Transcript Result

Electrode Selection for Noninvasive
Fetal Electrocardiogram Extraction
using Mutual Information Criteria
R. Sameni (1,3) , F. Vrins(2) , F. Parmentier (2), C. Hérail (4), V. Vigneron (4),
M. Verleysen (2), C. Jutten (1), and M. B. Shamsollahi (3)
(1) Laboratoire des Images et des Signaux (LIS) – CNRS UMR 5083, INPG, UJF, Grenoble, France
(2) Machine Learning Group (MLG), Microelectronics Laboratory, Université Catholique de Louvain
(UCL), Louvain-La-Neuve, Belgium
(3) Biomedical Signal and Image Processing Laboratory (BiSIPL), School of Electrical Engineering,
Sharif University of Technology, Tehran, Iran
(4) Laboratoire Systèmes Complexes (LSC) – CNRS FRE 2494, Evry, France
MaxEnt 2006
July 10th 2006, Paris, FRANCE
Overview




Introduction
Backgrounds
Methods & Results
Summary & Conclusions
Overview




Introduction
Backgrounds
Methods & Results
Summary & Conclusions
Introduction
Objective
 The noninvasive
extraction of fetal ECG
(fECG) from an array of
electrodes placed on
the abdomen of a
pregnant woman
Introduction
Perspective:
Array Recorded Signals
Spatial Filtering
Temporal Filtering
(Blind Source Separation)
(Dynamic Bayesian Filter)
Noninvasive Fetal ECG Extraction
Introduction
Perspective:
Array Recorded Signals
Spatial Filtering
Temporal Filtering
(Blind Source Separation)
(Dynamic Bayesian Filter)
Noninvasive Fetal ECG Extraction
Introduction
The Array Recording System
Introduction
Challenging issues in fECG
extraction








No direct access to the fetus
Weakness of the fECG
Maternal ECG, EMG, Diaphragm, and Uterus noises
Attenuation of the fECG in the maternal body
Fetal movement and rotation
Necessity of a canonical fECG representation
fECG of twins and triplings
…
Noninvasive fECG extraction is a challenging
application for the ICA community
Introduction
Why use ICA?
 By using the array recordings we
compensate the low fECG SNR by
the spatial diversity of the
electrodes
Introduction
Problems with high-dimensional
signals





Curse of dimensionality
High processing cost
Redundancy
Sensitivity to noise
Spurious components extracted by ICA
Introduction
General Perspective
 Record high-dimensional data
 Select the channels containing the most
information about the fetal heart
 Extract the fetal components using ICA
(a canonical representation of the fetal
ECG)
 Dynamically re-select the channels
according to the fetal movements
Overview




Introduction
Backgrounds
Methods & Results
Summary & Conclusions
Backgrounds
The electrical activity of the heart
 The contraction of the heart muscle
is due to the periodic stimulation of
the cardiac nervous system.
Backgrounds
The electrical activity of the heart
 Single dipole model: A rotating time-variant vector
located at the heart.
 Other Models: Moving dipole, Multipole, Activation
maps, …
Backgrounds
What is the ECG?
 The Electrocardiogram (ECG) is the
overall electrical activity of the heart
recorded from the body surface
Backgrounds
What is the Vectorcardiogram?
 The Vectorcardiogram (VCG) is a
3D representation of 3 orthogonal
ECG leads
Backgrounds
A dynamic model for the generation of
synthetic maternal abdominal signals
Overview




Introduction
Backgrounds
Methods & Results
Summary & Conclusions
Methods & Results
Channel selection vs. projection
 The fECG components are very weak,
and will be removed by projection
 For noisy signals, ICA can artificially
extract signals which do not correspond
to any physiological source
Methods & Results
Typical Signals Extracted by ICA
Maternal
ECG
Fetal
ECG
Noise
Systematic
Noise
Methods & Results
Which measure of selection?
 We require a measure for the selection of
the most- and least- informative leads.
 As we use the channel selection as a
preprocessing for ICA the Mutual
Information (MI) between each lead and
the maternal and fetal components is a
reasonable candidate.
Methods & Results
MI results on simulated data
Methods & Results
Mutual Information (MI)
X and Y can be either scalars or vectors
F and G are Invertible Transformations
Methods & Results
Mutual Information for ECG and
VCG signals
Result: The MI calculated between any body surface
recording and the VCG signals is ‘rather’ robust to the
locations of the VCG electrodes
Methods & Results
Previous sensor selection strategy
 Rejection of the channels with the
most MI with the maternal ECG:
Maternal reference
I ( X , mECGref )
Methods & Results
Typical ECG recordings
Methods & Results
New Channel Selection Strategy
 A three step selection with multiple
reference channels:
1. Classification of the electrodes according to
their correlation with the maternal ECG
2. Rejecting the channels with the most MI with
the maternal ECG
3. Among the remaining channels, keeping the
ones with the most MI with the fetal ECG
Methods & Results
1. Classification of electrodes based on
the maternal contribution:
Methods & Results
2-1. Ranking of electrodes based on
the maternal contribution:
(Rule #1)
Methods & Results
2-2. Ranking of electrodes based on
the maternal contribution:
(Rule #2)
Methods & Results
3. Ranking of electrodes based on the
fetal contribution:
Methods & Results
Typical fECG signals extracted from by
using the electrode selection rules
fECG extracted from the whole data set
fECG extracted from 20 selected leads
fECG extracted from 10 selected leads
Overview




Introduction
Backgrounds
Methods & Results
Summary & Conclusions
Summary & Conclusions
Summary & Conclusions:
 We proposed a channel selection
algorithm for the selection of the most
informative sensors corresponding to the
fetal ECG signals
 By using the MI with appropriate models
for the heart signals we can effectively
reduce the number of channels with
minimal loss of information
Thanks for your
attention!