Performance evaluation of diverse T-wave alternans estimators under variety of noise characterizations and alternans distributions

Prognostic significance of microvolt T-wave alternans (TWA) has been established since their inclusion among important risk stratifiers for sudden cardiac death. Signal processing schemes employed for TWA estimation have their peculiar theoretical assumptions and reported statistics. An unbiased comparison of all these techniques is still a challenge. Choosing three classical schemes, this study aims to achieve holistic performance evaluation of diverse TWA estimators from a three dimensional standpoint, i.e., estimation statistics, alternan distribution and ECG signal quality. Three performance indices called average deviation (? _L ), moment of deviation (? _m ) and coefficient of deviation ( $\varphi$ ) are devised to quantify estimator performance and consistency. Both synthetic and real physiological noises, as well as variety of temporal distributions of alternan waveforms are simulated to evaluate estimators’ responses. Results show that modification of original estimation statistics, consideration of relevant noise models and a priori knowledge of alternan distribution is necessary for an unbiased performance comparison. Spectral method proves to be the most accurate for stationary TWA, even at SNRs as low as 5 dB. Correlation method’s strength lies in accurately detecting temporal origins of multiple alternan episodes within a single analysis window. Modified moving average method gives best estimation at lower noise levels (SNR >25 dB) for non-stationary TWA. Estimation of both MMAM and CM is adversely effected by even small baseline drifts due to respiration, although CM gives considerably higher deviation levels than MMAM. Performance of SM is only effected when fundamental frequency of baseline drift due to respiration falls within the estimation band around 0.5 cpb.