Home SVCC                                                  Area: English - Español - Português

Risk Stratification for Myocardial
Infarction Patients Using
Neural Networks

Sepúlveda, Jose; Soria, Emilio;
Sanz, Gines; Marrugat, Jaime;
Gomez, Luis
for the RESCATE investigators

Universidad de Valencia, Departamento de Ingeniería Electrónica,
Burjassot, Valencia, España

SUMMARY
Introduction: The identification of factors that can be used to predict prognosis has been a challenge since the late 1960s. Careful risk assessment for each patient aids clinicians in assessing prognosis and may therefore be a useful guide in management, providing valuable information.
Objectives: The use of classification methods to predict prognosis in patients with acute myocardial infarction, allows for better management and may further decrease in-hospital mortality. The efficacy of early reperfusion therapy, either thrombolysis or primary angioplasty, depends on the patient risk and the delay in applying the treatment; therefore, early assessment of risk is mandatory. The purpose of the present study was to develop a clinical score for risk assessment to determine the profile of every patient with acute myocardial infarction (MI) using Neural Networks.
Methods: A cohort of 1,318 consecutive patients with a first MI admitted to four referral teaching hospitals (one with tertiary facilities ) were followed up for 6 months after admission. Only patients initially admitted to each of the study hospitals were retained for analysis in the original hospital's cohort. To classify patients an Artificial Neural Network (ANN), called Multilayer Perceptron was used with the backpropagation as training algorithm. This method is used to analyze a collection of simple clinical markers.
Results: The subjects of the study have been divided in two groups 2/3 of the population for the training and 1/3 for validation. ANNs have achieved high values of sensibility and specificity in the classification of the patients, training (82.85% sensitivity, 82.14% specificity), validation (83.33% sensitivity, 80.31% specificity). During the neural network training, several modifications of the learning algorithm were used to improve results.
Conclusion: We can conclude that ANN is reliable and precise tools for solving this problem.

Top

INTRODUCTION
   Despite the advances in the treatment of acute myocardial infarction (MI) there is high in-hospital mortality. The introduction into clinical practice of effective treatments, such as thrombolysis, aspirin, ß-blockers, and angiotensin-converting enzyme ACE inhibitors, has changed the prognosis of the disease. More aggressive interventions, such as direct percutaneous transluminal coronary angioplasty (PTCA) might, for selected patients, further decrease in-hospital mortality. Practitioners have a wide variety of reperfusion strategies, to interrupt the evolving myocardial event, but the efficacy of therapeutic intervention in acute myocardial infarction is strongly time dependent. To allocate every patient to the most beneficial treatment, the risk profile of every single patient should be available immediately when a patient enters the medical care system. Careful risk assessment for each patient aids clinicians in assessing prognosis and may therefore be a useful guide in management, providing valuable information.

OBJECTIVES
   The use of classification methods to predict prognosis in patients with acute myocardial infarction, allows for better management and may further decrease in-hospital mortality. The efficacy of early reperfusion therapy, either thrombolysis or primary angioplasty, depends on the patient risk and the delay in applying the treatment; therefore, early assessment of risk is mandatory. The purpose of the present study was to develop a clinical score for risk assessment to determine the profile of every patient with acute myocardial infarction (MI) using Neural Networks.

METHODS
   Study design. The study was designed as a 6-month follow-up study of patients admitted to one hospital with and three without angiography or coronary surgery facilities. All four participating hospitals were public teaching institutions. Only patients initially admitted to each of the study hospitals were retained for analysis in the original hospital's cohort.

   Inclusion criteria. Between May 1992 and June 1994, all patients with a first MI up to the age of 80 years admitted to the four participating hospitals within 72 h of onset of symptoms of MI were included. MI was diagnosed when two of the following criteria were present: 1) abnormal Q waves, 2) increase in cardiac enzyme levels (more than twice the upper normal value), and 3) typical chest pain >20 min in duration.

   Exclusion criteria. Residence outside the study areas or any of the following condition: 1) life-threatening diseases other than the index event; 2) previous CABG or PTCA; 3) or coronary angiography in the past 6 months. Patients enrolled in ongoing clinical trials were not excluded to reproduce actual care scenarios more faithfully.

    To classify patients an Artificial Neural Network (ANN), called Multilayer Perceptron was used with the backpropagation as training algorithm. This method is used to analyze a collection of simple clinical markers. The neural model uses several units, called neurons, whose structure is shown in Fig. 1.

   A neuron has the following elements
· Synaptic weights , which vary with the learning algorithm.
· Adder, all the inputs multiplied by their weights are added.
· Activation function, a non-linear function. Several activation functions can be used, however, in the current study we have used the hyperbolic tangent.

   Multilayer Perceptron uses several neurons in a multilayer structure. In this structure, the output of the neurons in one layer is the input of the neurons in the following layer; no direct connection between neurons in non consecutive layers and no feedback has been considered. Backpropagation was the learning algorithm used to train the neural network; this learning algorithm uses descent-gradient methods. This technique has to minimize a monotonic increasing function of the network error. In order to get good generalization levels, the training algorithm will stop using the cross-validation criteria [6].

RESULTS
   During the training of the Multilayer Perceptron initialisation of the weights, the risk of falling in a local minimum and learning rate selection were some of the problems to face. The cross-validation method together with variation in the number of hidden neurons (HN), the weight initialisation range and the learning rate were used to determine the best topology

   The subjects of the study have been divided in two groups 2/3 of the population for the training with whom the model was created and 1/3 for validation or test. ANNs have achieved high values of sensibility and specificity in the classification of the patients, training (82.85% sensitivity, 82.14% specificity), validation (83.33% sensitivity, 80.31% specificity). During the neural network training, several modifications of the learning algorithm were used to improve results.

CONCLUSION
   The use of simple clinical markers readily available at admission of patients with myocardial infarction allows artificial neural networks to give a reliable prediction of risk for in-hospital and 6-month mortality. We can conclude that ANN are inexpensive, quick and precise tools for assessment of risk for in-hospital mortality.

REFERENCES

1. C.M. BISHOP, "Neural Networks for Pattern Recognition." Oxford University Press,UK, 1995.

2. C. Fresco, F. Caricini, A.P. Maggioni, A. Ciampi, A. Nicolucci, E. Santoro, L. Tavazzi, G. Tognonia. Very early assesment of risk for in-hospital death among 11,483 patients with acute myocardial infaction. (Am Heart J 1999;138:1058-64)

3. J. Marrugat, G. Sanz, R. Masiá, V. Valle, L. Molina, M. Cardona, J. Sala, L. Serés, L. Szescielinski, X. Albert, J. Lupón, J. Alonso. Six-Month Outcome in Patients With Myocardial Infarction Initially Admitted to Tertiary and Nontertiary Hospitals(J Am Coll Cardiol 1997;30:1187-92)

 

Top

Your questions, contributions and commentaries will be answered
by the authors in the Prevention and Epidemiology list.
Please fill in the form (in Spanish, Portuguese or English) and press the "Send" button.

Question,
contribution
or commentary
:
Name and Surname:
Country:
E-Mail address:

Top


2nd Virtual Congress of Cardiology

Dr. Florencio Garófalo
Steering Committee
President
Dr. Raúl Bretal
Scientific Committee
President
Dr. Armando Pacher
Technical Committee - CETIFAC
President
fgaro@fac.org.ar
fgaro@satlink.com
rbretal@fac.org.ar
rbretal@netverk.com.ar
apacher@fac.org.ar
apacher@satlink.com

Copyright© 1999-2001 Argentine Federation of Cardiology
All rights reserved

 

This company contributed to the Congress: