2017 International Conference on Information Technology

IMTC 2005 – Instrumentation and Measurement

Technology Conference
Ottawa, Canada, 17-19 May 2005

A Machine Learning-based Method for Autism Diagnosis Assistance

in Children
Sushama Rani Dutta1, Soumyajit Giri2, Sujoy Datta3, Monideepa Roy4
School of Computer Engineering,
KIIT University,
Bhubaneswar, Odisha, India

Abstract –Predicting the onset of Autism from its early symptoms is

a difficult proposition. In this paper we have proposed a machine
particular symptom, but if we build the symptom set by going
learning algorithm for assisting the doctors in the diagnosis of the through a step by step symptom testing along with
disease in affected individuals. The Machine learning Association confirmation, then it can lead to a sufficiently reliable disease
Rule (AR), along with the minimum Redundancy-Maximum- diagnosis. The main challenge here is to find the most
Relevance (mRMR) method, is used to pull the symptoms which need probable associated symptoms set for a particular disease that
to be further tested for the prediction process, from a prior needs to be examined further. Here we have used a machine
database. We have used the Mutual Information Difference (MID) learning based Association rule (AR), minimum Redundancy
method for selecting the additional symptoms which can strengthen Maximum Relevant (mRMR), Mutual Information Difference
the symptom set, through the information of previously cases. The
(MID), and Highest Information Gain (HIG) to identify the
proposed system needs one or two primary symptoms as preliminary
inputs (measured through sensors or through mobile applications)
most appropriate associated symptom to be tested along with
and then it will automatically pull the appropriate associated the preliminary symptom. The tests and algorithms have been
symptoms from the domain knowledge and the previous patterns of a implemented for the autism disease. Children with the
patient’s database. The symptoms are further tested and confirmed. Autism Spectrum Disorder (ASD) mostly face many
We have tested this process by collecting a dataset of 500 autism difficulties while communicating during social interaction.
disease patients. Our experimental results show that the average This is due to abnormal sensory integration. It is estimated by
accuracy of correctly predicting the diagnoses for the various types the ‘Centre for Disease Control’ that more than 10% of the
of autism, by using our technique, is around 83%. Indian children population is affected by the autism spectrum
disorder (ASD). Autism is a neuro-developmental disorder
that generally starts in the childhood. We have observed that
Keywords – Machine learning; minimum Redundancy Maximum Autistic children in the villages are misguided because of
Relevant; Association Rule; Autism prediction; Mutual Information
improper treatment and poor diagnosis. We feel that this
Difference; Electronic Health Record
system will help doctors and the medical assistants to identify
the disease more easily. The system needs any one abnormal
I. INTRODUCTION symptom of the patient as a preliminary input. The patient
symptoms are measured through sensors or taken input
through smart phone and the data is integrated in a mobile
device and further sending to a server for diagnosis. The
An examination of preliminary symptoms is usually not framework can pull other symptoms from the domain
sufficient for a good prediction of the onset of autism in knowledge database automatically to decide which class of
children. Preliminary symptoms indicate different diagnoses, disease it belongs to. The data stored in the database is
which can be then narrowed down to the most probable designed based on the domain experts’ knowledge with the
disease and the initial treatment maybe initiated. An experiences and the previous patient’s symptoms sets. We
Electronic Health Record (EHR) system storing the previous have tested the framework on a dataset of suspected autistic
symptom-patterns of existing patients, needs to be available patients to identify the type of autism in a child, collected
for implementing our proposed algorithms. In this paper, we from the Centre for Autism Therapy Counselling and Help
propose a framework which will help to enhance the (CATCH, Bhubaneswar, India) to predict the onset of autism
diagnosis system by pulling the most appropriate additional in children. The rest of the paper is organised as follows:
symptoms for a disease. We have used the EHR along with Section II describes the related work. Section III explains
the database of the domain experts’ decisions for the proposed framework with the different stages of disease
symptom and disease correlation. Using them, we can diagnosis procedure starting with preliminary symptom.
observe the possible chances of particular diseases based on Section IV discusses experimental results and Section V
the symptoms. There may be various reasons for one discusses conclusion and future work.

