4. Tele ECG in the World
“A wireless ECG system for continuous event recording and
communication to a clinical alarm station”
Fensli, R.; Gunnarson, E.; Hejlesen, O
Engineering in Medicine and Biology Society, 2004. IEMBS apos;04. 26th
Annual International Conference of the IEEE.
Online Tele-ECG System in a remote Brazilian Emergency Room: The
evaluation of the time interval from door to discharge of cardiac patients
Adolfo L F Sparenberg1,2, Thais Russomano1,2 , Eleonora R Soares2,
Tatiane Schaun2, Ricardo B Cardoso1, Robert Timm2,
Scientists of the Bhabha Atomic Research Centre India have
developed a compact, low-cost and portable tele-electro cardiogram
(ECG) system that could be controlled by mobile phone by means of a
bluetooth connection in Feb 2009.
5. Aim of our research
Provide Automatic cardiac disease Analysis.
Remove the need of a specialized cardiologist
Make expert and accurate judgments based on patient’s past
record
Remote diagnosis during absence of medical personnel and
experts
A simple, easy and cost effective diagnosis
Provide continuous storage and assessment of patient’s ECG
6. Main features of Neural Networks
Neural networks learn by experience rather than by
remodeling or reprogramming ………Intelligence
They have the ability to generalize
They do not require a prior understanding of the
process
They are robust to noisy data
VLSI implementation is very easy
7. Current and future research
This research is ongoing in various parts of world
There are annual international competitions for
getting better results in this research area
Highly accurate results have not been obtained till
today
8. Main Goals of our research
To carry out research into Automated interpretation
of ECG signal.
To find the efficient features of the ECG signal
To use Neural Networks for signal classification
9. Software tools and ECG data used in
this research
Simulation based research, in MATLAB
Signal processing toolbox of Matlab SPtool
Neural Network tool box for Matlab NNtool
Availability of real time ECG data of patients at MIT-BIH
database
10. Experimental tasks of our research
Signal pre-processing of ECG for noise removal
Interpretation of ECG
QRS complex and other fiducially obtained point
detection
Efficient feature extraction for input to NN
Classification of cardiac problems using Neural
Networks
Fiducial
SIGNAL
QRS Feature NN
PRE- Point
Detection Extraction Classification
PROCESSING Detection
11. Electrocardiogram (ECG)
Normal ECG values for
healthy persons
Wave Duration (s)
P 0.08-0.10
QRS 0.06-0.10
T 0.12-0.16
Shape of ECG
14. Right Bundle Branch Block Beat
Diagnosis of RBBB is mainly based on widened
QRS 0.12 seconds or more
15. Atrial Premature Contraction Beat (APC)
A premature beat, appears early than expected
Decrease in R-R interval
16. Data for NN Classification
Feature Type Source Description
1 M R, S waves R-S interval
2 M P, R waves P-R interval
3 M QRS complex QRS width
4 M Q, T waves Q-T interval
5 M R waves R amplitude
6 M R waves R-R interval (HBR)
7 S QRS complex QRS energy
8 S ECG waves Auto correlation coefficient
9 S ECG waves Mean or expectation vector
10 S Histogram Maximum Amplitude of the signal
11-23 C ECG waveform 13 compressed ECG sample
M= Morphological points S=Statistical points C=compressed points
17. Compressed Signals
ECG contains 360 points for one beat but to reduce
complexity we have taken 52 essential points and
further the signal is compressed to 4:1
18. Data for NN Classification
Feature Type Source Description
1 M R, S waves R-S interval
2 M P, R waves P-R interval
3 M QRS complex QRS area
4 M Q, T waves Q-T interval
5 M R waves R amplitude
6 M R waves HBR
7 S QRS complex QRS energy
8 S ECG waves Auto correlation coefficient
9 S ECG waves Mean or expectation vector
10 S Histogram Maximum Amplitude of the signal
11-23 C ECG waveform 13 compressed ECG sample
M= Morphological points S=Statistical points C=compressed points
19. Experimental tasks of our research
Signal pre-processing of ECG for noise removal
Interpretation of ECG
QRS complex and other fiducially obtained point
detection
Efficient feature extraction for input to NN
Classification of cardiac problems using Neural
Networks
Fiducial
SIGNAL
QRS Feature NN
PRE- Point
Detection Extraction Classification
PROCESSING Detection
20. Simple architecture interconnections
Architecture of Neural Networks made in our
research. This is feed forward back propagation.
R-S
N
P-R
A
Q-T
V
R
Comp 1
L
Comp 2
P
23 element Input Hidden Layer 6 neurons Output
Layer Layer
21. Network used in our research
Training function
Learning function
Number of hidden layers
Number of Neurons in hidden layers
Number of nodes in input and output layer
Time for training
Number of epochs
23. Experimental tasks of our research
Signal pre-processing of ECG for noise removal
Interpretation of ECG
QRS complex and other fiducially obtained point
detection
Efficient feature extraction for input to NN
Classification of cardiac problems using Neural
Networks
Fiducial
SIGNAL
QRS Feature NN
PRE- Point
Detection Extraction Classification
PROCESSING Detection
24. Pre-Processing of ECG signal
Filtering of the ECG signal
Removing High Frequency components
Removing Low Frequency components
Removing Power line interference
31. Experimental tasks of our research
Signal pre-processing of ECG for noise removal
Interpretation of ECG
QRS complex and other fiducial point detection
Efficient feature extraction for input to NN
Classification of cardiac problems using Neural
Networks
Fiducial
SIGNAL
QRS Feature NN
PRE- Point
Detection Extraction Classification
PROCESSING Detection
42. Data for NN Classification
Features Type Source Description
1 M R, S waves R-S interval
2 M P, R waves P-R interval
3 M QRS complex QRS area
4 M Q, T waves Q-T interval
5 M R waves R amplitude
6 M R waves HBR
7 S QRS complex QRS energy
8 S ECG waves Auto correlation coefficient
9 S ECG waves Mean or expectation vector
10 S Histogram Maximum Amplitude of the signal
11-23 C ECG waveform 13 compressed ECG sample
M= Morphological points S=Statistical points C=compressed points
43. Training of Network
Training performance of Network.
Training Samples: 100
Testing Samples: 100
Network Input Output Hidden Epochs Elapsed MSE
Layer layer Layer Time (sec)
Classifier 23 6 5 12 9.0650 0.00484
Training Performance of Network
44. Classification Results
Training Samples: 100
Testing Samples: 100
Network Input Neurons Output Neurons Heart Problem Recognition rate % Average rate %
N 88
A 96
V 91
Classifier 23 6 90.6
R 98
L 84
P 87
N = Normal Beat
A = Atrial Premature Beat
V = Ventricular Premature Beat Beat
R = Right Bundle Branch Block Beat
L = Left Bundle Branch Block
P = Paced Beat
45. Conclusion and Future Recommendations
The aim of the research was to find the efficient
features which are given to NN for classification.
We have worked on only 6 beats and in future this can
be extended to greater number of beats. We used
100 number of Training and testing samples and this can be
extended, for greater average rate of recognition.
This is ongoing research and is conducted in different
part of world and different competitions are held
worldwide for best classification at www.physionet.org