Comparison of methods for automatic identification of ECG T-wave alternans

Microvolt T-wave alternans (TWA), consisting of every-other-beat changes in ECG T-wave morphology, is an index of susceptibility to malignant ventricular arrhythmias, requiring automatic techniques to be identified. Among these, the fast-Fourier-transform spectral method (FFTSM), the complex demodulation method (CDM), the modified-moving-average method (MMAM), the Laplacian-likelihood-ratio method (LLRM), the correlation method (CM), the enhanced-modified-moving-average method (EMMAM) and the adaptive-match-filter method (AMFM) were applied in this thesis to simulated and sample clinical data. The aim of the thesis was to compare individual methods ability to properly identify stationary and time-varying TWA, even in the presence of corrupting factors, such as noise, baseline wanderings, respiration modulation, T-wave misalignment and false R-peak detections. In the specific we first focused on the CM and EMMAM reliability to identify TWA. Only these two techniques, indeed, incorporate T-wave alignment procedures in their algorithm. More precisely, the CM accomplishes the T-wave alignment by means of a cross-correlation technique, while the EMMAM involves the dynamic-time-warping alignment procedure. This comparative analysis showed that alignment procedure including in the CM is more reliable than the one incorporating in the EMMAM. Consequently, for the analysis of other TWA-identification techniques, it is reasonable to consider a preprocessing stage including the cross-correlation procedure for T-wave alignment. Further comparative analysis involved the FFTSM, CDM, MMAM, LLRM and AMFM. The CM and the EMMAM were not considered any longer: the former because overcome by the AMFM, as found in literature, the latter because equivalent to the MMAM with exception of T-wave alignment procedure. The simulated study highlighted that the FFTSM did not provide false-positive results but, in the presence of time-varying TWA, it showed limitations in recognizing non-stationary features of TWA phenomena. The CDM was able to detect transient TWA, limited in rate variation, though; moreover, it showed limitations in the presence of simulated T-wave misalignment, both in the absence and in the presence of TWA. The MMAM provided false-positive TWA when applied to simulated ECGs affected by amplitude variability (possibly caused by interferences corrupting the tracing), but TWA. Stationary TWA was properly quantified by the MMAM and, occasionally, underestimated by all other methods. The LLRM showed limitations in the following cases: quick sinusoidal TWA, presence of high-frequency baseline in the tracing, and T-wave misalignment. The AMFM properly identified time-varying TWA and, in the presence of corrupting factors, did not provide any false-positive TWA. Altogether, the AMFM accomplished the best compromise between the needs to avoid false-positive TWA and to detect and characterize true-positive TWA in different conditions. Results of our simulation approach were useful to explain different levels of TWA measured by the methods applied to real situations constituted by sample Holter ECGs from healthy subjects and acute-myocardial-infarction patients.