978-1-5386-2924-6/17 $31.00 © 2017 IEEE 36

DOI 10.1109/ICIT.2017.26
 II. RELATED WORK patient is suffering from them or not. If there are only one or
two symptoms in the initial symptom set, then this method
Human life expectancy has been increased by the modern
needs to collect other possible additional symptoms. It can
healthcare system. Advanced healthcare nurtures the mental
then form a probable symptom set for the easy diagnosis of
and physical conditions of the population [1]. Automation in
the disease. The finding of associate symptoms is done
healthcare using alert systems through WBANs is now
though the association rule of machine learning.
helping in improving the quality of services provided. In [2]
an automation system has been developed, which uses
different types of sensors for the monitoring of fitness, by
hypertension monitoring and sending SMSs to the care
provider depending upon the emergency of the patient.
Autism is a mental disorder which causes limited interaction
and communication with the society and is known as social
phobia. These types of children face a deficiency in speech
development and social behavior [3]. Autism is generally
reported to affect a person permanently, but many cases have
reported that some children suffering from ASC have
regained their social and communication skills and have
reached the range of cognitive, adaptive and social skills to
the level of typically developing (TD) children [4].The early
assessment of any disease from previously stored patterns and
time series information from the clinical databases is
explained in [5]. It uses sequential rules mining along with
classification modeling techniques. A peculiar identification
of autism patient’s feature “reach-and-throw” movements has
been done by using the SVM of machine learning [6].
Support Vector Machines (SVMs) is used to form a groups of
classes by quadratic programming with linear constraints [7]
[8]. Autism is recognized by the sensor parameters. Many
sensors like a skin conductance sensor is used for the Fig 1. Overview of our system
diagnosis of autistic children at the very beginning of the first
week of their birth, by measuring the galvanic skin According to the Association Rule (AR) of machine
conductance [9]. learning:
→Y
→ A Patient’s initial symptom identified from an
III. THE PROPOSED FRAMEWORK unknown set S
Our diagnosis method takes inputs from multiple Y →The target symptom to be pulled from a known symptom
sensors and smart phones. The design of our framework is set
shown in Fig.1.The sensor data is fetched through an Y is present in the domain knowledge database, and might
application run on a mobile node. By running the application, have been experienced by some other patient earlier.
the preliminary symptoms try to pull the associated In order to find Y we rank the rule in the form of
symptoms in various iterations from the database through the ‘interestingness’ at that particular time period.
Internet. The database contains records of the patient’s ‘Support’ and ‘Confidence’ are two factors of interestingness.
previous symptom patterns along with the domain expert’s Confidence (Conf) of rule →Y
decisions of symptoms set for the specified class of disease. Probability of Conf( →Y) =
A. Additional symptoms selection using the AR, mRMR,
MID and HIG techniques

In this system we have collected data from multiple sensors Probability of Support of is:
and smart phones for the primary investigation input for an Support ( ) = Number of time ‘ ’ experienced.
autistic patient. We have considered 50 predefined Autism We considered the rule for minimal support with highest
symptoms from which, any symptom can be present or absent confidence condition for selecting the probable additional
for the patient. On the basis of that initial symptom, other symptoms with the AR.
additional symptoms can be automatically compiled and a list We have used the minimum-Redundancy-Maximum-
of additional symptoms can be made to check whether the Relevance (mRMR) method to find the most appropriate

37
mutually independent symptoms which have not yet been iteration of the mRMR calculation. For disease prediction, we
examined. The output symptoms of mRMR (a) can identify if have predefined 6 different Autism cases using classifiers.
the symptom belongs to the class of disease (b) must be We have set the medium and high conditions of the chance of
mutually uncorrelated at the particular time. It derives the the disease. If the chance of having the disease is 40% then
mutual information between the symptoms. This mRMR this is identified as medium and 80% is identified as a high
method is used along with Mutual Information Difference chance of the disease. If the chance of the disease occurring is
(MID) in our proposed work. below 40% then it is not considered as the probable case of
the disease, and if the chance is above 80% then it is
B. The minimum Redundancy Maximum relevance (mRMR) considered as a highly probable case. If any condition holds
based symptom selection between 40% and 80% of chances i.e. between medium and
high chance of disease then the next step is followed. Again
we will try to select at least one more additional symptom to
The non-linear relationships between an experienced and
make a more reliable prediction of the disease. This condition
a yet-to-be experienced symptom’s theoretic ranking was
is called ‘grey zone’. The prediction of the disease under this
taken into consideration. We found that the minimum
condition is not confident. Then the Highest Information
Redundancy Maximum Relevance (mRMR) [11] method, can
Gain (HIG) method is applied to choose the additional
be used to choose the optimal symptom for a class. The
symptom to find the most confident case. According to HIG,
mutual information between target Y and the individual
the training data set is prepared by recalculating the values
feature should be maximized while searching for the
and providing a reweighted data set. So the gray zone
maximum relevant symptoms. Let D(S, Y) be the mean of the
symptoms are assigned with higher weightages. If any
mutual information between individual symptom and target
condition is still in the gray zone then the procedure
Y. The equation for the above statement is:
terminates and selects the high probability symptom sets for
1) max D( S,Y ) = the diagnosis. This will identify the most probable disease.
The symptom sets stored in the database are predefined by
If two symptoms X and Y are correlated, then they may the domain expert or the pattern that have previously
not necessarily belong to the same class of disease. They may occurred in past patients.
cause two different classes of diseases. So domain experts For example the autism patient symptoms sets
should not include two correlated symptoms in a symptom abbreviated by the following letters for the diseases are:
set. If the two symptoms are highly correlated then we do not d1, d2, -------dn.
include these symptoms in the same symptom set. For the Let the symptoms set of the diseases are
minimum redundancy criteria, the selection of symptoms is in d1={G,P,C,A,D,W,H},d2={R,M,K,O,F,T},---
such a way that, they mutually show maximum dissimilarity. ,d7={C,R,M,G},--- ,dn={D,H,R,F,K,I,V}.
Let R(S,Y ) be the mean of the mutual information According to our proposed technique we started
between pairs of symptoms in S. The equation for the above diagnosis with 2 preliminary symptoms of a patient i.e. ‘C’
statement is: and ‘R’ and finally got the symptom set {C,R,M,G} of
disease d7 which is already mentioned in the database.
2) min R(S,Y ) = =
Fig.2 shows the representation of the highest probability of G
The combination of both the above statements is called the by C, R, M. Fig.3 uses the AR, mRMR and MID to pull the
minimal-redundancy-Maximal-relevance (mRMR). If we associated symptoms with our patient input i.e. ‘C’and ‘R’.
simultaneously minimize R(S,Y) and maximize D(S, Y) then According to our example the support of ‘C’ is 13. ‘R’ is 15
the mRMR feature will be obtained. The mRMR technique and ‘M is 12.
combines two features to form a single criterion function at a
time. The equation for the Mutual Information Difference
(MID) criterion is as follows:

The iteration of mRMR will be repeated for finding the

proper symptoms to be pulled for which the value of the
function is maximized. MI ( , C) is the mutual information
between ‘ ’, a symptom of unknown symptoms set S and
‘C’ is the variable of class i,e the diseases under a class. The
‘Y’ is a symptom from the set of known symptom set ‘T’
Fig.2 Representation of highest probability of G by C, R, M
Containing |T| symptoms. When symptom ‘Y’ is selected and
added to the additional set of symptoms, then the symptoms
are treated as one of the known symptoms in the next

38
 Probability of conf(C,R→G) =0.02
Probability of conf(C,R→P) =0.04
Hence C,R satisfies the minimum redundancy with maximum
relevant method with the ‘M’ symptom of any unknown
disease not yet found. Now C,R pulls the symptom ‘M’ by
following the above technique. The process terminates if the
set of symptoms {C,R,M} is found in the database. If they are
not found then the mutual information of C,R,M has to be
calculated by applying the above three AR,MRMR and MID
methods. We found that:
Probability of conf(C,R,M→G) =0.6
Probability of conf(C,R,M→A) =0.04,
Hence C,R,M pull the next associative symptom
‘G’, where the MID equation is maximized and checks for
the disease from the database with the symptom set
{C,R,M,G}. Fig.2 represents the highest support of G by C,
R, M. According to the Association rule it was found that the
disease d7 has the symptom set {C,R,M,G}.Hence we
predicted that the disease is d7. The technique helps to
identify more number of symptoms according to the input
pattern and pattern present in the database. If any symptom
set has problem of confidence level, then it will follow HIG
method to get a confident symptom set for the disease
prediction.

IV. EXPERIMENTAL RESULTS

Fig.3. Process of symptom pulling using AR, mRMR, MID
This proposed technique can be applied to general diseases
diagnosis as well as for a specific class of diseases. In order
Probability of conf(C→G) = =0.61
to apply this technique the database used should contain
specific autism patient data and autism domain expert
Probability of conf(C→P) = =0.15 decision symptom set for disease. A screenshot of the part of
the database for the symptom set i.e. used for the diagnosis of
Probability of conf(C→A) = 0.15 disease is given in TABLE-I. It was collected from the
Autism Therapy Counselling and Help (CATCH,
Probability of conf(C→Q) = 0.07 Bhubaneswar, India) in April 2017, for experimental purpose.
All the training and testing data have also been collected from
Probability of conf(C→M) = = 0.84 the same source. However for reasons of privacy related
We consider highest Probability of conf(C→M) = 0.84 issues, we are not at present in a position to publish more
Likewise calculated details of the organization data. For our experiment purpose
we have collected domain knowledge of 10 different Autism
Probability of conf(R→A) = =0.06 conditions with 50 different symptoms. The table of 500
disease-symptom-sets was tested with 200 autistic children.
Probability of conf(R→G) = =0.66 From the experiment, for Autism with Intellectual Disability
(ID) we got 95% accuracy in the predictions, whereas the
Probability of conf(R→M) = 0.9 actual prediction accuracy was 86%. Likewise for Asperger’s
Syndrome (AS), we got 79% (predicted) and the number of
Probability of conf(R→P) = 0.2 actually predicted cases was 70%. For Vagus Nerve Disorder
(VND), we got 80% but actually it was 89%. For Multiple
Probability of conf(R→D) = 0.2 Sclerosis (MS), we got 88% but actually it was 90%. For
Pervasive Developmental Disorder (PDD), we got 70% but
We consider the highest actually it was 85 %. For Fear of elevators (FoE), we got 86%
Probability of conf(R→M) = 0.9 but actually it was 90%.We tested for 6 Autism cases
According to mRMR with MID, the mutual information of according to our procedure and finally we obtained an
‘C’and ‘R’ is as follows: average diagnosis accuracy of 83%. Fig.4. shows the data
Probability of conf(C,R→M) =0.72 tested for Intellectual disability. This test begins by giving the

