmax,
transforming normalized values lower than cmin
to 0, higher values than cmax to 1, and linearly
increasing values from 0 to 1 in the interval
(cmin; cmax). This type
of preprocessing makes it possible to perform nonlinear classification
with simple single-layer perceptrons (cf. (25)), the basic
functional unit of a neural network. Neural network architectures of
diverse complexity [single-layer perceptrons (SLPs) and multi-SLPs
with fuzzified input variables, multilayer perceptrons (MLP) with 2 and
3 hidden units, and Elman recurrent
networks1
with 2 and 10 feedback units] were evaluated. The preprocessed
measured values of the infarct markers (myoglobin, CK-MB, and Tn-T) and
in some cases also their respective rate of change, expressed as the
difference 
to previous value, were used as input variables. The
number of markers was also varied to find the optimum set of input
variables. The SLPs and the MLPs all used sigmoidal transfer functions
and were trained with the Levenberg–Marquardt method to minimize a
cost function (sum of squared residuals) as implemented in the Neural
Network Toolbox 2.0 (The Mathworks). Recognizing that the phenomenon
known as "overtraining" does not occur in models of sufficiently
low complexity (cf. (24)), the SLPs were trained until the
training set was classified with a minimum error. The training of the
MLPs was stopped, to avoid overtraining, when the performance of the
networks was optimal, as measured by the objective function of the sum
of squared residuals on the cross-validation set used as a validation
set4. In this case the performance of the MLPs thus
approximates an upper bound of what is attainable with the technique of
"early stopping" for maximizing generalization. The
back-propagation algorithm (in the Neural Network Toolbox 2.0) was used
for training the Elman recurrent networks. The conventional method for
early stopping, with a validation set during training for monitoring of
the generalizing capability of the network, was not necessary because
the performance on the cross-validation sets (used as validation sets)
measured by the sum of squared residuals generally settled at a
constant concentration after the training set was classified with a
minimum error. Thus, the Elman networks as used in this application
were considered not to be overtrained.
To meet the requirement of early detection of AMI, the candidate neural
networks were tuned with ROC curve analysis to rule in all AMI patients
within 2 h (five sets of measurements) after admission. The
preferred neural network architecture was selected by comparing the
respective cumulated sensitivity and specificity of the candidate
networks for each consecutive set of measurements ((1)(2)(3)(4)(5)). The
preferred neural networks for prediction of infarct size and estimation
of time from onset, respectively, were selected on the basis of highest
sensitivity at comparable specificities (>0.80) with ROC curve
analysis.
The selected neural networks and sets of input variables for the
various classification problems are summarized in Table 1
. Tn-T did not contribute significantly to the diagnostic
performance of the neural network classifiers and was therefore
excluded from the list of input variables. Three SLPs were used in
parallel (multi-SLP) for classification of AMI/non-AMI; each SLP was
individually trained to detect AMI on the basis of measurements from a
specified time interval (<6 h; 6–8 h; >8 h) since onset of
infarction, thus making it possible to handle the nonlinearity due to
the pronounced temporal variation in concentrations of the selected
biochemical markers. If the output from any one of the individual SLPs
should indicate an AMI, that was taken to be the actual state. If AMI
was not detected after a predefined optimal period of monitoring (2 h
or five sets of measurements), AMI was excluded. Prediction of infarct
size was performed with a SLP without fuzzification (identical to
logistic regression). Estimation of the "time from onset,"
performed when AMI was diagnosed, involves a SLP with fuzzification and
3 output units trained to indicate whether the onset occurred before a
specified number of hours after infarction (5 h; 6 h; 7 h). The
training and the cross-validation sets were both extended in this case;
all sets of measurements from patients correctly classified as AMI and
measured within 12 h from onset of infarction were included. The
predictive values were calculated according to realistic prevalences
from the Uppsala University Hospital/CCU database.
