Time series prediction on electrocardiogram data by radial basis function neural networks

Date

1994-12

Journal Title

Journal ISSN

Volume Title

Publisher

Abstract

This thesis presents an investigation of time series prediction on Electrocardiogram (ECG) data by Radial Basis Function (RBF) neural networks. Both the healthy and unhealthy ECG data are investigated for single-point and multi-point prediction. Using RBF to process signals is actually a curve fitting __ p_~9~ed.ure in a mqlti_~ . - . ·- - ~ ~ ... --- -- ~- -. - -. -~ dimensional spac~ RBF can be trained very fast, but it has the problem in selecting a

  • -· -- ---- ------· ...____ __ -- ------- proper set of basic centers for good generali~ation. The Orthogonal Le~§t - ~-q!l_are ---- • ----- • • • • • •. - + (OLS) algorithm provides a systemati~ way to_ select _th~ RB;F -~_enters. Based on the
  • · - - -----~ ------ ~--- --- -- ----· - - ~ ---- . --- - · step by step studies on sine-wave data, Mackey-Glass (MG) equation data, German

unemployment data and ECG data, the OLS is evaluated as an RBF center selector. ~------------------ - ~ - - · - The prediction results show that OLS ... is a_ us~ful tool for RBF center selecti<;>!'J.. The --------- - "": . ---- ... -------·- . - repeat of OLS improves the center selection and forms an improved basis for the space. The sine-wave data is a regular time series with a strict cycle. This data set shows accurate prediction results using 2 as the input dimension. The prediction extends accurately far into the future. As a well-behaved chaotic time series, the prediction of MG data is also very good with input dimension as 17. The German unemployment data is a random time series from the real world process. Its single-point prediction performs well with input dimension as 13. Its multi-point prediction performs well in a short term, but cannot fit the real data curve farther away. The prediction error increases with ------ - __. __.. - - -t. he time st-· -e- p. . incre~~-~-

Description

Rights

Availability

Unrestricted.

Keywords

Electrocardiography -- Data processing, Neural networks (Computer science), Time-series analysis

Citation