A comparison of feature extraction methods for prediction of neuropsychological scores from functional connectivity data of stroke patients

Brain Informatics

Table 1 Results of elastic-net regression

Cognitive domain	Method	\(R^{2}\)	MSE	BIC	\(\lambda\)	\(\alpha\)	k	NZ
Language (\(n=95\))	PCA	0.52	0.48	404.10	0.22	0.001	45	45
	ICA	0.51	0.49	305.60	0.11	0.25	25	23
	DL	0.45	0.55	335.52	0.09	0.25	30	27
	NNMF	0.44	0.56	404.95	0.04	0.50	45	42
Spatial memory (\(n=78\))	PCA	0.23	0.76	295.45	0.11	1	50	22
	ICA	0.24	0.75	395.18	0.56	0.001	45	45
	DL	0.20	0.79	285.88	0.09	1	40	19
	NNMF	0.21	0.78	371.87	0.09	0.75	75	39
Verbal memory (\(n=78\))	PCA	0.34	0.65	327.62	0.09	0.75	45	32
	ICA	0.28	0.72	391.29	0.44	0.001	45	45
	DL	0.18	0.81	444.48	0.56	0.001	55	55
	NNMF	0.10	0.88	451.38	1.42	0.001	55	55

Performance of elastic-net regression models in the prediction of neuropsychological scores as a function of the feature extraction method. The value of the optimized parameters (\(\lambda\), \(\alpha\), and k) and the number of non-zero features (NZ) are also reported. \(R^{2}\): percentage of variance explained.
MSE mean squared error, BIC Bayesian information criterion