methods for comparison with clinicians
For comparing the performance of the computer method with the
clinicians, the 88 patients were divided into one "training set" of
50 patients and one "test set" of 38 patients. These two sets are
presented in Table 2
. The characteristics of the AMI patients are further detailed
in Table 3
. The relative numbers of AMI and non-AMI in the test set are
fairly close to the corresponding figures of 22% and 78% calculated
from the Uppsala University Hospital/CCU database (RiksHIA) for the
same category of patients admitted to the CCU during 1994 (with arrival
<8 h after onset of symptoms and without thrombolytic therapy). Three
clinicians with several years of experience of working in the emergency
department and the CCU were presented with the series of measured
values of biochemical markers in time order by using a computer. They
were asked to classify the patients as AMI or non-AMI and to predict
the infarct size. To make the evaluation clinically relevant, the
clinicians also had access to the 12-lead ECG on admission, in
parameterized form, as well as the time from onset of symptoms, age,
and sex. By following this procedure, the assessments put forward by
the clinicians could be compared with the assessment by the computer
method in terms of conventional measures of diagnostic performance.
View this table:
[in this window]
[in a new window]
|
Table 2. Clinical characteristics of patients in the training and
test sets used in comparison of the computer method with the
clinicians.
|
|
View this table:
[in this window]
[in a new window]
|
Table 3. Characteristics of AMI patients in the training and test
sets used in comparison of the computer method with the
clinicians.
|
|
methods for comparison with empirical rules for rule-in/-out of ami
The performance of the computer method was also compared with the
empirical rules proposed in a previous report from the BIOMACS study
(13); one rule for the early assessment of AMI: plasma
CK-MB 
20 µg/L or myoglobin above the upper reference limit and
CK-MB 10 µg; a second rule for exclusion of AMI 3 h after
admission: no sample of myoglobin above the upper reference limit; and
a third rule for the exclusion of AMI 6 h after admission: no
sample of CK-MB 8 µg/L.
|
Results
|
|---|
comparison with clinicians
In the test set of 38 patients, the computer method detected AMI
earlier than the clinicians; AMI was correctly detected on the first
measurement in 90% of the cases, compared with 70% for the
clinicians. For the remaining cases, the computer system required one
more measurement (additional 30 min) for correct diagnosis, and the
clinicians required on average two more measurements (additional 60
min). Sensitivity, specificity, and predictive values for detection of
AMI were high and not significantly different for the computer method
and the clinicians when disregarding the time for diagnosis (Table 4
).
The rate of correctly ruled-in AMI patients, estimated by the
cross-validation procedure and expressed as cumulated sensitivity
± 2SEM (n = 100), was 0.77 ± 0.03, 0.89 ± 0.02,
0.94 ± 0.02, 0.97 ± 0.01, and 0.99 ± 0.01 for the
first five measurements, respectively, with a corresponding cumulated
specificity between 0.93 ± 0.01 and 0.91 ± 0.01.
For exclusion of AMI, the computer method was optimized to use a fixed
number of five measurements. The clinicians required five measurements
on average but with a large variation (1–10 sets of measurements).
The computer method made a correct prediction of infarct size earlier
than the clinicians (Table 5
). The computer method and the clinicians did not differ
significantly (considering the low number of test cases) in terms of
diagnostic sensitivity, specificity, and positive and negative
predictive values (Table 6
).
comparison with empirical rules for rule-in/-out of ami
The empirical rules, when applied to the test set, required more
measurements than the computer method to rule in AMI. On admission,
80% of the cases were correctly detected. The remaining cases needed
two more measurements on average. At 3 h after admission, the
empirical rules had correctly ruled out 63% of the patients with
non-AMI in the test set, 26% had an uncertain diagnosis, and the
remaining 11% were incorrectly classified as AMI in the previous step.
The same figures for the computer system 2 h after admission were
93%, 0%, and 7%, respectively. At 6 h after admission, the
empirical rules had correctly ruled out 85% of the patients with
non-AMI, which was in parity with the computer method 4 h earlier.
estimation of time from onset of infarction
The performance of the computer method for diagnostic sensitivity,
specificity, and predictive values is presented in Table 7
.
