5.5 Performance Analysis using Feature Based on DWT Coefficients of
5.5.3 Multiple Heartbeat Analysis
In Table 5.4, performance of the proposed method based on DWT coefficient of ACF of ECG for multiple heartbeat analysis is summarized. Comparing Table 5.3 and 5.4, it can be inferred that using multiple heartbeat for calculating DWT coefficient of ACF of ECG results in lower FRR, no FAR , very high authentication
Fig. 5.6: Error performance using DWT approximate coefficients of ACF of ECG for single heartbeat analysis
Fig. 5.7: Authentication and Identifi cation accuracies in percentage (%) as a function of distance threshold using DWT approximate coefficients of ACF of ECG for single heartbeat analysis
and identification accuracies, higher precision, sensitivity and specificity with respect to the parameters results from the same feature derived using the single heartbeat.
Again, analyzing the results shown in Table 5.4 with the results obtained for DWT coefficients of ECG signals presented in Table 5.2, it can be clearly observed that the feature based on DWT coefficient of ACF of ECG results in improved performance in terms of all the parameters when multiple heartbeat analysis is performed in both the cases. In particular, in this case, all accepted beats are classified with their true identity thus providing an identification accuracy of 100%. Also, both sensitivity and specificity are attained as 1 meaning no testing beat is misclassified and no intruder is able to access system.
Table 5.4: Performance of the proposed method based on DWT coefficient of ACF of ECG for multiple heartbeat analysis
5.6 Performance Analysis using Feature Based on Cross-Correlation in Curvelet Domain
5.6.1 Results using Reduced Cross-Correlation Feature Based on Dominant Energy Bands
Single Heartbeat analysis
The performance of proposed method using reduced cross-correlation feature based on dominant energy bands is examined and the results are given in Table 5.5.
It is seen from the table that, although 113 are heartbeats are falsely rejected among 1140 heartbeats giving FRR of 9.91 %, whereas 23 beats out 280 are falsely accepted as intruders resulting in a FAR 8.07%. Thus, the proposed feature provides authentication and identification accuracies of 91.01 % and 94.8 %, respectively. It can be also concluded that, a precision of 93.6 %, sensitivity of 0.96 and specificity of 0.92 are achieved while using a single heartbeat for extracting the proposed reduced feature set.
Fig. 5.8 demonstrates the tradeoff s between FAR and FRR for different distance threshold values. From the fi gure, it can be stated that at lower threshold value, less intruder gets access to the system giving low FAR, but at the same time the rate of rejection of genuine user becomes very high giving high FRR. The FRR and FAR are optimized with an approximate EER of 9 %. The authentication and identification accuracies in percentage (%) are shown in Fig. 5.9. This figure shows that as the threshold increases, number of FA increases and these beats match with the trained person falsely. Thus, the values of both authentication and identification accuracies reduce gradually.
Table 5.5: Performance of the proposed method using reduced cross-correlation feature based on dominant energy bands for single heartbeat analysis
Fig. 5.8: Error performance using reduced cross-correlation feature based on domi- nant energy bands for single heartbeat analysis
Fig. 5.9: Authentication and Identifi cation accuracies in percentage (%) as a function of distance threshold using reduced cross-correlation feature based on dominant energy bands for single heartbeat analysis
Multiple Heartbeat Analysis
In Table 5.6, FRR, GAR, FAR , authentication accuracy, identification accuracy, precision, sensitivity and specificity obtained using the reduced cross-correlation feature based on dominant energy bands for multiple heartbeats is shown.
Comparing Tables 5.5 and 5.6, it is vivid that both authentication accuracy and identification accuracy improve with values 95.13% and 97.16%
respectively, when decision is made after applying majority voting, the FAR and FRR has reduced signifi cantly, such as 5.26 % and 4.47 %, respectively.
Consequently, precision, sensitivity and specificity have increased to higher values, such as 96.47 %, 0.98 and 0.95 in comparison to the case while using the same feature derived from the single beat.
Table 5.6: Performance of the proposed method using reduced cross-correlation feature based on dominant energy bands for multiple heartbeat analysis
5.6.2 Results using Reduced Cross-Correlation Feature Based on PCA
In view of reducing computational complexity, dimension reduction of the feature space plays an important role. In the proposed method, the task of feature dimension reduction is performed using PCA. The effect of feature dimension reduction upon diff erent performance parameters is shown using single heartbeat and multiple heartbeat analyses.
