Applying Machine Learning Techniques to Brain Imaging Characteristics to. Distinguish Between Individuals with Autism and Neurotypical Controls. Sarah E.
Applying Machine Learning Techniques to Brain Imaging Characteristics to Distinguish Between Individuals with Autism and Neurotypical Controls Sarah E. Schipul, Sandesh Aryal and Marcel A. Just Center for Cognitive Brain Imaging, Carnegie Mellon University Can machine learning algorithms distinguish between autism and neurotypical individuals based on their brain characteristics? Machine learning algorithms were run on feature sets including measurements of brain structure, activation, and synchronization from two neuroimaging experiments. Multiple classification algorithms were used to predict whether a given participant was in the autism group or the control group.
Introduction Currently, autism is diagnosed based solely on interviews and evaluation by trained clinicians.
Experimental Paradigm – Study 2
Selected Feature Sets for Each Study
Study 2
Task:
Study 1 used 20 features:
All of the classifiers were able to predict a diagnosis of autism with accuracy well above chance (0.59 for 5% cutoff accuracy) as seen in the graph and table below.
Watch a pair of videos and choose the liar.
3 voxel counts during fixation
In each pair, one video displayed truthful cues and the other displayed lying cues.
1 functional connectivity pair during the low condition
Conditions Pretest: first 12 video pairs, no feedback
43 Individuals with high functioning autism
12 functional connectivity pairs during the posttest condition The classifier model we want to train is a set of functions f of the form f:structAndFunctFeats → Yi ; i = 1:2 where Yi ={autism, control} where structAndFunctFeats are the set of structural and functional data features from fMRI scanning sessions.
Data Analysis
Cross-Validation
The data were analyzed using SPM2. Statistical analysis was
The classifier was tested on one subject from each group. This procedure was reiterated for all possible combinations (folds) of leaving one out from each group.
Scanner Procedure
In both studies, participants with autism met diagnostic criteria on the Autism Diagnostic Interview-Revised, on the Autism Diagnostic Observation Schedule, and by expert clinical judgment.
Participants were run on a 3.0T Siemens Allegra scanner using a CP transmit/receive head coil at the Brain Imaging Research Center, University of Pittsburgh/Carnegie Mellon University.
Experimental Paradigm – Study 1
A 160-slice axial 3D MPRAGE volume scan was acquired for each participant to be used in segmenting the corpus callosum into anatomically predefined regions.
Task: Read a sentence and indicate whether it is true or false. Conditions Fixation Low imagery: sentences contain little to no imagery
Features Used in the Machine Learning Analyses
High imagery: sentences contain high imagery
Stimuli – Study 1
A 3 turned backwards looks like a capital letter e.
Participants – Study 2 18 Individuals with high functioning autism Means: Age 22.4; FSIQ 107.6; 17 M: 1 F 18 matched neurotypical participants Means: Age 22.4, FSIQ 111.2; 17 M: 1 F
Diffusion-weighted data were acquired with diffusion-weighting gradients applied in six orthogonal directions. Diffusion tensors for each voxel were calculated and reduced to fractional anisotropy maps.
True
For each dataset, we chose the 18-20 most distinguishing features out of the following measures: Structural measures: Volume measurements of white and grey matter Area measurements of corpus callosum segments Diffusion tensor imaging fractional anisotropy measurements Activation measures: Activated voxel counts for all conditions in a set of ROIs Synchronization measures: ROI-pair synchronization measures for all conditions
Results
0.4 0.2 0.0 3 Nearest Neighbors
Logistic Regression
Support Vector Machine
Gaussian Naïve Bayes
Classifiers
Classifiers Test Accuracy Train Accuracy 3 Nearest Neighbors 0.917 0.979 Logistic Regression 0.889 0.966 Support Vector Machine 0.861 0.985 Gaussian Naïve Bayes 0.833 0.858
Conclusions
Study 1 All of the classifiers were able to predict a diagnosis of autism with accuracy above chance (0.59 for 5% cutoff accuracy) as seen in the graph and table below.
Several machine learning algorithms were able to reliably distinguish between autism and control participants based on brain structure, activation, and synchronization in two distinct datasets. This preliminary investigation suggests that functional and structural brain imaging data have potential to play a role in diagnosing autism.
Classification Accuracy: Study 1 1.0
Furthermore, the predictive potential of measures of functional and structural brain connectivity provides support for the underconnectivity theory of autism (Just et al., 2007).
0.8
Accuracy
Means: Age 24.9, FSIQ 112.7; 41 M: 2 F
0.6
Classification algorithms included Logistic Regression, Support Vector Machine, Gaussian Naïve Bayes, and Nearest Neighbor.
correlation between the average time courses of all voxels in each member of a pair of functional ROIs defined by activation at the group level.
43 neurotypical participants
False
Classification Model
Functional connectivity was computed for each participant as a
Means: Age 23.3; FSIQ 101.6; 38 M: 5 F
0.8
Posttest: last 12 video pairs, no feedback
performed on individual and group data using the general linear model (Friston et al., 1995).
Participants – Study 1
Study 2 used 18 features: 6 diffusion tensor imaging fractional anisotropy measurements
Previous studies have explored the potential of structural brain imaging data to predict a diagnosis of autism (Ecker et al., 2009; Fahmi et al., 2007).
Methods
1.0
1 brain volume measurement
Training Session: middle 12 video pairs, with feedback
Video Stimuli – Study 2
Classification Accuracy: Study 2
1 corpus callosum measurement
Fixation
However, brain imaging studies have revealed that individuals with autism show unique characteristics in brain structure and function.
This study explores the potential of both structural and functional brain imaging data to predict a diagnosis of autism.
14 functional connectivity pairs during the high condition
Accuracy
Overview
0.6
References
0.4 0.2 0.0 5 Nearest Neighbors
Support Vector Machine
Gaussian Naïve Bayes
Logistic Regression
Classifiers
Classifiers Test Accuracy Train Accuracy 5 Nearest Neighbors 0.698 0.795 Support Vector Machine 0.686 0.851 Gaussian Naïve Bayes 0.663 0.671 Logistic Regression 0.663 0.729
Ecker C, Rocha-Rego V, Johnston P, Mourao-Miranda J, Marquand A, Daly EM, Brammer MJ, Murphy C, and Murphy DG. (2009). Investigating the predictive value of whole-brain structural MR scans in autism: A pattern classification approach. NeuroImage, 49(1):44-56. Fahmia R, Elbazb A, Hassana H, Faraga AA, Casanova MF. (2007). Structural MRI-Based Discrimination between Autistic and Typically Developing Brain. Proc. of Computer Assisted Radiology and Surgery (CARS'07), Berlin, Germany, pp. 6-8. Just MA, Cherkassky VL, Keller TA, Kana RK, Minshew NJ. (2007). Functional and anatomical cortical underconnectivity in autism: Evidence from an fMRI study of an executive function task and corpus callosum morphometry. Cerebral Cortex, 17, 951-961. Conflict of Interest: None