Decoding Autism Diagnosis: A Journey Towards Transparency with

XAI in ML Models

Shivani Pandya

and Swati Jain

Institute of Technology, Nirma University, Sarkhej Gandhinagar Hwy, Ahmedabad, Gujarat, India

Keywords:

Autism Spectrum Disorder, Explainable Artiﬁcial Intelligence, LIME, SHAP, Machine Learning.

Abstract:

Autism Spectrum Disorder (ASD) is a developmental condition that manifests within the ﬁrst three years

of life. Despite the strides made in developing accurate autism classiﬁcation models, particularly utilizing

datasets like AQ-10, the lack of interpretability in these models poses a signiﬁcant challenge. In response to

this concern, we employ eXplainable Artiﬁcial Intelligence (XAI) techniques, speciﬁcally Local Interpretable

Model-agnostic Explanations (LIME) and Shapley Additive exPlanations (SHAP), to enhance transparency.

Our primary aim, following the commendable accuracy achieved with the AQ-10 dataset, is to demystify the

black-box nature of machine learning models used for autism classiﬁcation. LIME provides locally faithful

explanations, offering a more nuanced understanding of predictions, while SHAP quantiﬁes the contribution

of each feature to the model’s output. Through instance-based analyses, we leverage these XAI techniques

to delve into the decision-making processes of the model at an individual level. Integrating LIME and SHAP

not only elevates the model’s trustworthiness but also helps a deeper comprehension of the factors inﬂuencing

autism classiﬁcation. Our results underscore the efﬁcacy of these techniques in unraveling the intricacies of

the model’s decisions, shedding light on relevant features and their impact on classiﬁcation outcomes. This re-

search contributes to bridging the gap between accuracy and interpretability in machine learning applications,

particularly within the realm of autism classiﬁcation.

1 INTRODUCTION

Autism Spectrum Disorder (ASD) is a behavioral

condition that affects how individuals interact with

society throughout their lifetime. Symptoms of ASD

typically appear during childhood and persist into

adolescence and adulthood (Hasin et al., 2013). ASD

patients exhibit repetitive behaviors, and analyzing

these behaviors can aid in early detection. The di-

versity of behaviors demonstrated by ASD patients

depends on age and ability. Common behavioral dis-

orders in ASD patients include deﬁcient expressive

gestures, non-responsiveness to sound, lack of proper

eye contact, no sensation of pain, repetition of words,

and agitation with changes in daily routines (Hasin

et al., 2013). Compared to healthy populations, sib-

lings of individuals with autism are at a ﬁfty times

greater risk of developing ASD (Joseph and Tager-

Flusberg, 1997). Additionally, males are 4-5 times

more likely to be affected than females. The World

Health Organization reports that 1 in 160 children

https://orcid.org/0000-0001-9348-1840

https://orcid.org/0000-0002-5708-7472

worldwide is prone to developing ASD at any given

time. Hong Kong, South Korea, and the United States

have the highest prevalence rates of ASD. In India,

the prevalence rate is 1 in 500, with an incidence rate

of 11,914 people yearly. In recent years, ASD preva-

lence has increased from 15 to 64 per 10,000 in India

(Geetha et al., 2019). According to the Autism Soci-

ety of America, the incidence rate of autism is rising

at a rate of 10-17% each year in the USA. In 2020, the

CDC (Centers for Disease Control and Prevention) re-

ported a 10% increase in autism incidence rate, with 1

in 54 children in the USA being diagnosed with ASD.

These statistics are alarming, considering that ASD is

a rare disease (Shaw et al., 2023).

While ASD cannot be fully cured, early detec-

tion of symptoms can help reduce the effects of

the disease. Machine learning (ML) has been ap-

plied to predicting and detecting various diseases with

good accuracy, including ASD, based on multiple

physical and physiological parameters (Raj and Ma-

sood, 2020). However, detecting and analyzing ASD

is challenging due to the existence of other men-

tal health problems with typical symptoms, which

700

Pandya, S. and Jain, S.

Decoding Autism Diagnosis: A Journey Towards Transparency with XAI in ML Models.

