Selection of parameters from power spectral density, wavelet transforms and other methods for the automated interpretation of the ECG

We investigate a range of alternative classifiers for automated interpretation of Frank lead ECG's derived from (i) power spectral density (PSD) estimates of a continuous (8-10 s) recording, (ii) PSD estimates derived from single QRS complexes centred in a data window of 250 ms, (iii) RMS value of signals reconstructed from a 6 level discrete wavelet transform (DWT) decomposition of the ECG using the Haar wavelet, and (iv) scalar parameters calculated from the QRS complex in a selected beat. A database of 500 fully diagnosed ECG records was used to develop the ROC curves for all possible combinations of disease conditions. Results demonstrated that PSD parameters derived from the X lead were generally the most sensitive discriminators between infarcts and ventricular hypertrophies, and were generally more sensitive than parameters derived from DWT decomposition. Surprisingly, neither PSD or DWT parameters provided as good discrimination between normal and abnormal records as did simple scalar lead parameters. A feedforward neural net trained with backpropagation to discriminate between normal and abnormal records was provided with the ten most discriminating parameters in each of the four parameter sets as input. Best results (84.4% correct classification) were obtained for the scalar lead parameters. The worst result (76.6%) was obtained for the continuous PSD. The parameters selected were in general highly discriminating between individual disease conditions and were not optimal for the classification of normal/abnormal records.