Single Heartbeat analysis
In Table 5.7, performance of the proposed method using reduced cross- correlation feature based on PCA for single heartbeat analysis is shown. The results are analyzed in terms of FRR, FAR, authentication accuracy, identification accuracy, precision, sensitivity and specifi city for different numbers of principal components, such as 4,8,16 and 24. The FRR and FAR reduce, authentication and identifi cation accuracies increase, precision, sensitivity and specifi city increase with the increase of principal components from four to sixteen. The performance parameters reach to optimum values while sixteen principal components are employed for authentication and identification. It is to be noted that the performance degrades if the principal components more than sixteen are adopted. However, for single heartbeat analysis it can be concluded that the proposed method using reduced cross-correlation feature based on sixteen PCA yields authentication and identification accuracies of 96.7 %
and 98.99 %, respectively, which are higher than that obtained using reduced cross- correlation feature based on dominant energy bands reported in Table 5.5.
Thus even with low feature dimension, an improved performance is achieved with lower computational burden.
Table 5.7: Performance of the proposed method using reduced cross-correlation feature based on PCA for single heartbeat analysis
A scatter plot using fi rst two principal components of the cross-correlation feature in curvelet domain for five different persons are shown in Fig. 5.10.
This figure highlights the inter-person separability and intra-person compactness of the two principal components corresponding to the proposed feature through clustering analysis, where forty beats for each person are considered. It is observable from the figure that the two principal components for forty beats of a person are concentrated to form a cluster and the clusters of different persons are found separable. The amplitude variation of all sixteen principal components derived for the cross-correlation feature in curvelet domain are plotted for five different persons in Fig. 5.11. In this plot, for each person, mean of principal components considering forty beats are presented. It is found from this fi gure that the reduced feature of size sixteen presented in terms of mean principal components is capable of providing separability among different persons.
Multiple Heartbeat analysis
In Table 5.8, performance of the proposed method using reduced cross- correlation feature based on 4,8,16 and 24 PCAs is shown in terms of FRR, FAR, authentication accuracy, identification accuracy, precision, sensitivity and specificity for multiple heartbeat analysis. As in case of single heartbeat analysis, the FRR and FAR reduce, authentication and identification accuracies increase, precision, sensitivity and specificity increase with the increase of principal components from four to sixteen. The performance parameters reach to optimum values while sixteen and twenty four
Amplitude of Principal Components2nd Principal Components
7
Person 1
6 Person 2
Person 3
5 Person 4
Person 5 4
3 2
10 2 4 6 8 10 12 14
1st Principal Components
Fig. 5.10: Scatter plot using reduced cross-correlation feature using two principal components for fi ve different persons
15 10 5 0
−5
−10 2 4 6 8 10 12 14 16
Principal Components
Fig. 5.11: Principal components of reduced cross-correlation feature for five different persons
principal components are employed for authentication and identification.
However, in order to keep the feature vector of small size, sixteen principal components are considered. Comparing Tables 5.7 and 5.8, it is vivid that both authentication ac- curacy and identification accuracy improve with very high values of 99.74 % and 100 %, respectively, when decision is made from the multiple heartbeats instead of using single heartbeat. After applying majority voting, the FRR has reduced to a very low value, such as 0.53 % and FAR is reduced to zero value. Consequently, precision, sensitivity and specificity have increased to the highest values, such as 100%, 1 and 1 in comparison to the case while using the same feature derived from the single heartbeat. However, for multiple heartbeat analysis it can be concluded that the proposed method using reduced cross-correlation feature based on sixteen
PCA yields higher authentication and identification accuracies than that obtained using reduced cross-correlation feature based on dominant energy bands reported in Table 5.6.
Table 5.8: Performance of the proposed method using reduced cross-correlation feature based on PCA for multiple heartbeat analysis
5.7 Performance Analysis using Feature Based on Mean of Column Elements in Curvelet Domain
5.7.1 Results using Reduced Mean of Column Elements Fea- ture Based on Dominant Energy Bands
Single Heartbeat Analysis
The performance of proposed method using reduced MC elements feature based on dominant energy bands is examined and the results are summarized in Table 5.9. It is seen from the table that, only 8 heartbeats are falsely rejected among 1140 heartbeats giving FRR of 0.70%, whereas 1 beat out 280 are falsely accepted as intruders resulting in a FAR 0.35 %. Thus, the proposed feature provides authentication and identification accuracies of 99.47 % and 99.15 %, respectively. It can be also concluded that, a precision of 98.94 %, sensitivity of 0.99 and specificity of 1 are achieved while using a single heartbeat for extracting the proposed reduced feature set.
