Multi-parametric MRI phenotype with trustworthy machine learning for differentiating CNS demyelinating diseases

Table 2 Diagnostic performance of our method compared with 8 SOTA feature selection algorithms

Category	Method	Roc_auc	Sensitivity	Specificity	Accuracy
Filter	Wilcoxon [28]	0.625 ± 0.120	0.604 ± 0.145	0.585 ± 0.137	0.593 ± 0.104
Filter	Anova [29]	0.879 ± 0.037	0.883 ± 0.042	0.726 ± 0.067	0.789 ± 0.036
Filter	mRMR [30]	0.846 ± 0.041	0.823 ± 0.091	0.754 ± 0.076	0.782 ± 0.050
Wrapper	RF [31]	0.846 ± 0.041	0.828 ± 0.091	0.750 ± 0.075	0.781 ± 0.048
Wrapper	SFS [32]	0.858 ± 0.038	0.782 ± 0.073	0.829 ± 0.121	0.810 ± 0.067
Embedded	Lasso [33]	0.873 ± 0.050	0.884 ± 0.046	0.737 ± 0.087	0.796 ± 0.056
Embedded	ElasticNet [34]	0.850 ± 0.064	0.868 ± 0.070	0.719 ± 0.135	0.779 ± 0.086
Embedded	RFE [35]	0.814 ± 0.072	0.846 ± 0.058	0.612 ± 0.141	0.705 ± 0.090
Ours	Clinical	0.573 ± 0.055	0.716 ± 0.190	0.446 ± 0.179	0.554 ± 0.070
Ours	T2 MRI	0.852 ± 0.053	0.887 ± 0.067	0.733 ± 0.057	0.795 ± 0.045
Ours	T1-MPR MRI	0.880 ± 0.033	0.798 ± 0.102	0.859 ± 0.073	0.835 ± 0.060
Ours	Multi-parametric MRI	0.902 ± 0.027	0.873 ± 0.083	0.869 ± 0.051	0.871 ± 0.044

The Bolded value indicates the highest value in each column
AUC area under the curve, Anova analysis of variance, mRMR maximum relevance minimum redundancy, RF random forest, SFS sequential forward selection, Lasso least absolute shrinkage and selection operator, RFE recursive feature elimination

ISSN: 1479-5876