39
 input as the patient stress level while trying to communicate symptom is negative .So the further pulling of symptom
i,e. with communication disorder and finally predicted the based on the positive nature of the symptoms. Fig.5.
disease Intellectual disability. The arrow mark (→) indicates represents the graphical presentation of comparison of the
the tested symptom is positive and × mark indicates the above cases.

TABLE I. Part of database used for Autism Disease diagnosis (Domain expert decision)

Fig.5 Accuracy of diagnosis based on the proposed

Fig.4 Tested diagnostic method on a real patient with Intellectual
technique and the real diagnosis
disability

40
IMTC 2005 – Instrumentation and Measurement REFERENCES
Technology Conference
Ottawa, Canada, 17-19 May 2005 [1] Salomon, Joshua A., et al. "Healthy life expectancy for 187 countries,
1990–2010: a systematic analysis for the Global Burden Disease Study
2010." The Lancet 380.9859 (2013): 2144-2162.
[2] Apiletti, Daniele, et al. "Real-time analysis of physiological data to
support medical applications." IEEE transactions on information
V. CONCLUSION AND FUTURE WORK technology in biomedicine 13.3 (2009): 313-321.
[3] “Diagnostic and statistical manual of mental disorders (5th Ed.),”
In Autism Disease diagnosis we have seen that the highest Washington, DC: American Psychiatric Association, 2013
[4] M. Helt, E. Kelley, M. Kinsbourne, J. Pandey, H. Boorstein, M.
accuracy is 83% using Machine learning Association Rule with Herbert, and D. Fein, “Can children with autism recover? if so, how?,”
minimum Redundancy Maximum Relevance (mRMR) method. Neuropsychology Review, vol. 18, pp. 339–366, December 2008
As we observed in the actual diagnosis of Pervasive [5] Cheng, Yi-Ting, et al. "Mining Sequential Risk Patterns From Large-
Developmental Disorder (PDD), the accuracy is 85%, but we Scale Clinical Databases for Early Assessment of Chronic Diseases: A
Case Study on Chronic Obstructive Pulmonary Disease." IEEE Journal
got 70% with our proposed algorithm. This is due to some of Biomedical and Health Informatics 21.2 (2017): 303-311.
missing data and problems in capturing data. (e.g. when we [6] Perego, Paolo, et al. "Reach and throw movement analysis with support
took the readings of the child he might not have exhibited some vector machines in early diagnosis of autism." Engineering in Medicine
of the symptoms, but in the actual case diagnosis time he may and Biology Society, 2009. EMBC 2009. Annual International
Conference of the IEEE. IEEE, 2009.
exhibit some of the symptoms). We have used the Mutual [7] Begg, Rezaul, and Joarder Kamruzzaman. "A machine learning
Information Difference (MID) method for selecting the approach for automated recognition of movement patterns using basic,
additional symptoms which can strengthen the diagnosis kinetic and kinematic gait data." Journal of biomechanics 38.3 (2005):
decision further. This system helps the doctors to diagnose the 401-408.
[8] Wu, Jianning, Jue Wang, and Li Liu. "Feature extraction via KPCA for
disease easily by seeing any one or two preliminary symptoms. classification of gait patterns." Human movement science 26.3 (2007):
Our future work is to reduce the time of finding the appropriate 393-411.
symptoms set and predict the disease more accurately. [9] Nehme, B., et al. "Developing a skin conductance device for early
Autism Spectrum Disorder diagnosis." Biomedical Engineering
(MECBME), 2016 3rd Middle East Conference on. IEEE, 2016.
ACKNOWLEDGEMENT [10] Yu, Lei, and Huan Liu. "Efficient feature selection via analysis of
relevance and redundancy." Journal of machine learning research 5.Oct
This work has been carried out with the support and funding of (2004): 1205-1224.
ITRA Media Lab Asia and DeitY through the project “Remote [11] Peng, Hanchuan, Fuhui Long, and Chris Ding. "Feature selection based
on mutual information criteria of max-dependency, max-relevance, and
Health: A Framework for Healthcare Services using Mobile and min-redundancy." IEEE Transactions on pattern analysis and machine
Sensor-Cloud Technologies”. intelligence27.8 (2005): 1226-1238.

41