DOI: 10.5220/0012456300003657

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 17th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2024) - Volume 2, pages 700-707

ISBN: 978-989-758-688-0; ISSN: 2184-4305

can result in false detection. To address this issue,

a machine learning model is proposed in this paper

for early prediction of ASD through screening test

datasets, with the application of Explainable Artiﬁcial

Intelligence (XAI) to identify the contribution of fea-

tures towards accurate prediction. Comparative case

studies have been conducted using available features

in the dataset and the most contributing features pre-

dicted by XAI through performance metrics: accu-

racy, precision, recall, and F1-score (Das and Rad,

2020).

1.1 Contribution

1. This research focuses on detecting ASD using

AQ-10 datasets. It provides a comprehensive

guide on implementing machine learning algo-

rithms speciﬁcally designed for this purpose.

2. Addresses interpretability issue in black box mod-

els with XAI techniques (LIME and SHAP).

3. Provides detailed analysis of each prediction and

identiﬁes speciﬁc attributes with the greatest im-

pact on the decision-making process.

The paper is divided into four main sections. Sec-

tion 2 explores the related work, providing a context

and existing knowledge in the ﬁeld. Moving forward,

Section 3 gives a concise overview of the dataset and

describes the pre-processing steps taken. Section 4

presents the proposed work and provides an in-depth

examination of the methodology used in our exper-

iments. Section 5 unveils the results of our exper-

iments and engages in a comprehensive discussion.

Finally, Section 6 delivers a conclusive summary and

ﬁndings, encapsulating the paper’s key insights and

future scope.

2 RELATED WORK

Based on the reviewed literature, Machine Learning

(ML) techniques have proven highly effective in pre-

dicting various diseases based on syndromes. For ex-

ample, Cruz et al. (Cruz and Wishart, 2006) used

ML to diagnose cancer, demonstrating its broad range

of applications in predicting diseases. Khan et al.

(Khan et al., 2017) also employed ML to predict dia-

betes, showcasing its versatility in addressing differ-

ent health concerns. Wall et al. (Wall et al., 2012) uti-

lized the Alternating Decision Tree (ADTree) method

to detect ASD traits and achieved high accuracy

within a speciﬁc age range (5-17 years). However,

their approach was limited in its ability to predict

ASD across diverse age groups. Bone et al. (Bone

et al., 2016) applied a Support Vector Machine (SVM)

to predict ASD traits, achieving notable sensitivity

and speciﬁcity. However, the study faced constraints

due to a wide age range (4-55 years). Allison et

al. (Allison et al., 2012) achieved over 90% accu-

racy in ASD screening by employing the ’Red Flags’

tool in conjunction with the Autism Spectrum Quo-

tient. Thabtah (Thabtah, 2017) conducted a compre-

hensive comparative analysis of previous ML algo-

rithms for autism trait prediction. In contrast, Hauck

and Kliewer (Hauck and Kliewer, 2017) identiﬁed

pivotal screening questions for the Autism Diagnos-

tic Observation Schedule (ADOS) and Autism Diag-

nostic Interview-Revised (ADI-R) methods, empha-

sizing the efﬁcacy of their combined use. Bekerom

(van den Bekerom, 2017) used various ML techniques

to discern ASD traits in children, accounting for fac-

tors like developmental delay, obesity, and physi-

cal activity levels. Wall et al. (Wall et al., 2012)

made progress in classifying autism through concise

screening tests, with the ADTree and functional tree

exhibiting promising results. Heinsfeld (Heinsfeld

et al., 2018) explored algorithms and neural networks

for identifying ASD patients, achieving classiﬁcation

accuracy ranging from 66% to 71%. Meanwhile,

Liu (Liu et al., 2016) investigated the potential of

face-scanning patterns to identify children with ASD,

achieving exceptional accuracy, speciﬁcity, sensitiv-

ity, and Area Under the Curve (AUC) metrics. Bone

et al. (Bone et al., 2015) conducted a meticulous

analysis of prior studies, addressing conceptual and

methodological issues and successfully replicating re-

sults using their ML approach. Despite the extensive

research in this ﬁeld, a deﬁnitive conclusion regard-

ing the universal applicability of ML-based autism

screening across age groups remains elusive. There-

fore, the existing tools and techniques call for devel-

oping a comprehensive app-based solution tailored to

different age demographics.

