Miha Amon (2011) Robust estimation of morphologic features and shape representation of electrocardiograms using orthogonal transforms. MSc thesis.
An important task in the field of electrocardiogram (ECG) signal processing is the development of effective discrete transforms, which can extract useful clinical information from source signals and represent it as morphologic feature vector time series. Such time series are then suitable for further machine processing as well as for visual diagnostic procedures by cardiologists. We have tested and enhanced the existing Karhunen and Loève transform (KLT) feature vector space based noise detection algorithm. The algorithm is robust and uses the skipped mean value as an estimator. We have generated new KLT base functions for the ST electrocardiogram segment with all 86 24-hour records of the international reference ECG database LTST DB (Long-Term ST Database) as a learning set. New covariance matrices are robust and based on the kernel-approximation method. In addition, a new transform was developed, based on the Legendre polynomials (LPT) which are visually similar to typical morphologic changes of the ST segment during myocardial ischemia. The later provides a direct insight into the transient morphology change type from the feature vector space. The noise detection algorithm and both transforms were used for generation of new ST segment feature vector time series for all LTST DB records. We have studied and compared characteristics of both transforms in terms of residual error distribution and characterized new feature vector time series behavior in the neighborhood of transient ischemic and false non-ischemic ST segment episodes. Using the new feature vector time series of both transforms and the existing LTST DB diagnostic parameter time series (heart-rate, ST segment level) we have characterized ischemic (deviation, sloping, scooping) and non-ischemic (shift of T-wave into the ST segment) ST segment morphology changes at the level of single heart beats as well as at the level of transient episodes (episode beginnings and extremes) with the motivation of evaluating the LPT transform as a new approach to effective ischemic and non-ischemic physiologic process differentiation. New graphical tools for ECG data visualizations and morphologic feature vector time series characterization, algorithms and software for base functions generation, KLT and LPT transform feature vector time series derivation and several other functions as the residual errors, statistical analyses, time series manipulations etc. were also developed for this work. New feature vector time series and the associated residual errors were added to the LTST DB database which is freely available on the Physionet servers.
Actions (login required)