Synopsis

Functional network from ICA using fMRI data has been applied to identify biomarkers of brain disorders. However, the networks from ICA might be slightly different, making the comparison of results across different studies/diseases difficult. We propose a data-driven framework to estimate functional network maps and their inter-connectivity for linking neuromarkers among different disorders and studies. Our method is capable of computing functional networks which are optimized for independence based on each coming individual-subject data, and remaining their correspondence across different subjects by using unbiased templates. The results show this approach is an effective method for studying and classifying multiple-disorders.

Introduction

Methods

We first generate robust network templates from large-sample healthy controls (HCs), and then based on the templates we can compute the individual-subject networks and the connectivity between networks using the group information guided ICA (GIG-ICA) [5-7]. 1005 HCs from GSP and 823 HCs from HCP were used separately to compute 100 group-level ICs, followed by selecting 54 reliable networks which were highly matched between the two groups. For each new individual, we calculate the network spatial maps and time-series using GIG-ICA. Functional network connectivity (FNC) matrix is then obtained by computing the correlations between post-processed time series. This allows us to capture individual variability, leverage a multivariate framework, while providing a straightforward implementation. Based on network maps and FNC, we can investigate biomarkers for different diseases.

Study 1: Investigate if schizophrenia (SZ) and autism spectrum disorder (ASD) show overlapping and unique connectivity changes

Using fMRI data of 2980 subjects including 1665 HCs, 537 SZs and 778 ASDs from BSNIP, FBIRN, COBRE, PK, ABIDEI and ABIDEII, we investigated group differences in network maps and FNC. For each network, voxel-wise one-sample t-test, ANOVA and two-sample t-tests were performed. For each element of FNC matrix, we performed ANOVA and two-sample t-tests to examine the group difference.

Study 2: Investigate if Alzheimer's disease (AD) shows more changes than mild cognitive impairment (MCI)

Four groups of fMRI data (275 HCs, 107 ADs, 279 early MCIs (EMCIs) and 190 late MCIs (LMCIs)) from ADNI were used to identify differences in FNC. Similar to Study 1, we performed ANOVA followed by two-sample t-tests on FNC to examine.

Study 3: Distinguish bipolar disorder (BD) from major depressive disorder (MDD)

Using data from University of Western, the network maps of 33 HCs, 32 BDs and 34 MDDs were used to train a SVM classifier with kernel subspace similarity [8]. Six most discriminative components were selected by nested cross-validation. This classifier was then applied to 12 patients with unclear diagnoses.

Results

Fig. 1 shows the 54 group-level network maps revealed by ICA, grouped into seven function domains including sub-cortical (SC: 5), auditory (AU: 2), sensorimotor (SM: 10), visual (VI: 9), cognitive control (CC: 16), default mode (DM: 7) and cerebellar (CB: 5) networks.

Study 1: The group differences in FNC and network maps are shown in Fig. 2 and Fig. 3, respectively. Compared to HC, SZ primarily shows decreased positive connectivity between SC and CB domains, between SM and VI domains. The shared functional impairments are present for SZ and ASD, mainly involving sub-cortical, visual, sensorimotor and cerebellar regions. ASD in general shows similar but weaker impairments compared to SZ, while ASD and SZ also have unique changes.

Study 2: Our results (Fig. 4) show that AD, EMCI and LMCI exhibit similar changes in brain connectivity, especially the decreased FNC within the CB domain. LMCI shows more similar changes with AD, compared with EMCI. Decreased FNC between SM and CB can be observed in both AD and LMCI, but the changes are much weaker in LMCI.

Study 3: Classification rate of the known-diagnosis group is 93.9% (90.6% sensitivity, 97.1% specificity), and the prediction of 12 independent unclear patients was accurate except for one individual who has an MDD first degree relative. The performance using the NeuroMark framework is slightly better than previous results [8], and the selected most discriminative networks (Fig. 5) can also be visualized.

Conclusion

Acknowledgements

References

1. Calhoun, V.D. and N. de Lacy, Ten Key Observations on the Analysis of Resting-state Functional MR Imaging Data Using Independent Component Analysis. Neuroimaging Clin N Am, 2017. 27(4): p. 561-579.

2. Lee, M.H., C.D. Smyser, and J.S. Shimony, Resting-state fMRI: a review of methods and clinical applications. AJNR Am J Neuroradiol, 2013. 34(10): p. 1866-72.

3. Smith, S.M., et al., Functional connectomics from resting-state fMRI. Trends Cogn Sci, 2013. 17(12): p. 666-82.

4. van den Heuvel, M.P. and H.E. Hulshoff Pol, Exploring the brain network: a review on resting-state fMRI functional connectivity. Eur Neuropsychopharmacol, 2010. 20(8): p. 519-34.

5. Du, Y.H. and Y. Fan, Group information guided ICA for fMRI data analysis. Neuroimage, 2013. 69: p. 157-97.

6. Du, Y.H., et al., A group ICA based framework for evaluating resting fMRI markers when disease categories are unclear: application to schizophrenia, bipolar, and schizoaffective disorders. Neuroimage, 2015. 122: p. 272-280.

7. Du, Y.H., et al., Identifying functional network changing patterns in individuals at clinical high-risk for psychosis and patients with early illness schizophrenia: A group ICA study. Neuroimage Clin, 2017.

8. Osuch, E., et al., Complexity in mood disorder diagnosis: fMRI connectivity networks predicted medication-class of response in complex patients. Acta Psychiatrica Scandinavica, 2018. 138(5): p. 472-482.