ST-T segment change recognition using artificial neural networks and principal component analysis

Any ST-T segment was here represented by using the principal component analysis, or Karhunen-Loeve Transform (KLT). A representative KL basis set was built from a database containing more than 97000 normal and abnormal ST-T segments. So it was possible to concentrate the 90% of the ST-T signal energy in the first KL coefficients. For the evaluation, the ST-T European Database was chosen, because of its large amount of ischemic episodes. The baseline was removed by using a cubic spline and an adaptive filter was applied in order to improve the signal-to-noise ratio in the final KL series, delivering an improvement of about 10 dB. Then a 3-layers feedforward neural network trained with backpropagation, was applied to the KL series to recognize ST-T level changes. Each input pattern consisted of 28 features, representing 7 ST-T segments, each one described by means of its first 4 KL coefficients. 3 output units were designed, one to describe ST depression, one ST elevation, and one to represent artefacts. The use of principal component analysis and of artificial neural networks allowed us to obtain a sensitivity of 77% and a positive predictive accuracy of 86% on the test set.