Biclustered Independent Component Analysis for Complex Biomarker and Subtype Identification from Structural Magnetic Resonance Images in Schizophrenia

Cota Navin Gupta, Eduardo Castro, Srinivas Rachkonda, Theo G.m. van Erp, Steven G. Potkin, Judith M. Ford, Daniel H. Mathalon, Hyo Jong Lee, Bryon A. Mueller, Douglas N. Greve, Ole A. Andreassen, Ingrid Agartz, Andrew R. Mayer, Julia M. Stephen, Rex E. Jung, Juan Bustillo, Vince D. Calhoun, Jessica A. TurnerFrontiers in Psychiatry 2017 · DOI

TL;DR

This paper introduces a novel data-driven framework, Biclustered Independent Component Analysis (B-ICA), to identify distinct subtypes of schizophrenia (Sz) directly from structural brain imaging data (gray matter concentration). By uncovering complex patterns of brain abnormalities, B-ICA reveals subtypes that correlate with specific clinical symptom profiles, offering a more biologically grounded approach to understanding Sz heterogeneity.

What Problem Does This Paper Address?

Schizophrenia is a complex disorder characterized by significant heterogeneity in its clinical presentation and underlying neurobiology. Traditional methods of subtyping Sz based on clinical symptoms alone have been criticized for their lack of stable neurobiological correlates and their tendency to capture transient symptom fluctuations. While neuroimaging techniques like Voxel-Based Morphometry (VBM) and Source-Based Morphometry (SBM) can identify group-level differences in brain structure (e.g., gray matter concentration - GMC), they often struggle to capture the complex, localized patterns of abnormality that might define distinct patient subgroups. This paper aims to bridge this gap by developing a method that can identify these biologically distinct subtypes directly from neuroimaging data, moving beyond symptom-based classifications.

Methods

The core of this paper is the development and application of a Biclustered Independent Component Analysis (B-ICA) framework. This approach combines elements of Source-Based Morphometry (SBM) and Independent Component Analysis (ICA) in a novel way.

Here's a breakdown of the key steps:

  1. SBM Decomposition: The process begins with SBM, a multivariate technique that decomposes the GMC data from Sz patients into a set of independent components. These components represent spatially distributed patterns of GMC variation.
  2. Component Selection: The authors selected two specific SBM components that had previously shown significant group differences between Sz patients and healthy controls and exhibited the largest effect sizes. These components represent distinct brain networks: the insula/superior temporal gyrus/inferior frontal gyrus (I-STG-IFG) and the superior frontal gyrus/middle frontal gyrus/medial frontal gyrus (SFG-MiFG-MFG).
  3. Loading Sorting and Subtype Identification: The "loadings" (which indicate how strongly each subject is associated with a particular component) for these two selected components were then analyzed. Subjects were grouped based on their loading patterns:
  4. S1: Subjects highly weighted on the I-STG-IFG component only.
  5. S2: Subjects highly weighted on the SFG-MiFG-MFG component only.
  6. Sinter: Subjects highly weighted on both components. This step is crucial as it moves from identifying group-level brain patterns to identifying subgroups of subjects characterized by specific patterns of these components.

  7. Subtype-Specific Component Reconstruction: Group Information-Guided ICA (GIG-ICA) was used to reconstruct subtype-specific components. This ensures that the identified components are not only spatially distinct but also maintain independence at the subtype level.

  8. Biclustering: The algorithm effectively performs biclustering by simultaneously identifying clusters of subjects and clusters of voxels (brain regions) that exhibit correlated patterns of activity or structure. This is visualized as rearranging the subject-by-voxel matrix to reveal overlapping biclusters.

What's novel here is the application of biclustering directly to neuroimaging data for subtype discovery in Sz, and the approach of deriving subtypes from brain data before correlating them with clinical symptoms, rather than the other way around.

Datasets

The study utilized a large, aggregated, multi-site dataset of 382 patients with schizophrenia (Sz). This dataset was compiled from three independent studies. The data underwent preprocessing including normalization, segmentation, and smoothing. Age, gender, and site effects were regressed out from the voxelwise GMC data.

Key Insights and Ideas

  • Neurobiological Subtypes are Detectable: The study demonstrates that distinct neurobiological subtypes of schizophrenia can be identified directly from structural MRI data, independent of pre-defined clinical symptom categories.
  • Biclustering for Complex Biomarkers: B-ICA effectively identifies complex biomarkers by considering the joint contribution of multiple brain regions (components) and their association with specific subject groups. This moves beyond univariate analyses that might miss subtle, distributed abnormalities.
  • Symptom-Specific Subtypes: The identified subtypes show significant associations with clinical symptoms, particularly positive symptoms. This suggests that these neurobiological differences have functional consequences.
  • Personalized Treatment Potential: The discovery of distinct subtypes with varying patterns of brain abnormalities hints at the possibility of personalized treatment strategies tailored to these specific subgroups.
  • Data-Driven Approach is Key: The paper advocates for data-driven approaches that leverage the richness of neuroimaging data to understand the heterogeneity of complex disorders like schizophrenia, rather than relying solely on clinical classifications.