|
Discussion
|
|---|
A computational method based on several individually trained
artificial neural networks was developed and validated on a clinically
relevant reference sample group of patients with suspected AMI. Our
results indicate that by using repeated blood sampling and analyses of
plasma myoglobin and CK-MB mass concentration during the first hours
after admission it is possible for the computer method not only to
reliably diagnose AMI and predict infarct size, but also to crudely
estimate the time from onset of infarction. Not surprisingly, Tn-T
measurements were found not to add to the diagnostic performance of
networks for early detection/exclusion of AMI, considering the similar
or slightly slower appearance of Tn-T compared with CK-MB in plasma
during the first 8 h. In this context, early Tn-T measurements are
of potential value for indication of old infarctions. Currently, its
most important role seems to be as a risk indicator of subsequent
cardiac events in patients with unstable coronary artery disease
(27). The patients in the test set were generally
correctly classified with regard to AMI and non-AMI, both by the
computer method and the clinicians. The major difference in performance
was in the time required to make a correct diagnosis. The mean time was
shorter and the variation smaller for the computer. In comparison with
an application of empirical rules based on plasma myoglobin and CK-MB
measurements (13), the neural network method was faster
both for detection and particularly for exclusion of AMI. This
comparison is biased in favor of the empirical rules, since they were
derived with the whole material (including the patients in the test
set). Performed very early, a crude prediction of the impending infarct
size might be of value for the clinician in deciding whether to use
infarct-limiting therapy (i.e., thrombolysis or acute percutaneous
transluminal coronary angioplasty). To achieve this, the computer
system was designed to perform the size prediction, requiring at most
one more set of measurements after the detection of AMI. The computer
system predicted infarct size approximately on the second measurement,
and 10 of 11 AMI patients were predicted correctly by the system. In
contrast, the clinicians required considerably longer time than the
computer system for predicting infarct sizes, four and six sets of
measurements on average for classification of major and minor
infarctions, respectively. Examples of the time curves for the
biochemical markers and time points when AMI was detected and size was
predicted by the computer and the three clinicians are shown in Fig. 1
. These results indicate that a correct early infarct size
prediction, although crude, is feasible with the computer method
suggested.
View larger version (32K):
[in this window]
[in a new window]
|
Figure 1. Time series measurement results with 95% confidence
intervals of infarct markers for a patient with (A) a major
AMI and (B) a minor AMI.
Time points are marked with arrows when the computer and the
three clinicians detected the infarction and predicted the infarct
size.  , myoglobin; •, CK-MB;  , Tn-T. Relative
concentration = plasma concentration (µg/L) divided by 90/57
µg/L (men/women) and 7 µg/L for myoglobin and CK-MB, respectively.
For Tn-T a decision limit of 0.2 µg/L was used, in lack of an
established upper reference limit for healthy individuals at the time
of this study.
|
|
Some patients have difficulties in stating the time for onset of chest
pain, and furthermore, the onset of chest pain might be an uncertain
estimate for onset of infarction in the individual patient
(28). Therefore, a "second opinion" from the computer,
based on the pattern of biochemical markers, might be of value, e.g.,
for the decision to give or not to give the patient thrombolytic
treatment. The results indicate that it might be possible to make an
interval estimate of the time from onset of infarction very early with
reasonable certainty.
Although thrombolytic treatment is only proven to be beneficial for
patients with ST elevation or bundle branch block in the ECG on
admission (29), it might well also be of benefit in
certain subgroups of patients with nondiagnostic ECGs
(30), e.g., patients with short delay and an impending
large infarction. Given the possibility to select such subgroups, e.g.,
with the use of neural network methods, elucidating the value of
thrombolysis in these subgroups will be an important area of research
(31).
Although the patient material was representative for a CCU population
and unique with the early frequent blood sampling, the number of
patients was limited in both the test and training sets. Therefore, the
results must be regarded as indicative of the potential benefit of a
neural network-based computer system in the early assessment of
patients with acute chest pain. Further studies in larger groups of
patients are needed.
|
Acknowledgments
|
|---|
We thank all of the clinicians and staff of each center involved
for their skillful assistance. This work was supported by grants from
the Swedish Heart and Lung Foundation; the Sellander's Foundation,
Uppsala; Pharmacia Biosensor AB, Uppsala, Sweden; and from the European
Commission under its AIM (Advanced Informatics in Medicine) Program,
contract A2028 Openlabs.
|
Footnotes
|
|---|
2 Nonstandard abbreviations: ECG, electrocardiogram; AMI,
acute myocardial infarction; CCU, coronary care unit; CK-MB, isoenzyme
of creatine kinase; Tn-T, troponin-T; SLP, single-layer perceptron; and
MLP, multilayer perceptron. 
