Reference: A Diery, D Rowlands, TRH Cutmore, D James, Automated ECG diagnostic Pwave Analysis using Wavelets, Computer methods and programs in biomedicine 101 (1), 3343 Abstract: Pwave characteristics in the human ECG are an important source of information in the diagnosis of atrial conduction pathology. However, diagnosis by visual inspection is a difficult task since the Pwave is relatively small and noise masking is often present. This paper introduces novel wavelet characteristics derived from the continuous wavelet transform (CWT) which are shown to be potentially effective discriminators in an automated diagnostic process. Characteristics of the 12lead ECG Pwave were derived using CWT and statistical methods. A normal control group and an abnormal (atrial conduction pathology) group were compared. The wavelet characteristics captured frequency, magnitude and variance components of the Pwave. The best individual characteristics (i.e. ones that significantly discriminated the groups) were entered into a linear discriminant analysis (LDA) for four different models: twolead ECG, threelead ECG, a derived threelead ECG and a factor analysis solution consisting of wavelet characteristic loadings on the factors. A comparison was also made between wavelet characteristics derived form individual Pwaves verses wavelet characteristics derived from a signalaveraged Pwave for each participant. These wavelet models were also compared to standard cardiological measures of duration, terminal force and duration divided by the PR segment. Results for the individual Pwave approach generally outperformed the standard cardiological measures and the signalaveraged Pwave approach. The best wavelet model on the basis of both classification performance and simplicity was the twolead model that uses leads II and V1. It was concluded that the wavelet approach of automating classification is worth pursuing with larger samples to validate and extend the present study. Discussion: This study set out to determine whether characterisation of the Pwave with wavelet decomposition would result in useful classification of abnormality. The results clearly indicate that this has been achieved, particularly with the twolead and threelead subsets of the clinical 12lead ECG configuration. In addition, the twolead wavelet characterization showed a 10% improvement in classification rate over the typical cardiological parameters. The results achieved using the signalaveraged Pwaves were in general poorer than the individual Pwave characterization approach. This might be attributed to the signal average being more susceptible to variance in the Pwave morphology that “smears” during the averaging process losing some informative features. Fig. 7provides some support for this hypothesis in that patient individual Pwave scores do indeed show considerable variation. Looking at the waveletbased two and threelead models in greater detail, it can be noted that the characteristics with the best discriminant potential are the median frequency and the peak frequency energy value. The median frequency is the centre of the wavelet scale ‘energy’ distribution. This is defined in Eq. (1)and relates the magnitude of the wavelet coefficients (expressed as ‘energy’) to the wavelet scale. Fig. 5c shows that normal Pwaves have a higher frequency than the abnormal Pwaves. This reflects the more compact shape of the normal Pwave. The duration measure used by cardiologists is sensitive to this morphological difference in abnormal Pwaves as well. It is underscored by the good performance achieved by the duration measure on its own (nearly as good as the composite of all the cardiological measures). The conduction of the pulse through the atria takes longer (i.e. extended in duration) due to atrial enlargement. It should be noted, however, that there are individual differences in the orientation of the longaxis of the heart. This would introduce variance in the lead II estimate of duration. One way to examine this influence would be to include an echo cardiogram measure of longaxis orientation in tandem with the Pwave measures used here, and correlate them. Only one measure of variance appeared in the best lead models and this was the QV measure on avF and lead II (Table 1a) It is also apparent, that for the threelead model, in Fig. 5e and f, the variance measures do not result in large group differences (only avF QV spans more than two SEMs). Although two of the variance measures did load on factor 3 of the factor analysis, this method did not perform quite as well as the models that did not include the IQR measure. One wavelet method that did not perform as well as might be expected was the Frank lead XYZ method. The XYZ configuration did perform the best when the averaged Pwave was used as an input, though this performance was inferior to the single Pwave two and threelead wavelet methods and the classical ECG measures. Although the orthogonal set is similar to some extent to the avF, V2, V5 set, this is only approximate and indeed these derived leads may not have been completely orthogonal. A recorded Franklead trial may well show better performance than we have shown here. Finally, the best wavelet method predictor of all was the simplest one using only leads II and V1. Adding a third lead whether by a factor analytic method or a derived lead set does not appear to offer classification advantage and, in fact, may add noise. The twolead set used II and V1. The lead II may be particularly important as it is sensitive to Pwave duration. Table 1a shows that four of the wavelet parameters were, in fact, based on this lead with median frequency capturing the duration information. The twolead computational load for a modern PC is also quite practical, particularly once a sound discriminant equation is arrived at. The present study used a modest sample size and has demonstrated the utility in using an automated Pwave classifier, showing it to outperform the usual duration and other cardiological measures by at least 10%. It is important however to extend the confidence in this analytic approach by conducting these analyses on a much larger data set. This can be a very costly enterprise if new data is collected, however online databases with ECG recordings are slowly becoming available [34]. While the effects of age on cardiological measures taken from normal subjects does not appear to change much beyond the mid 50s [21] it would inspire greater confidence to have a larger control sample – particularly for getting a measure of specificity. Clearly these findings, with a modest sample size, warrant further study to extend the generality to other laboratory and clinical settings. Furthermore, our suggestions of the relationship of Pwave anomalies to underlying atrial structure should be followed up in a replication ECG study that also includes echo cardiogram to more directly validate, on a case by cases basis, possible structural abnormalities that correlate to waveletbased Pwave morphology. This work is now underway in our laboratory in collaboration with a cardiology clinic. To be useful to practicing cardiologists and GPs with ECG recording capabilities, the present system could be incorporated into current ECG software. This would be as a thirdparty plug in, for example to Cardioview™. A large set of cases should be compiled (future work on this is needed) to provide a robust LDAtrained and classified reference set. An interface could be easily added that allows the user to specify the patient, the database of LDAtrained cases and the form of the output, such as the distance scores for each of the diagnostics groups in the selected database as well as an interpretation of the Pwave morphology based on the wavelet parameter analysis. In conclusion, the success of this waveletbased approach may prove advantageous in the diagnosis of low amplitude and dispersed Pwaves, which are commonly observed in elderly participants [11]. This can make diagnosis by visual inspection of the ECG a difficult task. Furthermore, the results of the wavelet methods described here compare favourably to the higher order system modelling approach [16] and are significantly better than the DFT method. One advantage of the present approach is that it does not rely on any initial assumptions about the generation of the Pwave and could therefore have wide applicability. References:

Publications > 2011 >