Scribd
Upload
12 views

Abstract

Abstract on autism diagnosis

Uploaded by

Viswanathan Puthanveettil
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
12 views

Abstract

Abstract on autism diagnosis

Uploaded by

Viswanathan Puthanveettil
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 2

Multimodal Behaviour Analysis for Early Diagnosis of

Autism in Children

Abstract

1 Background:
Early detection of autism is crucial for deciding the appropriate
educational and behavioural intervention at the most suitable time.
The lack of biological markers, the complexity of the diagnostic
procedure, dependence on behavioural observations and
developmental history, and subjective diagnosis demand the
introduction of assistive technologies in diagnosing autism in
children. Recent research has shown increased interest in leveraging
machine learning paradigms for identifying biological and
behavioural markers to aid the early detection of autism.

2 Objectives:
Previous research has shown that linguistic biomarkers have the
potential to efficiently classify children with autism (CwA) from the
typically developing (TD) population. Similarly, it has also been
found in the literature that the CwA exhibits atypicalities in
emotional perception and expression. Accordingly, in this work, we
propose a framework for multi-modal behaviour analysis, specifically
the speech and facial expression behaviour in children, to aid in the
early diagnosis of autism.

3 Methods:
For the speech analysis, we used the transcripts of ASD and TD
children from the Eigsti, Nadig, and Flusberg dataset provided by
Child Language Data Exchange System (CHILDES)
databank(MacWhinney, 2000). This consists of dyadic interaction
between the child and a conversation partner of 76 children (ASD-35,
TD-41). We extracted semantic and pragmatic features of the

1
child-partner interaction for the binary classification using a Support
Vector Machine (SVM) to predict if the child has ASD or not.
Further, we analysed the facial expressions exhibited by the children
to identify their emotional state, which could be used to enable
efficient interactions during the assessment sessions. Due to
confidentiality and privacy concerns, the facial expression data for
autistic children was not available. Hence, for our experiments, we
used a lab-curated dataset collected from YouTube, consisting of
facial emotion expressions (FER) of TD children of the age group
3-10 years under the six labels - happy, sad, anger, surprise, fear and
neutral. For the emotion recognition from FER data, we employed a
transfer learning-based approach for classifying the emotions from
the facial expression data collected as images.

4 Conclusions:
The results show that speech-based biomarkers supported by a
machine-learning paradigm are ideal candidates for the automatic
assessment of autism. Similarly, once validated among CwA, FER
can also aid the diagnosis. The next step is to develop a system to
evaluate the proof of concept by integrating the modules into an
assessment system and conducting a pilot study with children with
autism.

You might also like