103.txt

MACHINE LEARNING APPROACH FOR
CHILDREN DISORDER UNDERSTANDING
By
Mariam Mostafa Mahmoud Hassan
A Thesis Submitted to the
Faculty of Computers and Artificial Intelligence
Cairo University
in Partial Fulfillment of the
Requirements for the Degree of
MASTER
in
Information System
Under the Supervision of
Prof. Hoda Mokhtar Omar Mokhtar
Department of Information System
Faculty of Computers and Artificial Intelligence, Cairo University

Abstract
Autism spectrum disorder (ASD) is a neurodevelopmental disorder that causes
deficits in cognition, communication and social skills. ASD, however, is a highly
heterogeneous disorder. This heterogeneity has made identifying the etiology of ASD
a particularly difficult challenge, as patients exhibit a wide spectrum of symptoms
without any unifying genetic or environmental factors to account for the disorder.
For better understanding of ASD, it is paramount to identify potential genetic and
environmental risk factors that are comorbid with it. Identifying such factors is of
great importance to determine potential causes for the disorder, and understand its
heterogeneity. Existing large-scale datasets offer an opportunity for computer
scientists to undertake this task by utilizing machine learning to reliably and
efficiently obtain insights about potential ASD risk factors, which would in turn
assist in guiding research in the field. Moreover, the large sample size of these
datasets helps to avoid the pitfalls and biased results associated with studying a
heterogeneous disorder through a small sample size. The purpose of this research is
to address these issues. In this research work, the main contribution is to investigate
the effect of different risk factors, including patient medical history and Family
Medical History, in the Autism domain. In addition, it is to examine differences
in gender, as the disparity in ASD diagnosis between males and females suggest
differences in underlying etiology. Because the Autism research has demonstrated
remarkable growth in the last 10 to 15 years, far outpacing research growth in
other comparable fields, thus planning and investment in ASD research should be
guided by clear objectives that account for the current state of the science and
the identification of critical research gaps, that’s why building Autism ontology
that represents the knowledge in Autism Spectrum Disorders in different aspects
including results from this study can help in clinical decision support and annotation
process to assist in guiding research by shedding light on gaps and opportunities.
The following methodology has been applied; composed of data capturing, data
pre-processing, machine learning implementation and building Autism ontology.
Data capturing include querying National Database for Autism Research (NDAR)
for Subject and Family Medical History where subjects are previously diagnosed
using trusted Autism diagnostic instrument, then using these datasets towards
building a database schema to facilitate mapping records and generate new datasets.
Data pre-processing handles the generated raw data to be more effectively and easily
processed by machine learning algorithms, which include data cleaning, unifying
terminologies and handling missing values using various python functions. Scikit-
learn a python machine learning library was used for constructing features and
patterns using DecisionTreeClassifier function. The algorithm performance was
measured using numerous matrices and measures. Results have been evaluated using
close datasets that covered similar medical conditions to the ones covered in the
datasets used in this study. To further validate the ASD associations obtained in
this study, 10 datasets were randomly generated from Family and Subject Medical
History datasets, with a ratio of 1:1 ASD and non-ASD. This subsampling technique
strategy was utilized to ameliorate the problem of significantly unbalanced classes.
xviResults were finally validated using medial studies. Finally, building the Autism
ontology to store the constructed features and patterns in a machine understandable
format to enable reusing it in decision support and annotation.
The results of the applied methodology identified eleven diagnosis rules and ten
medical conditions out of sixty nine used in the analysis that were highly associated
with ASD diagnoses in patients; furthermore, the analysis was extended to the
Family Medical History of patients and reported twelve rule sets and ten potentially
hereditary medical conditions from eighty one input conditions associated with
ASD. Associations reported had a 90% accuracy. Meanwhile, gender comparisons
highlighted conditions that were unique to each gender and others that overlapped.
Those findings were validated by medical studies in the Autism domain, thus opening
the way for new directions for the use of machine learning algorithms to further
understand the etiology of Autism. The second contribution in this study was
developing a domain Autism ontology that formally conceptualizes both domain
and clinical Autism knowledge, and facilitates access to precise Autism knowledge
for both decision support and annotation purposes. To construct and build such
ontology, many clinical and medical research works have been investigated in various
Autism areas. The ontology was build using existing ontologies, relevant articles,
standard textbooks, and clinical studies. Such sources were used to link key concepts
and build a semantic map among classes. Focusing on classes, properties between
classes and properties between classes and their instances. Ontology sub classes were
extended based on clinical cases and pilot studies. The ontology main contribution
in this work, is enable reusing knowledge extracted from classification approaches
in clinical decision support systems. The Autism ontology was implemented using
Protégé, an ontological framework, formatted by Web Ontology Language (OWL)
and Semantic Web Rule Language (SWRL). The ontology bring an opportunity
for researchers to have a panoramic view of Autism domain topics, assign their
knowledge into the ontology, compare it with other related work done in the same
topic and finally reasoning it over class individuals. Such knowledge can be used
and evaluated in clinical decision systems.
The experimental results show the effectiveness of the proposed approach. ASD
related factors and gender differences identified in this study can shed light on
areas for future research and investigation concerning the etiology of ASD, and
potential venues for prevention and treatment. This knowledge is stored in a machine
understandable format to enable reusing it in decision support and annotation. Using
ontology classes and SWRL rules provide numerous benefits to the domain users as
human or machine. The Autism ontology in this study will not only store constructed
patterns and rules but will help in conceptualization of knowledge, unification of
terminology, easy access to information, while assisting in expert decision making.
xvii