Fig. 5.12 demonstrates the tradeoffs between FAR and FRR for different distance threshold values. From the figure, it can be stated that at lower threshold value, very less number of intruder gets access to the system giving very low FAR, but at the same time the rate of rejection of genuine user becomes very high giving high FRR. The FRR and FAR are optimized with an approximate EER less than 1 %. The authentication and identification accuracies in percentage (%) are shown in Fig. 5.13. This figure shows that the threshold increases, less number of FA occurs and less beats match with the trained person falsely.
Therefore, the values of both authentication and identification accuracies reach to more than 99 % at the point of EER.
Comparing Tables 5.5 and 5.9 and Figs. 5.8, 5.9, 5.12, 5.13, it can be concluded that performance of the proposed method using the reduced mean of column elements feature based on dominant energy bands is better in comparison to that obtained using the reduced cross-correlation feature based on dominant energy bands.
Table 5.9: Performance of the proposed method using reduced mean of column elements feature based on Dominant Energy Bands for single heartbeat analysis
Fig. 5.12: Error performance using reduced mean of column elements feature based on dominant energy bands for single heartbeat analysis
Multiple Heartbeat Analysis
In Table 5.10, FRR, GAR, FAR , authentication accuracy, identification accuracy, precision, sensitivity and specificity obtained using the reduced mean of column elements feature based on dominant energy bands for multiple heartbeats is shown. Comparing Tables 5.9 and 5.10, it is clear that both authentication accuracy and identification accuracy improve with very high values of 100 % and 99.79 %, respectively, when decision is made from the multiple heartbeats instead of using single heartbeats. After applying majority voting, the FAR and FRR has reduced to zero values. Consequently, precision, sensitivity
and specificity have increased to very high values, such as 99.74 %, 1 and 1 in comparison to the case while using the same feature derived from the single beat.
Comparing Tables 5.6 and 5.10, for multiple beat analysis is seen that using reduced MC elements feature based on dominant energy bands provides improved performance parameters relative to that obtained using reduced cross-correlation feature based on dominant energy bands.
Table 5.10: Performance of the proposed method using reduced mean of column elements feature based on Dominant Energy Bands for multiple heartbeat analysis
5.7.2 Results using Reduced Mean of Column Elements Feature Based on PCA
Single Heartbeat analysis
In Table 5.11, performance of the proposed method using reduced mean of column elements feature based on PCA for single heartbeat analysis is shown. The results are analyzed in terms of FRR, FAR, authentication accuracy, identification accuracy, precision, sensitivity and specificity for different numbers of principal components, such as 4,8,16 and 24. The FRR and FAR reduce, authentication and identification accuracies increase, precision, sensitivity and
Specificity increase with the increase of principal components from four to sixteen. The performance parameters reach to optimum values while sixteen principal components are employed for authentication and identification. It is to be noted that the performance degrades if the principal components more than sixteen are adopted. However, for single heartbeat analysis it can be concluded that the proposed method using reduced mean of column elements feature based on sixteen PCA yields authentication and identification accuracies of 99.69 % and 100 %, respectively, which are higher than that obtained using reduced mean of column elements feature based on dominant energy bands reported in Table 5.9.
Table 5.11: Performance of the proposed method using reduced mean of column elements feature based on PCA for single heartbeat analysis
A scatter plot using first two principal components of the MC elements feature in curvelet domain for five different persons are shown in Fig. 5.14.
This figure highlights the inter-person separability and intra-person compactness of the two principal components corresponding to the proposed feature through clustering analysis, where forty beats for each person are considered. It is observable from the figure that the two principal components for forty beats of a person are concentrated to form a cluster and the clusters of different persons are found separable. The amplitude variation of all sixteen principal components derived for the MC elements feature in curvelet domain are plotted for five different persons in Fig. 5.15. In this plot, for each person, mean of principal components considering forty beats are presented. It is found from this figure that the reduced feature of size sixteen presented in terms of mean principal components is capable of providing separability among different persons.
