Skip to main content

Table 1 Comparison between different ICA algorithms

From: MetICA: independent component analysis for high-resolution mass-spectrometry based non-targeted metabolomics

 

MetICA

FastICA

FastICA

icapca

icamix

kernel-ICA

Bayesian

18 Clusters

'Parallel'

'Deflation'

  

Gaussian

[1] Variance Explained

90 %

90 %

90 %

75.7 %

99 %

99 %

99 %

[2] Component Extracted

18

20

20

7

43

43

9

[3] Maximal Kurtosis

44.1

43.9

44.1

44.1

43.6

3.9

29.8

[4] Minimal abs(Kurtosis)

1.6

3.4

1.9

0.5

15.1

0.008

0.007

[5] Minimal Kurtosis

−1.6

3.4

−2

0.5

15.1

−1.7

−0.9

[6] Stable Components

12/18

20/20

9/20

3/7

43/43

0/43

1/9

[7] Model Selection

HCA

-

-

LOO-CV

Likelihood

-

BIC

[8] Component Order

Bootstrap

-

Deflation

Variance

-

Deflation

Kurtosis

  1. Seven ICA algorithms were compared based on [1] maximal percentage of variance the algorithm could handle (depending on the computer memory), [2] optimal number of components that the algorithm suggests, [3] kurtosis of the most super-Gaussian component [4] kurtosis of the most Gaussian component, [5] minimal kurtosis of components (the most sub-Gaussian when it is negative), [6] number of consistent components extracted in all 10 algorithms runs with an absolute Spearman's correlation between them higher than 0.8 and on whether the algorithm suggests [7] model selection criteria [8] importance order of components