Numerous studies have delved into the application

of machine learning techniques to predict ASD from

high-dimensional datasets. Archana et al. (Archana

et al., 2023) introduced a comprehensive methodol-

ogy encompassing dimensionality reduction, feature

extraction, and classiﬁer selection, achieving notable

advancements in ASD prediction. Their approach

addresses the challenge posed by large feature sets

in high-dimensional data. Additionally, similar ef-

forts have been made by several researchers. They

demonstrated promising results in reﬁning classiﬁers

for enhanced ASD prediction. Furthermore, it exam-

ined various classiﬁers, emphasizing the potential of

Decision Tree classiﬁers in achieving high accuracy

rates in ASD prediction. Building upon these pre-

Decoding Autism Diagnosis: A Journey Towards Transparency with XAI in ML Models

701

ceding works, the current study reﬁnes and extends

these methodologies, achieving a remarkable testing

accuracy of 97.47% with a reduced feature set of 254

features and a model training time of 23.508 seconds.

This research contributes signiﬁcantly to the evolving

landscape of machine learning-based ASD prediction,

offering promising avenues for improving early diag-

nosis and intervention strategies for individuals with

ASD.

3 DATASET DESCRIPTION AND

PRE-PROCESSING

Data is a crucial component in AI, and to achieve sub-

stantial efﬁciency, a large amount of data must be an-

alyzed. Numerous resources have been dedicated to

collecting data, including Dr. Fadi Fayez Thabtah

(Thabtah, 2019; Thabtah et al., 2018), who has fo-

cused on autism assessment in young children. The

dataset used in this study was curated through the

ASDTests mobile application, which can be accessed

at ASDTest. This dataset represents a signiﬁcant ad-

vancement in screening autism, particularly in infants,

and highlights several inﬂuential factors that require

further exploration. These factors are essential not

only for detecting autistic traits but also for reﬁning

the categorization of ASD. As part of the diagnosis

process for ASD, behavioral traits are recorded using

the Q-chat-10 questionnaire.

This dataset has been designed speciﬁcally for

classiﬁcation tasks in ASD. It consists of 704 in-

stances, with each instance containing 21 attributes.

These attributes include a mix of categorical, con-

tinuous, and binary data. The dataset provides de-

mographic details such as age and gender, as well

as contextual factors like ethnicity and family history

of Pervasive Developmental Disorders (PDD). Addi-

tionally, it includes speciﬁcs about the screening pro-

cess, such as who completes the test, the country of

residence, and whether a screening app was used pre-

viously. The dataset also describes screening methods

by age category and captures binary responses (0 or 1)

to ten questions as mentioned in table 1 integral to the

screening process.

In the process of developing a classiﬁcation

model, the attributes, denoted as [X], serve as in-

dependent variables. These features are utilized to

construct the model, where the target variable, rep-

resented as [Y], is binary. Speciﬁcally, [Y] indicates

the presence (1) or absence (0) of autistic traits.

4 PROPOSED WORK

Our primary objective is to build a model that gener-

ates predictions and provides insight into its decision-

making process. We aim to understand the reasons be-

hind the model’s decisions and the factors contribut-

ing to its predictions. This requires us to examine

each output and explain the basis of each decision. As

we implement machine models, it is crucial to con-

sider interpretability in the model’s outcomes.

4.1 LIME

Local Interpretable Model-agnostic Explanations

(LIME) (Ribeiro et al., 2016) is a technique that aims

to explain the predictions of machine learning mod-

els, particularly those considered black-box models.

The main goal of LIME is to provide interpretable

explanations for individual predictions, which can

help humans understand why a speciﬁc prediction

was made. The process involves selecting a particular

instance and generating perturbed versions of that

instance by slightly and randomly modifying its

features. These perturbed instances are then used

to obtain predictions from the black-box model.

After that, a locally interpretable linear regression

model is trained on the perturbed instances and their

corresponding black-box model predictions. The

optimization problem can represent the working

equation of LIME:

minimize L( f , g, π

) = E

′

∼π

[L ( f , g, x

′

)]+Ω(g)

Here, f represents the black-box model, g is the

interpretable model, π

is the probability distribution

over perturbed instances, x

′

is a perturbed instance,

L is a loss function measuring the difference be-

tween black-box and interpretable model predictions,

and Ω(g) is a regularisation term discouraging model

complexity. This equation encapsulates the optimiza-

tion problem that LIME addresses, seeking an inter-

pretable model g that effectively captures the behavior

of the complex model f for the chosen instance while

promoting simplicity through regularisation.

