Maša Strobl (2002) . Prešeren awards for students.

Full text not available from this repository.

Abstract

In diagnostic cardiology, there was an increasing demand for automatic assessment of presence or absence of transient myocardial ischaemia in long 24-hour electrocardiographic recordings. For this purpose, there were previously developed algorithms for searching the reference points of heart cycle like the position of the isoelectric level and the position of the J point. These two positions serve for assessment of the ST-segment level amplitude and thus presence or absence of ischaemic heart disease. Unfortunately, automatic estimation of the positions of the two reference points yield errors with regard to the positions manually set by expert cardiologists. The developed algorithms needed to be optimized. We first built an user interface which allows one to view functioning of the algorithms on single electrocardiographic recordings. Besides, it was necessary to define estimators, or performance measures, which reflect actual quality of the algorithms. The performance measures chosen were aggregate average mean error and its standard deviation for the position of the isoelectric level, the position of the J point, and for the ST-segment level amplitude. It was also necessary to study the effectiveness of the algorithms in details and to select those parameters which would influence the performance results at most. For the optimization of the algorithms, we used the LTST database containing 24-hour ECG recordings annotated by expert cardiologists. By modifying different parameters of the algorithms, we made nine iterations of working of the algorithms each time using the entire database. As the best result, we select that iteration which showed to be the most robust, i.e., the iteration which demonstrated pretty good estimation of the reference points with a little higher error in all problematic recordings, and because of this the algorithms can make also a little higher error in non-problematic recordings. In real clinical environments, it is more important robust, and more or less accurate functioning of the algorithms in all recordings, and not complete failure of the algorithms in rare problematic recordings. The results are excellent. We are the most satisfied with automatic searching for the position of the isoelectric level and a bit less with searching for the position of the J point. For the latter, we think better performance could be obtained with a few changes of the algorithm itself. It is important to emphasize also following. Functioning of the algorithms was not problematic on normal waveforms in non-pathological recordings. Problems appear in pathological recordings, like recordings containing conduction defects. In these recordings, expert cardiologists themselves are having problems in annotating in spite off many year's experience and deep medical knowledge.

Actions (login required)