1 3 "Elman networks" are a type of recurrent artificial neural networks with one hidden layer suitable for analysis of time series data [26].
|
References
|
|---|
-
Acute myocardial infarction: pre-hospital and
in-hospital management. The task force on the management of acute
myocardial infarction of the European Society of Cardiology. Eur
Heart J 1996;17:43–63..
-
Emergency department: rapid identification and treatment
of patients with acute myocardial infarction. National Heart Attack
Alert Program Coordinating Committee, 60 Minutes to Treatment Working
Group. Ann Emerg Med 1994;23:311–29..
-
Newby LK, Gibler WB, Ohman EM, Christenson RH. Biochemical markers in suspected acute myocardial infarction: the need for early assessment [Editorial]. Clin Chem 1995;41:1263-1265.
[Free Full Text]
-
Roberts R, Kleiman NS. Earlier diagnosis and treatment of acute myocardial infarction necessitates the need for a 'new diagnostic mind-set'. Circulation 1994;89:872-881.
[Abstract/Free Full Text]
-
Katus HA, Remppis A, Neumann FJ, Scheffold T, Diedrich KW, Vinar G, et al. Diagnostic efficiency of troponin T measurements in acute myocardial infarction. Circulation 1991;83:902-912.
[Abstract/Free Full Text]
-
Roxin LE, Cullhed I, Groth T, Hallgren T, Venge P. The value of serum myoglobin determinations in the early diagnosis of acute myocardial infarction. Acta Med Scand 1984;215:417-425.
[Web of Science][Medline]
[Order article via Infotrieve]
-
Mair J, Artner DE, Dienstl A, Lechleitner P, Morass B, Smidt J, et al. Early detection of acute myocardial infarction by measurement of mass concentration of creatine kinase-MB. Am J Cardiol 1991;68:1545-1550.
[Web of Science][Medline]
[Order article via Infotrieve]
-
Mair J, Morandell D, Genser N, Lechleitner P, Dienstl F, Puschendorf B, et al. Equivalent early sensitivities of myoglobin, creatine kinase MB mass, creatine kinase isoform ratios, and cardiac troponins I and T for acute myocardial infarction. Clin Chem 1995;41:1266-1272.
[Abstract/Free Full Text]
-
Bakker AJ, Koelemay MJW, Gorgels JPMC, van Vlies B, Smits R, Tijssen JG, Haagen FD. Failure of new biochemical markers to exclude acute myocardial infarction at admission. Lancet 1993;342:1220-1222.
[Web of Science][Medline]
[Order article via Infotrieve]
-
Lee TH, Juarez G, Cook EF, Weisberg MC, Rouan GW, Brand DA, Goldman L. Ruling out acute myocardial infarction. A prospective multicenter validation of a 12-hour strategy for patients at low risk. N Engl J Med 1991;324:1239-1246.
[Abstract]
-
Jonsbu J, Aase O, Rollag A, Liestol K, Erikssen J. Prospective evaluation of an EDB-based diagnostic program to be used in patients admitted to hospital with acute chest pain. Eur Heart J 1993;14:441-446.
[Abstract/Free Full Text]
-
Collinson PO, Rosalki SB, Flather M, Wolman R, Evans T. Early diagnosis of myocardial infarction by timed sequential enzyme measurements. Ann Clin Biochem 1988;25:376-382.
-
Lindahl B, Venge P, Wallentin L. Early diagnosis and exclusion of acute myocardial infarction by biochemical monitoring. Coron Artery Dis 1995;6:321-328.
[Web of Science][Medline]
[Order article via Infotrieve]
-
Baxt WG. Use of an artificial neural network for the diagnosis of myocardial infarction. Ann Intern Med 1991;115:843-848.
-
Baxt WG. Analysis of the clinical variables driving decision in an artificial neural network trained to identify the presence of myocardial infarction. Ann Emerg Med 1992;21:1439-1444.
