Antonio J. Serrano López⁽¹⁾, Gustavo Camps i Valls⁽²⁾, EMILIO SORIA OLIVAS⁽³⁾, NICOLÁS VICTOR JIMÉNEZ TORRÉS⁽⁴⁾, José David Martín Guerrero^(5)
(1)Dpto. Electrónica. Universidad de Valencia - Burjasot. Spain
⁽²⁾Universitat de València - Burjassot, Valencia. Spain
⁽³⁾DPTO INGENIERÍA ELECTRONICA. FACULTAD DE FISICAS - BURJASSOT/VALENCIA. Spain
⁽⁴⁾DPTO FARMACIA Y TECNOLOGÍA FARMACEÚTICA. FACULTAD DE FARMACIA - BURJASSOT/VALENCIA. Spain
⁽⁵⁾G.P.D.S. Departament d´Enginyeria Electrònica. Universitat de València - Burjassot. Spain

			[Health Informatics]
	[New Technology]	[Cardiolovascular Diseases]

INTRODUCTION

This work proposes the use of tools for helping medical decisions on the intoxication prediction for digoxin, decisions that allows to customize the doses, bringing the benefits of this drug in cardiologic problems and avoiding its intoxication risk.

METHODOLOGY

The developed tool is an artificial neural network (ANN) that is trained with a part of the available patterns, so the network extracts the information that allows itself to predict if there will or will be not risk of intoxication. Later on, some patterns that have not been seen by the network, are chosen to check the goodness of the carried out learning.

The ANN used is a Multilayer Perceptron (MLP), that is a network formed by one input layer, at least a hidden layer and one output layer, as can be seen in Fig. 1. Each layer is formed by a set of neurons. The input and output layers have the number of neurons determined by the problem inputs and outputs and the number of neurons in the hidden layer is chosen by the network designer. Fig.1

The aim of this network is to carry out a transformation of the input variables to the output space, so the MLP has associated a mathematical function that allows to estimate the output for a certain input.

In our problem, the input variables are some physiological parameters of the patient and the output is the intoxication risk, so if the function is the best, it will be able to predict if intoxication risk exists or not.

The network is trained with the Backpropagation learning algorithm, based on the minization of a monotonic increasing function of the error made by the network on each iteration. This function is denominated cost function and in this problem, quadratic error function has been chosen.

RESULTS

The training group is formed by 172 patients (39 of these with intoxication risk) and the one of generalization is formed by 85 patients (15 of these with intoxication risk).

The best obtained result has been obtained with a neural network with 7 hidden neurons:

Training: Especifity: 92 % / Sensibility: 100 %

Generalization: Especifity: 83 % / Sensibility: 87 %

CONCLUSIONS

The intoxication risk is predicted with high level of trust, obtaining much better results that those obtained with classical statistical tools (Logistic Regression). Neural networks constitutes an effective tool in the aid on medical decission due to highly nonlinear relationships between independents and dependents variables. The proof of this assert is the high specifity and sensibility obtained with neural networks.

ACKNOWLEDGEMENTS

This work has been supported by FEDER project 1FD97-0935.

REFERENCES

Soria E., Serrano A.J., Herreros A., Albert A., Jiménez N.V., Guerrero J. "Aplicación de diferentes técnicas en la modelización de la probabilidad de riesgo de intoxicación por digoxina". Actas del XV Congreso Anual de la Sociedad Española de Ingeniería Biomédica, CASEIB 97, Valencia, November 1997.
Haykin S. "Neural Networks: To Comprehensive Foundation". McMillan, 1994.
Kung S.Y. "Digital Neural Networks". Prentice Hall, 1995.
Bishop C.M. "Neural Networks for Pattern Recognition". Clarendon Press, 1995.

Discussion Board

Any Comment to this presentation?

[ABSTRACT] [INTRODUCTION] [METHODOLOGY] [RESULTS] [FIGURES] [CONCLUSIONS] [ACKNOWLEDGEMENTS] [REFERENCES] [Discussion Board]

---

			[Health Informatics]
	[New Technology]	[Cardiolovascular Diseases]

Antonio J. Serrano López, Gustavo Camps i Valls, EMILIO SORIA OLIVAS, NICOLÁS VICTOR JIMÉNEZ TORRÉS, José David Martín Guerrero
Copyright © 1999-2000. All rights reserved.

Last update: 14/01/00