4.2 SHAP

SHapley Additive exPlanations (SHAP) (Lundberg

and Lee, 2017) is an algorithm that helps explain the

output of machine learning models fairly and consis-

tently. It uses the concept of Shapley values from

cooperative game theory, where each feature is con-

sidered as a ”player” in a game, with the prediction

being the outcome. The Shapley value provides a

HEALTHINF 2024 - 17th International Conference on Health Informatics

702

Table 1: Q-chat-10-Toddler Features Corresponding to Variables in Dataset.

Variable

Dataset

Corresponding Q-chat-10-Toddler Feature

A1 Does your child look at you when you call his/her name?

A2 How easy is it to make eye contact with your child?

A3 Does your child point to indicate that s/he wants some-

thing? (e.g., a toy that is out of reach)

A4 Does your child point to share interest with you? (e.g.,

pointing at an interesting sight)

A5 Does your child pretend? (e.g., care for dolls, talk on a

toy phone)

A6 Does your child follow where you’re looking?

A7 If you or someone else in the family is visibly upset, does

your child show signs of wanting to comfort them? (e.g.,

stroking hair, hugging them)

A8 Would you describe your child’s ﬁrst words as:

A9 Does your child use simple gestures? (e.g., wave good-

bye)

A10 Does your child stare at nothing with no apparent pur-

pose?

way to distribute the contribution of each feature to

the model’s prediction fairly.

The algorithm considers all possible feature com-

binations for each prediction using a permutation-

based approach. It calculates the difference in the

model output when a feature is included versus when

it is excluded, representing the marginal contribution

of that feature. The Shapley value for a speciﬁc fea-

ture is then computed by taking the weighted average

of these marginal contributions over all possible com-

binations. The weights are determined by the number

of ways each combination can occur.

Mathematically, the equation for SHAP values

for a particular feature i in a speciﬁc prediction

instance involves a summation of all possible subsets

of features. This equation ensures that the Shapley

values satisfy essential properties such as consistency

and linearity.

( f ) =

∑

S⊆N\{i}

|S|!·(|N|−|S|−1)!

|N|!

[ f (S ∪ {i}) − f (S)]

In this equation, N represents the set of all fea-

tures, S is a subset excluding feature i, and f (S) and

f (S ∪ {i}) denote the model output for subsets S and

S with feature i included, respectively. The Shap-

ley value for feature i is determined by considering

its marginal contribution to all possible subsets, en-

suring a comprehensive and fair explanation of the

model’s predictions. The workﬂow of both algo-

rithms is shown in ﬁgure 1.

For Feature Importantace and

Ranking

LIMESHAP

ML/ DL model Classification

Dataset

Autistic

Non-

Autistic

Result

But why?

Machine Learning workflow

Explainable AI

Black box

Relevent Feature and score

(geasture)

A7 (Sign of

comfort)

A6 (Looking

where point)

0.085085

0.048019

0.052081

Figure 1: Proposed Methodology for XAI.

5 RESULT AND DISCUSSION

We used various machine learning algorithms on the

AQ-10 dataset, a questionnaire designed to measure

autistic traits. During the development of a model, the

data preprocessing stage is carried out on the whole

dataset. This stage includes tasks such as handling

missing values and scaling features. After this, the

dataset is divided into two sets: 80% for the train-

ing set and 20% for the testing set. Hyperparame-

ter tuning is then performed. Following this, cross-

validation is applied to the training set to assess the

model’s generalization across different subsets.

The ﬁnal model is then trained on the entire train-

ing set. Its evaluation is carried out on a separate

testing set that was not used in the preceding stages.

while evaluations are based on metrics such as accu-

Decoding Autism Diagnosis: A Journey Towards Transparency with XAI in ML Models

703

racy, precision, recall, and F1 score. The performance

of each model on the AQ-10 dataset is summarized

below in the table. 2.

Following the implementation of machine learn-

ing algorithms on the dataset, we speciﬁcally chose

the Random Forest algorithm to interpret the model

predictions. The selection was motivated by the en-

semble nature of Random Forest, which consists of

multiple decision trees. Understanding individual de-

cision trees can be challenging due to their com-

plex structures, prompting our preference for Random

Forests as they provide a diverse collection of trees,

contributing to a more interpretable and robust model.

