Classification of Autism Spectrum

Disorder using Machine Learning

Utkarsh Patel
Third-year Undergrad
Department of E&ECE, IIT Kharagpur

Mentor
Debasis Samanta
Associate Professor
Department of Computer Science, IIT Kharagpur
Abstract
• In this study, we implemented a data-driven approach to classify ASD
patients and typically developing (TD) participants by using resting-
state fMRI data.
• SVM and KNN were used for classification purpose.
• Neural Networks were also used (just for comparison)
• Each classifier was fine-tuned via grid search on hyper-parameters via
cross-validation.
• Finally, a stacked model was also used comprising of fine-tuned
classifier as its base models.
What is Autism Spectrum Disorder?
Autism Spectrum Disorder
• It is a complex developmental condition that involves persistent
challenges in social interaction, speech and nonverbal
communication, and restricted/repetitive behaviors.
• One in 59 children is estimated to have autism in United States.
Why Machine Learning?
Autism Spectrum Disorder
• There is no medical test for autism.
• Early diagnosis and treatment are important to reducing the
symptoms of autism and improving the quality of life for people with
autism and their families.
• This procedure usually takes a lot of time like 5-10 years.
• Hence, the objective is to use machine learning algorithms for
classification.
Why Functional Connectivity Analysis important?
Role of Functional Connectivity
• Past extensive brain imaging studies have reported that ASD is
associated with brain connectivity.
• It has been found that in case of autism, some brain regions have
weaker functional connectivity (than normal) which may be
responsible for their social behaviour, while some of them are highly
correlated (than normal) which may account for their exceptional
ability to concentrate.
• Studies have found that most of the gold medallists in IMO, IOI, IPhO,
etc. have autism.
The Dataset
ABIDE Dataset
• For this task, we used ABIDE dataset
• It contains aggregated functional and structural brain imaging data
collected from laboratories around the world to accelerate our
understanding of the neural bases of autism.
• The ABIDE 1 dataset contains fMRI images of 1112 subjects, 539 from
individuals with ASD, and 573 from typical controls.
Pre-processing
Pre-processing
• For this task, I used the data provided by NYU only. Reasons being:
• Data provided from different laboratories have different configuration for
their instruments, way of examining patients, etc.
• Of all the sources, NYU had maximum number of subjects (172)
• For pre-processing, I used CPAC pipeline (as the task involves
functional connectivity analysis).
nilearn
• nilearn is the library for doing ‘Statistics for Neuroimaging in Python’
• Uses sklearn as backend
• This library is used in this project for:
• Extracting pre-processed ABIDE 1 dataset
• Building connectivity matrices from the time-series signals obtained from
fMRI images
Brain MAP / Atlas
Brain Atlas
• As per Wikipedia, a brain atlas is composed of serial sections along
different anatomical planes of the healthy or diseased developing or
adult animal or human brain where each relevant brain structure is
assigned a number of coordinates to define its outline or volume.
• Brain atlases are contiguous, comprehensive results of visual brain
mapping and may include anatomical, genetical or functional
features.
• In this project, I used AAL atlas, which has 116 ROIs (functional
clusters).
Extracting data and building
connectivity matrices
Snippet
import pandas as pd
import numpy as np
import matplotlib.pyplot as plt
import nilearn

from nilearn.datasets import fetch_abide_pcp

from nilearn.connectome import ConnectivityMeasure, sym_to_vec

data = fetch_abide_pcp(derivatives=['rois_aal'], SITE_ID=['NYU’])

conn_est = ConnectivityMeasure(kind='correlation')
conn_matrices = conn_est.fit_transform(data['rois_aal'])
Connectivity Matrix b/w ROIs given by
correlation of its time-series signal
Connectivity Matrix b/w ROIs given by
correlation of its time-series signal
This matrix is then flattened to a vector and is used in training the model
Model Selection
Model Selection
• After flattening the connectivity matrix, we have feature vectors
𝑥𝑖 , 𝑖 = 1, 2, … , 172 and each 𝑥𝑖 is a 6786 dimensional vector.
• We cannot use dimensionality reduction because every correlation is
important for classification. (Tried it, however doesn’t increase
accuracy)
• Tried various algorithm, however only Support Vector Machines and K
Nearest Neighbours classifier gave satisfactory results.
• Later tuned the SVM and KNN classifier.
Classification using Support Vector
Machines
Support Vector Machines
• Support Vector Machines are a class of supervised learning, in which
a hyperplane which produces maximum margin between the class is
used for classification.
• It has various hyperparameter:
• Kernel – linear, radial, sigmoid, polynomial
• Regularization factor C
• Fitting parameter 𝛾
• A careful tuning is required to get best results
Tuning SVM
• Tuning regularization constant C
• For this, we varied C from 2−5 to 215
• Tuning fitting parameter
• For this, we varied 𝛾 from 2−15 to 23