[Web of Science][Medline]
[Order article via Infotrieve]
-
Furlong JW, Dupuy ME, Heinsimer JA. Neural network analysis of serial cardiac enzyme data. A clinical application of artificial machine intelligence. Am J Clin Pathol 1991;96:134-141.
[Web of Science][Medline]
[Order article via Infotrieve]
-
Pedersen S, Jorgensen J, Pedersen J. Use of neural networks to diagnose acute myocardial infarction. II. A clinical application. Clin Chem 1996;42:613-617.
[Abstract/Free Full Text]
-
Ripley B. Neural networks and related methods for classification. J R Statist Soc B 1994;56:409-456.
-
Jørgensen PJ, Hørder M, Selmer J, Bøtker HE. Analytical evaluation of a sensitive enzyme immunoassay for determinations of creatine kinase isoenzyme MB. Clin Chem 1990;36:1502-1505.
[Abstract/Free Full Text]
-
Roxin L-E, Venge P, Wide L. A fast and sensitive radioimmunoassay of human myoglobin for use in the early diagnosis of heart infarction. Clin Chim Acta 1980;107:129-134.
[Web of Science][Medline]
[Order article via Infotrieve]
-
Katus HA, Looser S, Hallermayer K, Remppis A, Scheffold T, Borgya A, et al. Development and in vitro characterization of a new immunoassay of cardiac troponin T. Clin Chem 1992;38:386-393.
[Abstract/Free Full Text]
-
Gillum RF, Fortman SP, Prineas RJ, Kottke TE. International diagnostic criteria for acute myocardial infarction and stroke. Am Heart J 1984;108:150-158.
[Web of Science][Medline]
[Order article via Infotrieve]
-
Ripley B. Pattern recognition and neural networks 1996:403 Cambridge University Press Cambridge, UK. .
-
Schenker B, Agarwal M. Cross-validated structure selection for neural networks. Comput Chem Eng 1996;20:175-186.
-
Forsström JJ, Irjala K, Selén G, Nyström M, Eklund P. Using data preprocessing and single layer perceptron to analyze laboratory data. Scand J Clin Lab Invest 1995;55(Suppl 222):75-81.
-
Elman JL. Finding structure in time. Cognit Sci 1990;14:179-211.
-
Lindahl B, Venge P, Wallentin L. Relation between troponin-T and the risk of subsequent cardiac events in unstable coronary artery disease. Circulation 1996;93:1651-1657.
[Abstract/Free Full Text]
-
Herlitz J, Hjalmarson A, Waldenstrom J. Time lapse from estimated onset of acute myocardial infarction to peak serum enzyme activity. Clin Cardiol 1984;7:433-440.
[Web of Science][Medline]
[Order article via Infotrieve]
-
Indications for fibrinolytic therapy in suspected acute
myocardial infarction: collaborative overview of early mortality and
major morbidity results from all randomised trials of more than 1000
patients. Fibrinolytic Therapy Trialists' (FTT) Collaborative. Lancet
1994;343:311–22..
-
Langer A, Goodman S, Topol E, Charlesworth A, Skene AM, Wilcox RG, Armstrong PW. Late assessment of thrombolytic efficacy (LATE) study: prognosis in patients with non-Q wave myocardial infarction. J Am Coll Cardiol 1996;27:1327-1332.
[Abstract]
-
Braunwald E, Cannon C. Non-Q wave and ST segment depression myocardial infarction: is there a role for thrombolytic therapy? [Editorial]. J Am Coll Cardiol 1996;27:1333-1334.
[Web of Science][Medline]
[Order article via Infotrieve]
The following articles in journals at HighWire Press have cited this article:
|
|
|
|
 
J. R. Lakowicz, J. Malicka, E. Matveeva, I. Gryczynski, and Z. Gryczynski
Plasmonic Technology: Novel Approach to Ultrasensitive Immunoassays
Clin. Chem.,
October 1, 2005;
51(10):
1914 - 1922.
[Abstract]
[Full Text]
[PDF]
|
|
|
Top
Abstract
Introduction
