This MSc thesis falls in the field of features extraction from ECG signals in order to build automatic analysis system to help cardiologist in their work. In this work we presented new approach for detecting features from ECG signals depending on nonlinear derivative scheme for one dimension 1D NLFS. We started by proposing a theatrical model for exploiting this scheme to detect features from free of noise synthetic ECG signal, then we developed this model to overcome challenges that appears in real noisy ECG signals to build new approach. We started detection by QRS complex peak detection as a first step, and then we extended the approach to detect the onset and end for P and T-waves in the beat. Performance evaluation of our approach have been conducted on records from MIT QT Standard Database that has manual annotations by experts for testing the sensitivity, positive predictivity, mean error and standard deviation for detection of each feature separately. The results obtained from this single lead-based ECG , analysis approach are promising especially in P and T-waves delineation. Because it is the first use of 1D NLFS approach in the field of feature extraction, there is a big chance to enhance the performance in the future and to extend the application of this approach for more type of signals.
Then we tested how much this approach is robust against noise, by applying this approach on synthetic ECG signals with different level of noise. It gives good performance against noise.
5. Motivation
Taleb al-Ashkar 5
Statistics
• 1/3 people in US has Cardiac Problem
• Main reason of mortality in developed countries
• Costs of healing an caring of patients
Fig 3. Cardiac Problems Costs in US
6. Motivation
Taleb al-Ashkar 6
Why we need automatic analysis system
• Decrease costs
• Increase efficiency of diagnosis systems
Fig 4. ECG Automatic Analysis
7. Methodology
Taleb al-Ashkar 7
Methodology of ECG Features Detection
• 1D Nonlinear Filtering Scheme (NLFS)
• Mathematical Model for ECG Features Detection
• Real Approach for ECG Features Detection
8. Methodology
Taleb al-Ashkar 8
1D Nonlinear Filtering Scheme (NLFS)
• Edge Detection Approach
• Decomposing signal into two signals by:
Y +(z)=T(F +(z)S(z))
Y -(z)=T(-F -(z)S(z))
T: Threshold to select the response
F(z): Detector Filter
S(z): Original Signal
13. Methodology
13
Mathematical Model for ECG Features Detection
• Onset of P or T-wave Detection
2. Y+ = Y+ [w0 ,w1 ]
3. Differentiation difY+
4. Linear search
P-wave
Y+
difY+
Fig 9. Onset Detection
14. Methodology
14
Mathematical Model for ECG Features Detection
• End of P or T-wave Detection
1. Defining search window (w0 ,w1 )
2. Y- = Y- [w0 ,w1 ]
3. Differentiation difY-
4. Linear search T-wave
Y-
difY-
Fig 10. End Detection
17. Methodology
17
Real Approach for ECG Features Detection
• Starting from previous Mathematical Model
• Modification to overcome challenges
Fig 13. Real ECG
18. Methodology
18
Real Approach for ECG Features Detection
• QRS Peak Detection
1. Smoothing ECG: Average Filtering
2. 1D NLFS: to get Y+
3. Y+ Differentiation
4. Thresholding
5. Linear Search: C0 is the end of each peak
6. Search window :
QRS peak = max (ECG[C0 -5, C0 +5])
5. Repeat 5, 6 steps up to end of ECG signal
6. Defining RR line
Fig 14. QRS Peak Detection
20. Methodology
20
Real Approach for ECG Features Detection
• Onset of P or T-wave Detection
1. Defining w0 ,w1
2. Y+ = Y+ [w0 ,w1 ]
3. Differentiation by 8 samples step
4. Onset is the index of max value of S(i)
Fig 16. S(i) fro Y+
21. Methodology
21
Real Approach for ECG Features Detection
• End of P or T-wave Detection
1. Defining w0 ,w1
2. Y- = Y- [w0 ,w1 ]
3. Differentiation by 8 S(i)
4. Index min value of S(i)
5. Shifting by 8 samples to get the End
point Fig 17. S(i) for Y-
22. Results
22
Testing on Standard Database
Testing Database
• 12 Records of QTMIT Standard Database
• Contains Manual Annotations by expert
• Each record contains about 30 annotated beats
23. Results
23
Testing on Standard Database
Evaluation Parameters
1. Sensitivity:
2. Positive Predictivity
3. Mean Error
4. Standard Deviation
24. Results
24
Testing on Standard Database
Standard Accepted Error
For deciding Automatic Detection is TP, FP or FN
Table 1. Maximum Accepted Error
25. Results
25
Testing on Standard Database
Testing Results
ME (ms) SD (ms) Se % P+ %
P-onset 1.48 11.55 75.16 75.16
P-end -1.747 13.57 71 71
R peak -3.251 2.487 98.43 98.88
T-end -7.93 12.396 90.7 90.7
Table 2. Results
26. Results
26
Testing on Standard Database
Comparing with other methods
Method Parameters P-onset P-end QRS T-end
This work
Se (%)
P+ (%)
m±s (ms)
75.16
75.16
1.48±11.5
71
71
-1.7±13.5
98.88
98.88
-3.2±2.48
90.7
90.7
-7.9±12.3
WT
Se (%)
P+ (%)
m±s (ms)
98.87
91.03
2.0±14.8
98.75
91.03
1.9±12.8
`99.92
99.88
NA
99.77
97.79
-1.6±18.1
LPD
Se(%)
P+ (%)
m±s (ms)
97.7
91.17
14±13.3
97.70
91.17
-0.1±12.3
NA
99.90
97.71
13.5±27.0
Bayes
Se (%)
P+ (%)
m±s (ms)
99.6
NA
1.7±10.8
99.6
NA
2.5±11.2
NA
100
NA
2.7±13.5
Table 3. Comparing Results
27. Results
27
Testing Against Noise
ECG signal with 8 Different Level of additive noise.
Noise Level L1 L2 L3 L4 L5 L6 L7 L8
PSNR (dB) 113 106.9 100.9 97.4 94.7 86.9 80.8 77.4
0
20
40
60
80
100
120
L1 L2 L3 L4 L5 L6 L7 L8
Se(%)
Noise Levels
Pon
Poff
R
Ton
Toff
Table 4. Noise Levels
Fig 18. Testing against noise results
28. Conclusion
28
Pros
1. Exploiting 1D NLFS in ECG features Detection
2. Fast & Robust to noise approach
3. Possibility to improve performance
Cons:
1. Less performance than other method
2. Not sufficient for special ECG cases