Noninvasive Cardiac Signal Analysis Using Data Decomposition Techniques - Modeling in Computational Biology and Biomedicine - page 93

The actual detection of the alternans effect can be carried out in different
manners. In any case, a sliding window of fixed length is applied over the output
of the PCA-based detector, i.e., the data displayed in Fig. 3.7 a. For each position of
the window, two ensembles are formed by splitting the odd and even samples. Their
means are compared assuming normal distribution laws with unequal variances.
The statistical significance ( p value) of their difference is verified by means of a
We l c h 's t test. For a window length of 16 beats, results are shown in Fig. 3.7 bwhere
the values 0, 1 and − 1 refer, respectively, to no detection, detection of sequence

“ + + ” (phase change). After

a short delay, the angioplasty procedure causes the alternans of the ventricular
repolarization, as clearly detected by the PCA-based method. Interestingly, phase
changes of the alternans sequence are present in this example. An alternative
approach applies PCA over the detection window instead of on the entire set of
Twaves.

− + − + − ” and detection of sequence “ − + − + −

3.3 Atrial Activity Extraction via Independent

Component Analysis

As presented in Sect. 3.1.2.2 , the problem of noninvasive atrial activity extraction
aims at suppressing the ventricular interference that masks the atrial signal observed
in the ECG during AF episodes. Classical methods for solving this problem compute
an average ventricular beat and subtract it from the recordings after suitable time
alignment and amplitude scaling [ 17 , 22 , 27 ]. This average beat subtraction tech-
nique relies on the quasi-repetitiveness of the QRST complex and its uncorrelation
with the atrial signal. The spatio-temporal QRST cancellation (STC) method of [ 17 ]
(also described in [ 27 ]) belongs to this family. Using a signal model reminiscent
of Eq. ( 3.12 ), each segmented QT interval is modeled as a linear combination of
the average ventricular beats of every lead (or their dominant subspace computed
via PCA) plus an additive noise term including atrial activity. An optimal linear
combination of the average beats is then subtracted from each QT interval and each
lead in the recording, canceling out the ventricular activity. Average beat subtraction
requires a previous beat detection and classification stage, and thus its performance
severely degrades in the presence of ectopic beats. By construction, this approach is
unable to cancel interference other than the ventricular activity itself. Other classical
techniques for atrial signal extraction are summarized in [ 25 ].

3.3.1 Linear Mixture Model

A recent alternative approach to atrial activity enhancement models the ECG
lead signals,

} i =1 , as a linear superposition of contributions from signal

{x i ( t )