A Data-Driven Clustering of Subjects Based on Brain Activation Reveals Phenotypic Group Differences

Conference Year

January 2019

Abstract

Introduction: Traditional approaches to understanding the neural basis of neuropsychiatric disease, substance use and cognitive processes focus on between group analysis with groups determined by phenotypic criteria (e.g., symptom profiles, diagnoses and levels of substance use). This approach has revealed activation differences with effect sizes that are typically small and difficult to replicate. Data driven approaches to identifying individual differences in patterns of brain activity during cognitive tasks may have more power to detect neural differences that are clinically relevant. The objective of this study was to explore the utility of Data Spectroscopic Clustering (DSC) in identifying groups based on patterns of BOLD fMRI signal during the Stop Signal Task (SST) and to determine if these groups have distinct behavioral phenotypes.

Methods Analysis was conducted on a subset of the IMAGEN study, a multi-site longitudinal study of neurodevelopment in adolescents. Data analysis focused on subjects, ages 19-to-20-years old, who had complete mental health and substance use information (n=726; 374 males). We employed DSC to group subjects based on patterns of BOLD response during successful stop trials in the SST. Beta weights from the stop success contrast were extracted using the Shen-268 atlas and then DSC was applied. Resultant groups were examined for differences on mental health and substance use assessments (the Development and Well-Being Assessment (DAWBA) and the European School Survey Project on Alcohol and Drugs (EPSAD) questionnaires, respectively). ANOVAs were used to determine if the groups differed on substance use or psychiatric symptoms.

Results The DSC identified six groups of subjects (Table 1) defined by similarities in task activation during the SST. The groups did not differ by sex, handedness or behavioral performance on the Stop Signal Task. Although group membership was significantly related to site (p = .032), all sites had representation in all groups. Interestingly, one of the groups, group 2, had significantly higher binge drinking behavior in the 30 days preceding their study visit compared to group 1 (F (1, 720) = 21.7, pp2 = .029), group 3 (F (1, 720) = 8.73, p = .003, ηp2 = .012), group 4 (F (1, 720) = 15.1, p p2 = .021), group 5 (F (1, 720) = 16.3, p p2 = .022) and group 6 (F (1, 720), pp2 = .027). A subset of participants had available Strength and Difficulties Questionnaire (SDQ) domain scores (n=1005). Some of the groups differed significantly by self-reported conduct symptoms. Levene’s test indicated unequal variances F (5, 1005) = 2.40, p = .04 so degrees of freedom were adjusted from 1005 to 195.72 and a brown-forsythe ANOVA was conducted F(5, 195.72) = 4.81 p

Conclusions Using DSC to identify groups of subjects with similar patterns of task-related brain activation may be useful for identifying groups that differ on clinically relevant variables. The next step in this approach focusses on investigating why the phenotypic characteristics of a group are related to that group’s particular pattern of brain activation. Longitudinal studies will be important to determine the predictive utility of grouping subjects based on patterns of neural activity during cognitive tasks.

Primary Faculty Mentor Name

hgaravan

Secondary Mentor Name

Nicholas Allgaier, PhD

Faculty/Staff Collaborators

Nicholas Allgaier, PhD (Post Doc), Philip Spechler (Graduate Student), Nicholas Fontaine (Undergraduate Research Assistant), Shana Adise, PhD (Post-Doc)

Status

Graduate

Student College

College of Agriculture and Life Sciences

Program/Major

Psychological Science

Primary Research Category

Social Sciences

Secondary Research Category

Biological Sciences

Abstract only.

Share

COinS
 

A Data-Driven Clustering of Subjects Based on Brain Activation Reveals Phenotypic Group Differences

Introduction: Traditional approaches to understanding the neural basis of neuropsychiatric disease, substance use and cognitive processes focus on between group analysis with groups determined by phenotypic criteria (e.g., symptom profiles, diagnoses and levels of substance use). This approach has revealed activation differences with effect sizes that are typically small and difficult to replicate. Data driven approaches to identifying individual differences in patterns of brain activity during cognitive tasks may have more power to detect neural differences that are clinically relevant. The objective of this study was to explore the utility of Data Spectroscopic Clustering (DSC) in identifying groups based on patterns of BOLD fMRI signal during the Stop Signal Task (SST) and to determine if these groups have distinct behavioral phenotypes.

Methods Analysis was conducted on a subset of the IMAGEN study, a multi-site longitudinal study of neurodevelopment in adolescents. Data analysis focused on subjects, ages 19-to-20-years old, who had complete mental health and substance use information (n=726; 374 males). We employed DSC to group subjects based on patterns of BOLD response during successful stop trials in the SST. Beta weights from the stop success contrast were extracted using the Shen-268 atlas and then DSC was applied. Resultant groups were examined for differences on mental health and substance use assessments (the Development and Well-Being Assessment (DAWBA) and the European School Survey Project on Alcohol and Drugs (EPSAD) questionnaires, respectively). ANOVAs were used to determine if the groups differed on substance use or psychiatric symptoms.

Results The DSC identified six groups of subjects (Table 1) defined by similarities in task activation during the SST. The groups did not differ by sex, handedness or behavioral performance on the Stop Signal Task. Although group membership was significantly related to site (p = .032), all sites had representation in all groups. Interestingly, one of the groups, group 2, had significantly higher binge drinking behavior in the 30 days preceding their study visit compared to group 1 (F (1, 720) = 21.7, pp2 = .029), group 3 (F (1, 720) = 8.73, p = .003, ηp2 = .012), group 4 (F (1, 720) = 15.1, p p2 = .021), group 5 (F (1, 720) = 16.3, p p2 = .022) and group 6 (F (1, 720), pp2 = .027). A subset of participants had available Strength and Difficulties Questionnaire (SDQ) domain scores (n=1005). Some of the groups differed significantly by self-reported conduct symptoms. Levene’s test indicated unequal variances F (5, 1005) = 2.40, p = .04 so degrees of freedom were adjusted from 1005 to 195.72 and a brown-forsythe ANOVA was conducted F(5, 195.72) = 4.81 p

Conclusions Using DSC to identify groups of subjects with similar patterns of task-related brain activation may be useful for identifying groups that differ on clinically relevant variables. The next step in this approach focusses on investigating why the phenotypic characteristics of a group are related to that group’s particular pattern of brain activation. Longitudinal studies will be important to determine the predictive utility of grouping subjects based on patterns of neural activity during cognitive tasks.