Amplitude of Principal Components2nd Principal Components
0
−0.05
−0.1
Person 1 Person 2 Person 3 Person 4 Person 5
−0.15
−0.20 0.1 0.2 0.3 0.4 0.5 0.6 0.7
1st Principal Components
Fig. 5.14: Scatter plot using reduced MC elements feature using two principal com- ponents for five different persons
1.5 1 0.5 0
−0.5
−1
−1.5
Person 1 Person 2 Person 3 Person 4 Person 5 Person 6
2 4 6 8 10 12 14 16
Principal Components
Fig. 5.15: Principal components of reduced MC elements feature for five diff erent persons
Multiple Heartbeat Analysis
In Table 5.12, performance of the proposed method using reduced MC elements feature based on 4,8,16 and 24 PCAs is shown in terms of FRR, FAR, authentication accuracy, identification accuracy, precision, sensitivity and specificity for multiple heartbeat analysis. As in case of single heartbeat analysis, the FRR and FAR reduce, authentication and identifi cation accuracies increase, precision, sensitivity and specificity increase with the increase of principal components from four to sixteen. The performance parameters reach to optimum values while sixteen and twenty four principal components are employed for authentication and identification. However, in order to keep the feature vector of small size, sixteen principal components are considered. Comparing Tables 5.11 and 5.12, it is vivid that both authentication accuracy and identification accuracy improve with the highest values of 100 % and 100%, respectively, when decision is
made from the multiple heartbeats instead of using single heartbeat. After applying majority voting, the FRR and FAR have reduced to zero values.
Consequently, precision, sensitivity and specificity have increased to the highest values, such as 100 %, 1 and 1 in comparison to the case while using the same feature derived from the single heartbeat.
However, for multiple heartbeat analysis it can be concluded that the proposed Method using reduced MC elements feature based on sixteen PCA yields higher authentication and identification accuracies than that obtained using reduced MC elements feature based on dominant energy bands reported in Table 5.10.
Table 5.12: Performance of the proposed method using reduced mean of column elements feature based on PCA for multiple heartbeat analysis
5.8 Performance Comparison
In the previous sections, performance analysis based on the proposed features ex- tracted using wavelet and curvelet domains is performed. In order to compare it’s performance with some of the existing methods, such as MFCC [21] and AC/DCT [14], another table is presented in 5.13. It is to be mentioned that the other comparison methods reported results only based on single heartbeat analysis.
Among the two methods based on DWT, the proposed method based on DWT approximate coefficients of ACF of ECG is capable of providing higher identification accuracy (%) compared to the MFCC and AC/DCT methods even with single heartbeat analysis and identification accuracy of the proposed feature reaches to 100 (%) using multiple heartbeat analysis.
In the curvelet domain, although the reduced cross-correlation feature based on dominant energy bands show lower identifi cation accuracy (%) for both single and multiple heartbeat analysis, the cross-correlation feature outperforms the comparison methods while it is reduced by using PCA even in case of single heartbeat analysis. The cross-correlation feature reduced based on PCA is able to produce 100 (%) identification accuracy while using multiple heartbeat analysis.
It is interesting to note that the mean of column elements feature in curvelet domain which is reduced based on dominant energy bands, yields higher identification accuracy (%) compared to the existing methods for both single and multiple heartbeat analysis.
It is as expected the proposed mean of column element feature in curvelet domain provides 100 (%) identification accuracy while it is reduced by employing PCA not only for multiple heartbeat analysis, also for single heartbeat analysis.
Table 5.13: Performance comparison of the other and the proposed methods using single and multiple heartbeat analysis
5.9 Conclusion
DWT approximate coefficients of ECG and DWT approximate coefficients of ACF of ECG are found as suitable feature sets for human identification. In the Curvelet domain considering the advantageous aspects of curvelet domain, cross-correlation and mean of column elements in curvelet domain are proposed for human identification. The feature sets are evaluated using standard MIT-BIH database in terms of several performance parameters. Detail simulation result show that the proposed feature sets are eff ective in identifying humans compared to some of the existing methods.
Chapter 6
CONCLUSION
6.1 Concluding Remarks
In this thesis, time frequency analysis like wavelet transform is utilized for calculating feature sets via approximate DWT coefficients. It is found that instead of considering the approximate DWT coefficients of ECG, taking approximate DWT coefficients of ACF of ECG provides a sufficient an effective feature set capable of identifying persons with higher accuracy. Considering the advantageous aspects of curvelet transform compared to wavelet transform, cross-correlation and mean of column elements in curvelet domain are proposed to achieve higher authentication and identification accuracy. In order to reduce complexity and computational burden, curvelet domain features are reduced based on dominant energy bands in coarse and detail layers. Employment of PCA on curvelet feature space are found more effective in reducing feature dimension and producing improved performance for human identification. Simulations are carried out to evaluate the performance of the proposed method in terms of FAR, FRR, authentication accuracy, precision, sensitivity, specificity and identification accuracy. It is shown that the proposed method outperforms some of the state- of-the-art methods with superior efficacy.
6.2 Contributions of the thesis
The major contributions of this thesis are:
1. A set of DWT approximate coefficients based feature is proposed for human identification using ECG signal.
2. Another set of DWT approximate coefficients based feature is developed by applying DWT on autocorrelation function of ECG signal.