Key Results

  • Two Exclusive Subtypes and One Overlapping Subtype: The B-ICA algorithm identified three distinct subject groups:
  • S1 (65 subjects): Highly weighted on the I-STG-IFG component.
  • S2 (62 subjects): Highly weighted on the SFG-MiFG-MFG component.
  • Sinter (53 subjects): Highly weighted on both components.
  • Positive Symptom Differences: Subtypes S1 and S2 showed a significant difference in positive symptom severity (PANSS positive scores), with S2 exhibiting higher positive symptoms (p = 0.005).
  • Trend for General Symptoms: The Sinter group showed a trend-level correlation (r = 0.25, p = 0.07) between their loadings on the SFG-MiFG-MFG component and general clinical symptoms (PANSS general).
  • Structural Network Connectivity (SNC): The three subtypes exhibited varying strengths of connectivity between the I-STG-IFG and SFG-MiFG-MFG components, with Sinter showing the strongest connectivity (R(51) = 0.93, p < 0.00001).
  • Subtype-Specific Brain Patterns: Reconstructed subtype-specific components showed subtle variations in brain regions compared to the overall group components, highlighting the localized nature of these subtype differences.

Limitations and Future Work

  • Antipsychotic Medication: All Sz participants were on antipsychotic medication, and the study acknowledges that the availability of standardized chlorpromazine equivalents (Cpz eqvt) across all datasets would have strengthened the analysis. The amounts and history of dosage were not consistently recorded.
  • Cognitive and Electrophysiological Data: Cognitive symptom tests and electrophysiological data were not available for inclusion in this analysis, which the authors suggest would have been beneficial.
  • Clinician Rating Heterogeneity: While site effects on PANSS scores were controlled, potential heterogeneity due to differences in clinician ratings remains a limitation.
  • Newness of B-ICA: The authors note that biclustering for neuroimaging analyses is relatively new, suggesting that further research and validation are needed.
  • Replication: The authors emphasize the need for future work to replicate these findings.
  • Future Directions: The authors suggest that future work should focus on:
  • Replicating these results.
  • Investigating the potential of these subtypes as endophenotypes for Sz, which would require further studies on their relationships across diagnostic boundaries, within families, and over the course of the illness.
  • Exploring the use of RDoC (Research Domain Criteria) frameworks.
  • Clustering multiple components to identify brain regions relevant to psychotic symptoms.
  • Conducting more clinically oriented studies to explore personalized treatment and targeted drug development for these identified subtypes.

Summary updated

Abstract

Clinical and cognitive symptoms domain-based subtyping in schizophrenia (Sz) has been critiqued due to the lack of neurobiological correlates and heterogeneity in symptom scores. We, therefore, present a novel data-driven framework using biclustered independent component analysis to detect subtypes from the reliable and stable gray matter concentration (GMC) of patients with Sz. The developed methodology consists of the following steps: source-based morphometry (SBM) decomposition, selection and sorting of two component loadings, subtype component reconstruction using group information-guided ICA (GIG-ICA). This framework was applied to the top two group discriminative components namely the insula/superior temporal gyrus/inferior frontal gyrus (I-STG-IFG component) and the superior frontal gyrus/middle frontal gyrus/medial frontal gyrus (SFG-MiFG-MFG component) from our previous SBM study, which showed diagnostic group difference and had the highest effect sizes. The aggregated multisite dataset consisted of 382 patients with Sz regressed of age, gender, and site voxelwise. We observed two subtypes (i.e., two different subsets of subjects) each heavily weighted on these two components, respectively. These subsets of subjects were characterized by significant differences in positive and negative syndrome scale (PANSS) positive clinical symptoms (<i>p</i> = 0.005). We also observed an overlapping subtype weighing heavily on both of these components. The PANSS general clinical symptom of this subtype was trend level correlated with the loading coefficients of the SFG-MiFG-MFG component (<i>r</i> = 0.25; <i>p</i> = 0.07). The reconstructed subtype-specific component using GIG-ICA showed variations in voxel regions, when compared to the group component. We observed deviations from mean GMC along with conjunction of features from two components characterizing each deciphered subtype. These inherent variations in GMC among patients with Sz could possibly indicate the need for personalized treatment and targeted drug development.
BibTeX
@inproceedings{gupta2017biclustered,
  author = {Cota Navin Gupta and Eduardo Castro and Srinivas Rachkonda and Theo G.M. van Erp and Steven G. Potkin and Judith M. Ford and Daniel H. Mathalon and Hyo Jong Lee and Bryon A. Mueller and Douglas N. Greve and Ole A. Andreassen and Ingrid Agartz and Andrew R. Mayer and Julia M. Stephen and Rex E. Jung and Juan Bustillo and Vince D. Calhoun and Jessica A. Turner},
  title = {{Biclustered Independent Component Analysis for Complex Biomarker and Subtype Identification from Structural Magnetic Resonance Images in Schizophrenia}},
  booktitle = {Frontiers in Psychiatry},
  year = {2017},
  doi = {10.3389/fpsyt.2017.00179},
}
Figure