Hybrid detector for the T wave alternans challenge

In the work presented here we propose an extended approach for the TWA modeling and detection based on a previous method presented in CinC07. To this aim, the ECG amplitude modulation and the baseline wander in the ST-T segment are approximated by a scaling factor applied to the T wave added to a constant. Thanks to this simplification, the proposed global model of each ST-T segment is constituted by the scaled T wave, an offset and the shape of the alternans. In addition to the modeling we also introduce a method to address the problem of the estimation of the T wave, the offset and the alternans shape and their respective coefficients. Thus, a model-based detector can be derived such as the generalized likelihood ratio test. Through this detection scheme, the orthogonality of the models is addressed and solved by imposing some constraints. The application of these constraints is driven by a student-t test output applied to the estimated parameters but also on the energy of the reconstructed T waves. An alternative detector based on a principal component analysis is also proposed. Here again the presence of a baseline residual is accounted in the analysis. Since the last stage of the latter detector utilizes a student-t the selection of an optimal threshold is avoided. These detectors and alternans wave estimators are applied to the data set proposed by the CinC Challenge. It is shown that when applied on even short duration alternans it produces positive results. It is also shown that under the scope of the challenge the T waves segmentation plays an important role.