450
S. Chen et al.
60
65
70
75
80
85
90
95
Accuracy(%)
0.1
0.15
0.2
0.25
0.3
0.35
0.4
Threshold
20
25
30
35
40
45
50
Number of Features
Accuracy with Features in AUC
Accuracy with Features in Threshold
Number of Features Selected
Fig. 6. Comparison between features in threshold and AUC. For each group of features computed under 15 different threshold values, the classifier was constructed and
evaluated.
The applicability of stability selection for feature selection: It was
demonstrated that logistic regression had a better performance in MCI classification when combined with stability selection [5]. In light of that, whether
stability selection would improve the performance of LASSO in feature selection
with the data of this study was examined. The area under the ROC curve and
the accuracy of the three curves shown in Fig. 5 are: LASSO + Stability Selection (28 features), 0.9755 and 86%; LASSO (28 features), 0.9783 and 87%; and
LASSO (36 features), 0.9603 and 84%. The result indicates that the area under
the ROC curve when combining LASSO with stability selection for classification
was slightly larger than that when using LASSO only (36 features were selected).
It is noteworthy that when employing top 28 features with the highest weight
value to control the variable of the number of features the same as LASSO with
stability selection, using LASSO showed even larger area under the ROC curve.
Combining LASSO with stability selection is time consuming for calculation in
loops, while could not bring obvious improvement to the classification. Therefore, LASSO has higher applicability in this study than which been combined
with stability selection.
4 Conclusion
In this study, based on complex brain networks, a machine learning method
was explored to discriminate MCI from CN. By calculating sigma values, it was
discovered that the brain of CN is more effective in information transmission
than that of MCI. Features in the form of AUC showed better performance in
classification than features under most thresholds. LASSO was demonstrated as
an effective method in feature selection, with an accuracy of 87.0% and an area
under ROC curve of 0.9783 by 28 optimal features in the form of AUC.
S. Chen et al.
60
65
70
75
80
85
90
95
Accuracy(%)
0.1
0.15
0.2
0.25
0.3
0.35
0.4
Threshold
20
25
30
35
40
45
50
Number of Features
Accuracy with Features in AUC
Accuracy with Features in Threshold
Number of Features Selected
Fig. 6. Comparison between features in threshold and AUC. For each group of features computed under 15 different threshold values, the classifier was constructed and
evaluated.
The applicability of stability selection for feature selection: It was
demonstrated that logistic regression had a better performance in MCI classification when combined with stability selection [5]. In light of that, whether
stability selection would improve the performance of LASSO in feature selection
with the data of this study was examined. The area under the ROC curve and
the accuracy of the three curves shown in Fig. 5 are: LASSO + Stability Selection (28 features), 0.9755 and 86%; LASSO (28 features), 0.9783 and 87%; and
LASSO (36 features), 0.9603 and 84%. The result indicates that the area under
the ROC curve when combining LASSO with stability selection for classification
was slightly larger than that when using LASSO only (36 features were selected).
It is noteworthy that when employing top 28 features with the highest weight
value to control the variable of the number of features the same as LASSO with
stability selection, using LASSO showed even larger area under the ROC curve.
Combining LASSO with stability selection is time consuming for calculation in
loops, while could not bring obvious improvement to the classification. Therefore, LASSO has higher applicability in this study than which been combined
with stability selection.
4 Conclusion
In this study, based on complex brain networks, a machine learning method
was explored to discriminate MCI from CN. By calculating sigma values, it was
discovered that the brain of CN is more effective in information transmission
than that of MCI. Features in the form of AUC showed better performance in
classification than features under most thresholds. LASSO was demonstrated as
an effective method in feature selection, with an accuracy of 87.0% and an area
under ROC curve of 0.9783 by 28 optimal features in the form of AUC.