Table 2: Autism Screening AQ-10 dataset Accuracy.

Algorithm Training Acc Testing Acc

Decision Tree 1.0 1.0

SVM 1.0 1.0

Logistic Regression 1.0 1.0

Random Forest 1.0 0.995

k-Nearest Neighbors (knn) 0.962 0.947

MLP 0.93 0.92

We have chosen two instances to compare. In In-

stance 1, the class label is ’Yes’ or ’1’, representing

the presence of autistic traits. In Instance 2, the class

label is ’No’ or ’0’, representing the absence of autis-

tic traits or non-autistic. We applied the random forest

algorithm and then used the LIME and SHAP meth-

ods to interpret the model’s decisions. Below are the

results.

5.1 LIME: Interpretation

Figure 2: LIME interpretation for Instance 1, where the in-

dividual identiﬁes as autistic Class: Yes.

LIME plays a crucial role in making autism classiﬁca-

tion model predictions more understandable through

visualizations that highlight positive and negative as-

sociations with speciﬁc features. Consider Figure. 2,

where the model establishes a positive relation be-

tween features A9, A7, and A6 and the likelihood of

predicting autism. In this visualization, a green color

signiﬁes this positive association. The intensity of the

green color indicates that higher levels of these fea-

Figure 3: LIME interpretation for Instance 2, where the in-

dividual identiﬁes as non-autistic Class: No.

tures are linked to an increased probability of autism

prediction. Now, consider another scenario, as shown

in Figure 3, where features negatively affect autism

prediction. In this case, LIME represents the negative

correlation with a red color. The increasing intensity

of red communicates that as these features increase,

there is a decreased likelihood of the model predicting

autism. These color-coded visualizations from LIME

offer valuable insights into the importance of speciﬁc

features in predicting outcomes for individual cases.

They contribute to a nuanced understanding of how

the model makes decisions in autism classiﬁcation.

Consider Instance 1, depicted in Figure 4. In this

case, the actual label is autistic. The prediction label

score is notably high at 0.99 for a ”YES” prediction,

indicating a 99% probability that the person is autis-

tic. This score is derived from a combination of dif-

ferent features, and the contribution of each feature is

visually represented in the ﬁgure. Conversely, when

the actual label is non-autistic (NO), as illustrated in

Figure 5, the prediction score is much lower at 0.071.

This implies that there is a mere 0.071% chance that

the person could be autistic, leading the model to pre-

dict a non-autistic classiﬁcation. Analyzing such re-

sults allows us to interpret the model’s predictions and

conduct a detailed analysis of the contribution of each

feature to the model’s decision-making process.

5.2 SHAP: Interpretation

SHAP is a global interpretation method that helps un-

derstand how a model operates. As shown in Figure

6, the summary plot is an excellent tool for visualiz-

ing how each feature contributes to the model’s out-

come. The plot uses a color-coding system to repre-

sent each feature based on its distance from the base-

line. Warmer colors are used for features that have

a more signiﬁcant impact on the model’s prediction.

In contrast, cooler blue colors indicate a negative or

lesser impact on the model’s outcome. The SHAP

summary plot aids users in comprehending the impact

of individual features on the model’s predictions, en-

hancing the interpretability of the model by providing

HEALTHINF 2024 - 17th International Conference on Health Informatics

704

Figure 4: LIME interpretation for autistic class.

Figure 5: LIME interpretation for non-autistic class.

insights into the relative importance and directionality

of each feature.

In this detailed analysis using SHAP values, we

focus on interpreting the predictions of a model for

two instances; for instance 1 7, we observe positive

SHAP values for attributes A1, A2, A3, A6, A7, A9,

and A10. This means that higher values in these at-

tributes positively predict Yes meaning the presence

of autistic traits. By this, we can observe that Features

such as A9 (with a SHAP value of 0.085085) have a

powerful positive impact on the model’s afﬁrmative

prediction.

