Analysis of ECG using Wavelets

1
Analysis of ECG using Wavelets

• By Aditya Saurabh
• 2001001

2
What to diagnose? Ischemia!

• WHO prediction By 2020, 75 of deaths due to
  non-communicable diseases and Coronary Heart
  Disease (CHD) will top the list.
• Physionet Challenge of 1000 and IEEE paper for
  the best classifier.
• Indo-Russian project for IIITA Development of
  new methods in search for spatiotemporal
  descriptive features in the Cardiac signals and
  images. To be specific, in this project we will
  target to find new spatiotemporal descriptive
  feature mining non-invasive biomedical
  signals/images to relate Ischemic Heart Diseases
  caused by Coronary Artery Blockade.
• Myocardial Ischemia precedes lethal arrhythmia

3
The normal electrocardiogram
So much information in our heart beats!! I
thought it only lub-dubbed
4
Creation of Ischemic template

• Myocardial ischemia- a shortage of blood flow and
  oxygen that is not quite severe enough to
  permanently damage the muscle. Ischemia is the
  precursor of CHD.
• Effect on ECG ST elevation/depression, T
  inversion, Q prominence.
• Three stages of coronary artery blockage and thus
  increasing myocardial damage
• Ischemia with inversion of T wave
• Injury with shift of ST segment
• Infarction or muscle death, recognized by the
  appearance of Q wave and decrease in amplitude or
  disappearance of R wave.

5
Other findings

• increase in the QRS amplitude
• subtle prolongation of QRS duration
• QRS axis shifts
• and/or T-wave morphology changes
• prominent mid-QRS peaks in the frequency range of
  40100 Hz

6
Template example
7
Why Wavelets?

• The complements of nature?
• Soft Computing solve NP complete problems by
  giving up precision decrease dimensions,
  decrease complexity.
• Wavelets where to give up precision.

Wavelets facilitate to look at information from
different distances
8
The process
ECG
Diagnosis
Preprocessing
Classification
Pattern Recognition
Reduce the data yet Increase information
Create Feature Vector
9
A Lead V4 ECG
Age 65 Sex M Mixed angina Inferior
myocardial infarction 3-vessel
disease Medications verapamil Recorder type
ICR 7200
There are 12 leads for our body of which 4 are
redundant!
10
Looking at different details (DWT)
Data reducing, yet information increasing
11
Much more details CWT
Mexican Hat
db4
12
Preprocessing

• Collection of Data
• Download and understand the long term ST database
  from physionet.org
• Preparation of Data
• Parse and Clip the ST Episodes
• Create the mapping structure
• Detection of QRS Complex
• Understand the signature of R peaks on wavelet
  basis
• Apply DWT to segment ECG into R-R intervals

13
Data Preparation
14
Appearance of zero crossings
15
Detection of QRS Complex

• Detection of R peak
• A robust method for R peak detection is proposed
  using the Detail Coefficients of level 1 DWT of
  the signal using haar wavelet.
• Steps
• Get the detail coefficients of level 1 dwt with
  haar wavelet, say cd1
• Normalize cd1 by dividing with its maximum value
• Find the zero crossings of normalized cd1 using a
  filter of -3 0 3
• Threshold the filtered values with tau2
• Find the windowed maximum of the generated
  signal, join the two max if they exist in a
  window of 25 samples. Say the max index is i.
• 2(i-1) gives the location of R peak in the
  original signal, since we were processing the
  sub-sampled dwt.
• Q inflection point
• S inflection point

16
R peak detection example
17
An erroneous signal
18
Failure of the detection algorithm
19
Ischemic Observation
20
Non-Ischemic Observation
21
Feature extraction using CWT and ischemia
detection

• Steps
• CWT using the Mexican hat wavelet is calculated
  for the signal.
• The just previous inflection point before the
  detected R peak gives the Q valley, just next S.
• Mean Mi of the CWT coefficients at scales of 8-12
  for the QRS complexes are calculated. The mean of
  these Mis, Mbar and their standard deviation is
  taken as two features.
• The two features are considered for
  classification using ANN and Clustering.

22
Results
23
Scope for Future Work

• The following extensions to my work should be
  fruitful
• The ischemic template could be used to construct
  the problem dependent wavelet and applied to the
  electrocardiogram signal to detect ischemia.
• Continuous Wavelet Transform can be used to
  completely segment the ECG as well using phase
  and absolute values.
• Many other feature extraction techniques could be
  implemented for disease detection.
• A supervised classification using modified neural
  networks could be explored by exploiting my
  feature extraction process.

24
References

• 1) http//www.physionet.org
• Surehka Palreddy, Ph. D. University of Wisconsin,
  ECG BEATS DATABASE Description, 1996.
• http//www.tararokpa.org/diagnose.htmlpulse.
• W. Sweldens, The lifting scheme A custom-design
  construction of biorthogonal wavelets," Appl.
  Comput. Harmon. Anal., vol. 3(2), pp. 186-200,
  1996.
• W. Sweldens and P. Schroder, Building your own
  wavelets at home," Tech. Rep. 19955, Industrial
  Mathematics Initiative, Department of
  Mathematics, University of South Carolina, 1995,
  (ftp//ftp.math.sc.edu/pub/imi 95/imi95 5.ps).
• I. Daubechies, Ten Lectures on Wavelets, CBMS-NSF
  Regional Conf. Series in Appl. Math., Vol. 61,
  Society for Industrial and Applied Mathematics,
  Philadelphia, PA, 1992.
• Marc Thuillard, Wavelets in Soft Computing,
  World Scientific Publishing Co. Pte. Ltd., 2001

The first link is always the best one!
25
Thank You Any Questions?
Opposites seem to contradict but complement each
other!