Until now, this approach had eluded scientists, who focused solely on finding salient features in a single language. Furthermore, on other occasions machine learning and autism had paired the one to diagnose the second .
Last June, a new study led by researchers from Northwestern University paved the way for a new way to detect the disease . Thanks to this branch of artificial intelligence, they have identified consistent language patterns in children with autism, in both English and Cantonese .
The underlying problem
Children with autism (ASD) often speak slower than typically developing children (TD). They also have other differences in intonation, tone and rhythm. Such inconsistencies (called by researchers "prosodic differences") are surprisingly difficult to characterize in a coherent and objective way . In fact, their origins have remained unclear for decades.
All this limits scholars: in fact, those who do such research are hampered by the bias of the English language . Not to mention the subjectivity of human beings in classifying the language differences between people with autism and those without.
To highlight these differences, a team led by Molly Losh and Joseph CY Lau together with Patrick Wong used the technique of supervised machine learning.
Machine learning and autism, the procedure used
This study included participants from two language groups, consisting of native speakers of American English and Hong Kong Cantonese . The English group included 55 people with ASD (English ASD group) and 39 people with TD (English TD group). The Cantonese group included 28 people with ASD (Cantonese ASD group) and 24 people with TD (Cantonese TD group).
Participants had to narrate (in their respective languages) a 24-page wordless picture book, “Frog, Where Are You?”. The text presents the story of a child and his dog, as they search for the child's missing frog. This book is widely used in studies of narrative speech in ASD and other neurodevelopmental disabilities.
As the screen displayed each page and the participants told the story, the researchers recorded their voices on audio files. The collected data, analyzed with Audio Toolbox MATLAB , provided the team with interesting insights which they then analyzed into the following acoustic characteristics :
- speech rhythm , ie the variations in duration between the syllables of a sentence that indicate linguistic and affective properties;
- intonation , i.e. the variation of the pitch of the voice over time.
Then harnessing the power of machine learning algorithms, the researchers examined these aspects of prosody across time or frequency domains.
Results and future applications
Published in the journal PLOS Uno , the study showed that using acoustic characteristics is worthwhile for the purposes of a reliable classification of autistic status, both in English and in Cantonese.
As one of the Northwestern team leaders put it:
When you have such structurally different languages, any similarity in language patterns found in autism in both languages is likely to involve traits strongly influenced by the genetic responsibility of autism
He also added that Machine learning could be useful for developing tools to identify aspects of speech suitable for therapeutic interventions ; as well as measuring the effect of such interventions in assessing a speaker's progress over time.
Finally, as mentioned, the study results would be useful for understanding the role of specific genes and brain processing mechanisms involved in genetic susceptibility to autism. Ultimately, the goal is to create a more complete picture of the factors that determine the language differences of people with autism.
In fact, one brain network involved is the subcortical auditory pathway, strongly linked to differences in the way the brain processes speech sounds. Different in individuals with autism than in those with typical development, in all cultures. The next step will be to identify whether these processing differences lead to the language behavioral patterns observed in the study and their underlying neural genetics.
The article Machine learning and autism, a study between two languages was written on: Tech CuE | Close-up Engineering .