On the other hand, for instance, 2, 8, which is

labeled as No, we observe negative SHAP values for

attributes A1 to A10. This indicates that higher values

in these attributes negatively inﬂuence the model’s

prediction of the absence of autistic traits. Features

like A9 have a signiﬁcant impact on predicting the

negative outcome. We can clearly understand the

interpretation of the SHAP values using the color-

coded visualization, where red highlights feature

with higher values that strongly impact positive

predictions. In comparison, blue indicates features

with higher values that have a mitigated impact on

negative predictions. This comprehensive analysis

Decoding Autism Diagnosis: A Journey Towards Transparency with XAI in ML Models

705

Figure 7: SHAP interpretation for Instance 1, where the individual identiﬁes as autistic Class: Yes.

Figure 8: SHAP interpretation for Instance 2, where the individual identiﬁes as non-autistic Class: No.

Figure 6: SHAP summary of overall model prediction.

provides a nuanced understanding of the importance

of features in individual predictions. It underscores

the transparency and interpretability offered by SHAP

values in unraveling the intricacies of the model’s

decision-making process.

Through an examination of LIME and SHAP

analyses, it becomes evident that features such as

”Does your child use simple gestures? (A9),” ”Does

your child comfort you? (A7),” and ”Does your child

follow where you point? (A6)” play a pivotal role

in both positive and negative predictions made by

the model. Notably, these features exhibit signiﬁcant

prominence in the SHAP values as well. This sub-

stantiates the conclusion that gestures, expressions of

comfort, and the ability to follow directions are ar-

eas where individuals with autism may experience

discomfort or challenges. The consistent emphasis

on these features in both LIME and SHAP analyses

strengthens the inference that they signiﬁcantly in-

ﬂuence the model’s predictions. This insight holds

valuable implications for tailoring child-oriented ther-

apies, as it underscores the importance of addressing

and supporting individuals with autism in the domains

of gestural communication, providing comfort, and

following cues. Leveraging these ﬁndings can con-

tribute to developing more targeted and effective in-

terventions for autistic children, fostering better un-

derstanding and support in areas where they may face

difﬁculties.

6 CONCLUSION AND FUTURE

SCOPE

In our approach to autism classiﬁcation through var-

ious machine learning algorithms, our goal is to

achieve consistent accuracy while acknowledging the

inherent complexity of these models. We employ De-

cision Trees, Random Forests, SVM, Logistic Regres-

sion, knn, and an MLP. Emphasizing interpretability,

we leverage two widely used eXplainable Artiﬁcial

Intelligence methods—SHAP and LIME. SHAP em-

ploys game theory principles to calculate Shapley val-

ues, providing insights into how features collectively

impact predictions. Meanwhile, LIME offers speciﬁc,

instance-level explanations. Looking ahead, we plan

to enhance explanations by incorporating domain-

speciﬁc knowledge, exploring advanced visualization

techniques, and staying current with emerging in-

terpretability methods in healthcare data. We aim

for continuous adaptation, real-time model monitor-

ing, and exploring novel approaches to boost trans-

parency. This includes integrating model-agnostic in-

terpretability methods with domain-speciﬁc knowl-

edge and staying informed about state-of-the-art tech-

niques. Through these efforts, we aspire to improve

transparency in our autism classiﬁcation model and

contribute to advancing interpretable machine learn-

ing in clinical applications.

HEALTHINF 2024 - 17th International Conference on Health Informatics

706

REFERENCES

Allison, C., Auyeung, B., and Baron-Cohen, S. (2012). To-

ward brief “red ﬂags” for autism screening: The short

autism spectrum quotient and the short quantitative

checklist in 1,000 cases and 3,000 controls. Journal

of the American Academy of Child & Adolescent Psy-

chiatry, 51(2):202–212.

Archana, P., Sirisha, G., and Chaitanya, R. K. (2023).

Prediction of autism spectrum disorder from high-

dimensional data using machine learning techniques.

Soft Computing, pages 1–7.

Bone, D., Bishop, S. L., Black, M. P., Goodwin, M. S.,

Lord, C., and Narayanan, S. S. (2016). Use of machine

learning to improve autism screening and diagnos-

tic instruments: effectiveness, efﬁciency, and multi-

instrument fusion. Journal of Child Psychology and

Psychiatry, 57(8):927–937.

Bone, D., Goodwin, M. S., Black, M. P., Lee, C.-C., Au-

dhkhasi, K., and Narayanan, S. (2015). Applying ma-

chine learning to facilitate autism diagnostics: pitfalls

and promises. Journal of autism and developmental

disorders, 45:1121–1136.

