Content uploaded by Jeremy Cohen
Author content
All content in this area was uploaded by Jeremy Cohen on May 04, 2017
Content may be subject to copyright.
Using Patterns of Functional Brain Connectivity to
Classify Autism Spectrum Disorder
Hakeem A. Brooks1, Jin Hyun Cheong2, Jeremy D. Cohen, Ph.D.1, Luke J. Chang, Ph.D.2
1Xavier University of Louisiana, Department of Psychology, New Orleans, LA, USA
2Computational Social Affective Neuroscience Laboratory, Department of Psychological and Brain Sciences, Dartmouth College, Hanover, New Hampshire, USA
• The strong positive correlations of the tempo parietal junction (TPJ),
fusiform face area (FFA), and ventromedial prefrontal cortex (vmPFC)
with predicting autism suggest that these regions have abnormal
connections in patients with ASD.
• The TPJ is known to have implications in representing others’ mental
states. Defections in this functional connectivity network would relate to
the disruption of nonverbal communications.
• The FFA is involved in processing faces and the vmPFC is implicated in
representing value, both of which have been shown to be impaired with
the ASD phenotype.
• These findings are consistent with previous studies that have identified
differences in ASD connectivity patterns.
• Future studies could investigate whether the inclusion of more robust
phenotypic information has an effect on classifier accuracy.
• Autism Spectrum Disorder (ASD) is a neurodevelopmental disorder
characterized by verbal and nonverbal communication deficits and
repetitive behaviors.
• Approximately 1 in 68 children have ASD.
• No reliable neurobiological biomarkers have been identified to date.
• One possible reason for this is the use of modest sample sizes,
which present issues of overfitting.
• Another possible issue is the employment of more features than
samples when building the model, which also introduces overfitting
and excessive noise variance into the classifier model.
• Here, we apply a multivariate statistical learning model to a large
open-access resting-state functional magnetic resonance imaging
dataset to elucidate the functional connections that are unique to
ASD.
Acknowledgements:
This study was supported the NIH BUILD grant TL4GM118968.
Participants
• We used data provided by the ABIDE Connectomes Project
• 886 subjects (478 typically developing (TD) and 408 ASD) were
included in this study
Image Processing
• Images were processed along the Configurable Pipeline for the
Analysis of Connectomes (C-PAC) pipeline. This included motion
correction, normalization, smoothing, applying global signal
regression and applying a band pass filter (0.01 - 0.1 Hz) to the
data, and removing variance associated with nuisance variables.
Data Analysis/Building the classifier
• Average time signals were extracted from all subjects using a meta-
analytic parcellation scheme.
• Correlation coefficients (CCs) were calculated between ROI in each
subject.
• A logistic regression with least absolute shrinkage selection operator
(LASSO) penalty was trained to classify ASD from TD.
• Features were standardized before performing the regression.
• Cross-validation of the model was done using a nested stratified k-
folds procedure.
• Model weights revealed which connections most strongly contributed
to classification.
Methods
Discussion
Parcellation regions (k=50). Brain regions that are most frequently reported
together in a database of 12,000+ published papers using k-means clustering.
Figure 4. Model Weights on Brain Region Connectivity
Average activation in each region (6 & 32) was extracted and functional connectivity
between each region over time was calculated using a pearson correlation.
Figure 3. Classifier Model Accuracy
A heat map depicts the
connection weights for all 50
ROIs retained by the LASSO
model Red squares indicate
positive connections. Blue
squares indicate negative
connections
Figure 2. Functional Connectivity between Regions
Figure 1. Brain Parcellation
References
• Nielsen, Jared A., Brandon A. Zielinski, P. Thomas Fletcher, Andrew L. Alexander, Nicholas Lange, Erin D. Bigler, Janet E. Lainhart, and Jeffrey S.
Anderson. 2013. “Multisite Functional Connectivity MRI Classification of Autism: ABIDE Results.” Frontiers in Human Neuroscience 7 (September): 599.
• Yahata, Noriaki, Jun Morimoto, Ryuichiro Hashimoto, Giuseppe Lisi, Kazuhisa Shibata, Yuki Kawakubo, Hitoshi Kuwabara, et al. 2016. “A Small Number of
Abnormal Brain Connections Predicts Adult Autism Spectrum Disorder.” Nature Communications 7 (April): 11254.
• Plitt, Mark, Kelly Anne Barnes, and Alex Martin. 2015. “Functional Connectivity Classification of Autism Identifies Highly Predictive Brain Features but Falls
Short of Biomarker Standards.” NeuroImage. Clinical 7: 359–66.
This figure shows the receiver
operator characteristic curve
which depicts the overall
model performance.
• Accuracy = 0.68.
• Sensitivity = .63
• Specificity = .64
Introduction
Sparse
Logistic
Model
ASD
TD
Feature 1
Feature 2
Feature 3
Diagram of Classifier Model
Figure 5. Brain Connectivity Predictive of ASD
Results
