Synopsis

A proper method for distinguishing Schizophrenia (SZ) patients from people belonging to normal controls is invaluable for diagnostics. Standard techniques of identifying SZ patients include a variety of clustering-based methods. For example, those based on independent component analysis (ICA). However, they typically focus on average connectivity. In this work, we propose a novel method for SZ patient identification based on the products of ICA of default mode network (DMN) and cognitive enhancement therapy (CET). Numerical experiments show that the proposed method achieves statistically significant results for differentiating the two groups of subjects.

INTRODUCTION

Central executive network (CEN) and default mode network (DMN) are a pair of anticorrelated networks¹. Under cognitive demanding tasks, CEN recruits necessary cognitive resources and exhibits high BOLD signal in fMRI images. In contrast, DMN is more active under task-free or resting condition. While both CEN and DMN abnormalities have been associated with schizophrenia individually², whether there is an alteration in the interaction of the two networks remains largely unknown. The present study examines the interaction using phase-space embedding of resting state fMRI data.

METHODS

Network interaction. One hundred and twenty one normal control (NC) subjects and 50 age- and sex-matched schizophrenia (SZ) patients from the UCLA Consortium for Neuropsychiatric Phenomics LA5c Study were included in the analysis. Standard prepossessing was performed on the rfMRI data using FSL. Group CEN and DMN maps were then identified using concatenated group ICA analysis implemented in FSL MELODIC. Time courses of individual CEN and DMN components was acquired using dual regression analysis. The time series were then converted to percentiles to reflect the relative activity strength across the scan. The interaction between the two networks was defined as the product of percentile time series:

T_t=D_t*C_t,t=1,...n (1)

where n is the total number of samples points, D_t and C_t represent DMN and CET percentile time series respectively. Phase-space embedding. Since phase space has the same dynamic property as the underlying system⁴, we reconstructed the phase space of the interaction time series T_t for each subject. For a given time series, phase-space embedding results in m time-consecutive variables in the new space and each variable contains a series of delayed points with interval of τ. Specifically, the embedding Y_l of the interaction time series T_t is constructed as follows:

Y=[T1...T1+(m-1)τ;... ... ...; TK ... Tn ](2)

Where m is the embedding dimension, is the delay, n is the number of points in the original time series, and K is the length of new variables, which equals n-(m-1)τ. The embedding dimension m can be determined by the false nearest neighbor (FNN) algorithm, and the parameter delay τ can be determined by minimizing mutual information between reconstructed new variables. Feature extraction of the embedding space. Followed by Seraj et al., we obtained the moment invariant features in the discriminant matrix D, including mean, standard deviation and major minors⁴, and compare them between NC and SZ using ANOVA. As an illustration, Figure 1 shows two phase-space embedding and their related ellipsoid fitting. It can be clearly seen that smaller first order major minor of matrix D corresponds to smaller ellipsoid and smaller phase-space compacity.

RESULTS

Embedding parameters. The embedding dimension m and the time delay τ based on 121 healthy control subjects and 50 schizophrenia subjects are depicted in Fig 2.(A) and (B) are optimal settings for this study.

With the optimal parameter setting, SPSS ANOVA analysis showed significant difference between the two groups. Post-hoc analysis using LSD revealed that the schizophrenia group had smaller values in all features, compared to the control group. To exclude the possible contamination of percentiling data, we repeated the whole pipeline using processed BOLD residuals rather than the percentile data for network interaction, and the results were similar.

Discussion and Conclusion

Acknowledgements

References

1 Rashid, Barnaly, et al. "Classification of schizophrenia and bipolar patients using static and dynamic resting-state fMRI brain connectivity." Neuroimage 134 (2016): 645-657.

2 Allen, Elena A., et al. "A baseline for the multivariate comparison of resting-state networks." Frontiers in systems neuroscience 5 (2011): 2.

3 Poldrack, Russell A. "Region of interest analysis for fMRI." Social cognitive and affective neuroscience 2.1 (2007): 67-70.

4 Takens, Floris. "Detecting strange attractors in turbulence." Dynamical systems and turbulence, Warwick 1980. Springer, Berlin, Heidelberg, 1981. 366-381.

5 Sayed, Khaled, et al. "Characterization of phase space trajectories for Brain-Computer Interface." Biomedical Signal Processing and Control 38 (2017): 55-66.

6 Gorgolewski, Krzysztof J., Joke Durnez, and Russell A. Poldrack. "Preprocessed Consortium for Neuropsychiatric Phenomics dataset." F1000Research 6 (2017).