• Basically, a grid search is performed that give us the best possible

accuracy
Linear kernel
Linear Kernel
• For linear kernel, the 5-fold cross-validation accuracy was found to be
59.92% for all values of 𝐶 and 𝛾
Radial kernel
Radial Kernel
• For radial kernel, the maximum 5-fold cross-validation was found to
be 67.97% for 𝛾 = 0.000122 and 𝐶 ≥ 4.
Sigmoid kernel
Sigmoid kernel
• The maximum 5-fold cross-validation accuracy was found to be
61.04% for 𝛾 = 0.000122 and 𝐶 = 2.
Quadratic kernel
Quadratic kernel
• The 5-fold cross-validation accuracy of 63.41% was observed for
values of 𝛾 and 𝐶 satisfying

2𝑙𝑜𝑔2 𝛾 + 𝑙𝑜𝑔2 𝐶 + 25 = 0 → 𝐶𝛾 2 = 2−25

Summary for SVM
Summary for SVM
Kernel Best Accuracy C 𝜸
Linear 59.92 ℝ ℝ
Radial 67.97 ≥ 21 2−13
Sigmoid 61.04 21 2−13
Quadratic 63.41 23 2−14
 Classification using
K Nearest Neighbours
K Nearest Neighbours
• In KNN, we have two hyper-parameters
• Value of 𝑘
• Distance function
• sklearn uses Minkowski metric for distance function, which has a
hyper-parameter 𝑝
1
𝑝
𝑝
𝑑𝑝 𝒙, 𝒚 = ෍ 𝑥𝑖 − 𝑦𝑖
𝑖
• 𝑝 = 1 corresponds to Manhattan distance, while 𝑝 = 2 corresponds
to Euclidean distance.
Tuning KNN classifier
• KNN classifier is tuned by varying:
• Value of 𝑘 from 1 to 50
• Value of 𝑝 in 1, 2
Accuracy observed for KNN
Optimal Hyper-parameters
𝑘 = 29, 𝑝 = 1

𝒑 Best Accuracy 𝒌
1 64.55 29
2 63.99 23
SVM vs KNN
SVM vs KNN
• The best accuracy observed for SVM is 67.97 for radial kernel.
• The best accuracy observed for KNN is 64.55 for k = 29, p = 1.

• Hence, SVM outperforms KNN over NYU data set.

 SVM + KNN
Stacked Model
Stacked Model
• In stacked ensemble learning, we have
• Base models (level 0)
• Meta models (level 1)
• I used the fine-tuned SVM, KNN and logistic regression model as base
models
• Logistic regression was also used in meta model
Stacked Model Architecture

LR
SVM
Base Models
KNN

Meta Model
LR
Performance of Stacked model

Standalone fine-tuned
SVM classifier
outperforms the
stacked model
Classification using Neural Networks
Neural Networks
• The best ML classifier of the task is SVM giving accuracy as high as
67.97.
• Let’s see whether Neural Networks can outperforms the SVM or not.
Tuning Neural Networks
• Neural Networks have a large number of hyper-parameters
• # of hidden layers
• # of nodes in each hidden layer
• Non-linear activation used in hidden layers
• Learning rate
• Optimization algorithms (Adam, Momentum, RMSProp, Adagrad)
• Regularization algorithms (Dropout, L2-regularization)
• .
• .
• .
Tuning Neural Networks
• For this task, I am using sklearn’s MLP API.
• Some background information:
• It uses ‘Adam’ optimizer by default.
• It uses SGD for optimization.
• We can vary the (initial) learning rate and the architecture
• I trained the model for 3000 epochs.
Performance of NN

We can observe that model with zero

hidden layer (a logistic regression) is
performing far-better as compared to
complex models over a wide range of
learning rates. However, the peak
accuracy is 63.75 which is less than
that observed for SVM (67.97)
Conclusion
Conclusion
• It is clear than a standalone fine-tuned SVM classifier is
outperforming stacked model as well as the neural networks for this
task.
• However, we may be able to further push the performance of the
classifier if we can figure out a new way to extract functional
connectivity instead of relying on correlation-based approach.
Future Work
• Apply new ways to extract functional connectivity.
• Check the proposed algorithm with different people according to
their gender and age.
Thank You.