Cruz, J. A. and Wishart, D. S. (2006). Applications of

machine learning in cancer prediction and prognosis.

Cancer informatics, 2:117693510600200030.

Das, A. and Rad, P. (2020). Opportunities and challenges

in explainable artiﬁcial intelligence (xai): A survey.

arXiv preprint arXiv:2006.11371.

Geetha, B., Sukumar, C., Dhivyadeepa, E., Reddy, J. K.,

and Balachandar, V. (2019). Autism in india: a case–

control study to understand the association between

socio-economic and environmental risk factors. Acta

Neurologica Belgica, 119(3):393–401.

Hasin, D. S., O’brien, C. P., Auriacombe, M., Borges, G.,

Bucholz, K., Budney, A., Compton, W. M., Crowley,

T., Ling, W., Petry, N. M., et al. (2013). Dsm-5 criteria

for substance use disorders: recommendations and ra-

tionale. American Journal of Psychiatry, 170(8):834–

851.

Hauck, F. and Kliewer, N. (2017). Machine learning for

autism diagnostics: applying support vector classiﬁca-

tion. In Int’l Conf. Heal. Informatics Med. Syst, pages

120–123.

Heinsfeld, A. S., Franco, A. R., Craddock, R. C., Buch-

weitz, A., and Meneguzzi, F. (2018). Identiﬁcation of

autism spectrum disorder using deep learning and the

abide dataset. NeuroImage: Clinical, 17:16–23.

Joseph, R. M. and Tager-Flusberg, H. (1997). An investiga-

tion of attention and affect in children with autism and

down syndrome. Journal of autism and developmental

disorders, 27(4):385–396.

Khan, N. S., Muaz, M. H., Kabir, A., and Islam, M. N.

(2017). Diabetes predicting mhealth application us-

ing machine learning. In 2017 IEEE international

WIE conference on electrical and computer engineer-

ing (WIECON-ECE), pages 237–240. IEEE.

Liu, W., Li, M., and Yi, L. (2016). Identifying children

with autism spectrum disorder based on their face pro-

cessing abnormality: A machine learning framework.

Autism Research, 9(8):888–898.

Lundberg, S. M. and Lee, S.-I. (2017). A uniﬁed approach

to interpreting model predictions. Advances in neural

information processing systems, 30.

Raj, S. and Masood, S. (2020). Analysis and detection

of autism spectrum disorder using machine learning

techniques. Procedia Computer Science, 167:994–

1004.

Ribeiro, M. T., Singh, S., and Guestrin, C. (2016). ” why

should i trust you?” explaining the predictions of any

classiﬁer. In Proceedings of the 22nd ACM SIGKDD

international conference on knowledge discovery and

data mining, pages 1135–1144.

Shaw, K. A., Bilder, D. A., McArthur, D., Williams,

A. R., Amoakohene, E., Bakian, A. V., Durkin, M. S.,

Fitzgerald, R. T., Furnier, S. M., Hughes, M. M., et al.

(2023). Early identiﬁcation of autism spectrum dis-

order among children aged 4 years—autism and de-

velopmental disabilities monitoring network, 11 sites,

united states, 2020. MMWR Surveillance Summaries,

72(1):1.

Thabtah, F. (2017). Autism spectrum disorder screening:

machine learning adaptation and dsm-5 fulﬁllment. In

Proceedings of the 1st International Conference on

Medical and health Informatics 2017, pages 1–6.

Thabtah, F. (2019). Machine learning in autistic spec-

trum disorder behavioral research: A review and ways

forward. Informatics for Health and Social Care,

44(3):278–297.

Thabtah, F., Kamalov, F., and Rajab, K. (2018). A new

computational intelligence approach to detect autistic

features for autism screening. International journal of

medical informatics, 117:112–124.

van den Bekerom, B. (2017). Using machine learning for

detection of autism spectrum disorder. In Proc. 20th

Student Conf. IT, pages 1–7.

Wall, D. P., Dally, R., Luyster, R., Jung, J.-Y., and DeLuca,

T. F. (2012). Use of artiﬁcial intelligence to shorten

the behavioral diagnosis of autism.

Decoding Autism Diagnosis: A Journey Towards Transparency with XAI in ML Models

707