Transcript ECG signal delination and compression

T-61.181 Biomedical Signal Processing
ECG Signal Delineation And
Compression
Chapters 6.2.6 – 6.3
18th November
Outline
I.
ECG signal delineation



II.
ECG signal compression


III.
Definition (What)
Clinical and biophysical background (Why)
Delineation as a signal processing (How)
General approach to data compression
ECG signal compression
(Intrabeat/Interbeat/Interlead)
Summary
Part I.
EGC signal delineation
Delineation - Overview
• Aim – Automatically decide/find onsets and
offsets for every wave (P, QRS, and T) from ECG
signal (PQRST-complex)
• Note! Experts (Cardiologist) use manual/visual
approach
Why?
• Why – Clinically relevant parameters such
as time intervals between waves, duration of
each wave or composite wave forms, peak
amplitudes etc. can be derived
• To understand this look how ECG signal is
generated
ECG Signal Generation
What Are We Measuring?
• ECG gives (clinical) information from
generation and propagation of electric
signals in the heart.
• Abnormalities related to generation
(arrhythmia) and propagation (ischemia,
infarct etc.) can be seen in ECG-signal
• Also localization of abnormality is possible
(12 lead systems and BSM)
Clinically Relevant Parameters
• QRS duration
Bundle brand block
depolarization
• ST segment
ischemia
• QT interval
ventricular
fibrillation
• PR interval
SA  ventricles
Signal Processing Approach to
Delineation (How)
• Clinical importance should now be clear
• Delineation can also be done manually by
experts (cardiologist)  expensive and time
consuming. We want to do delineation
automatically (signal processing)
• No analytical solution  performance has
to be evaluated with annotated databases
Building Onset/Offset Detector
 Many algorithms simulate cardiologist
manual delineation (ground truth) process:
 Experts look 1) where the slope reduce to
flat line 2) respect maximum upward,
downward slope
 Simulate this: define the boundary
according to relative slope reduction with
respect maximum slope  LPD approach
Low-Pass Differentiated (LPD)
• Signal is 1) low-pass filtered i.e. high
frequency noise is removed (attenuated) and
2) differentiated dv/dt
• New signal is proportional to slope
• Operations can be done using only one FIR
filter :
y(n) x(n) * h(n)
LPD cont.
• Each wave has a unique frequency band
thus different low-pass (LP) filtering
(impulse) responses are needed for each
wave (P, QRS, and T)
• Design cut-off frequencies using Power
Spectral Density (PSD)
• Differentiation amplifies (high freq.) noise
and thus LP filtering is required
LPD cont..
• Waves w={P,QRS,T} are segmented from
the i:th heart beat.
 y ( n)

ywi  
0


n  ˆi  W0 ,...,ˆi  We
, oteherwise
• Using initial and final extreme points
thresholds for can be derived
 o wi  y o wi / K o w
 e wi  y e wi / K e w
LPD cont...
• Constants are control the boundary detection they
can be learnt from annotated database
• Search backwards from initial extreme point.
When threshold is crossed  onset has been
detected
• Search forward from last extreme point and when
threshold is crossed  offset is detected.
Part II.
EGC signal compression
General Data Compression
• The idea is represent the signal/information
with fewer bits
• Any signal that contains some redundancy
can be compressed
• Types of compression: lossless and lossy
compression
• In lossy compression preserve those
features which carry (clinical) information
ECG Data Compression
1) Amount of data is increasing: databases,
number of ECG leads, sampling rate,
amplitude resolution etc.
2) ECG signal transmission
3) Telemetry
ECG Data Compression
• Redundancy in ECG data: 1) Intrabeat 2)
Interbeat, and 3) Interlead
• Sampling rate, number of bits, signal
bandwidth, noise level and number of leads
influence the outcome of compression
• Waveforms are clinically important
(preserve them) whereas isoelectric
segments are not (so) relevant
Intrabeat Lossless Compression
• Not efficient – has mainly historical value
• Sample is predicted as a linear combination
of past samples and only prediction error is
stored (smaller magnitude):
xˆ p (n)  a1 x(n  1)  ...  a p x(n  p)
e p  x(n)  xˆ p (n)
Intrabeat Lossy Compression
Direct Method
• Basic idea: Subsample the signal using
parse sampling for flat segments and dense
sampling for waves:
(n,x(n)), n=0,...,N-1  (nk,x(nk)), k=0,...,K-1
Example AZTEC
• Last sampled time point is in n0
• Increment time (n) As long as signal in
within certain amplitude limits (flat)
xmin (n)  min{x(n0 ), x(n0  1),, x(n)}
xmax (n)  max{x(n0 ), x(n0  1),, x(n)}
xmax (n)  xmin (n)  
1
y (nk )  ( xmin (nk )  xmax (nk ))
2
Intrabeat Lossy Compression
Transform Based Methods
• Signal is represented as an expansion of
N
basis functions:
x   wk k
k 1
• Only coefficients need to be restored
• Requirement: Partition of signal is needed
(QRS-detectors)
• Method provides noise reduction
Interbeat Lossy Compression
• Heart beats are almost identical (requires
QRS detection, fiducial point)
• Subtract average beat and code residuals
(linear prediction or transform)
xi (n)  X (n  ˆi )
n  0,..., N  1
1 L
sˆi (n)   x(n ˆi  j )
L j 1
yi  xi (n)  sˆi (n)
n  0,..., N  1
Interlead Compression
• Multilead (e.g. 12-lead) systems measure
same event from different angles 
redundancy
• Extend direct and transform based method
to multilead environment
– Extended AZTEC
– Transform concatenated signals
 x1 
x 
2

x

 
 x12 
Summary - part I
• Delineation = automatically detect waves
and their on- and offsets (What)
• Clinically important parameters are
obtained (Why)
• Design algorithm that looks relative slope
reduction (How)
• LPD-method – Differentiate low-pass
filtered signal
Summary - part II
• Compression = remove redundancy:
intrabeat, interbeat, and interlead
• Why – Large amount of data, transmission
and telemetry
• Lossless (historical) and lossy compression
• Notice which features are lost (isoelectric
segments don’t carry any clinical
information)
Summary - part II cont.
• Intrabeat 1) direct and 2) transform based
methods
– 1) Subsample signal with non-uniform way
– 2) Use basis function (save only weights)
• Interbeat subtract average beat and code residuals
(linear prediction or transform-coding)
• Interlead extend intrabeat methods to multilead
environment
Thank you!