Artificial intelligence diagnoses autism spectrum better than specialists

Credit: Adobe Stock
Credit: Adobe Stock

Artificial intelligence diagnoses the autism spectrum better than specialists. The AI achieved sensitivity and specificity rates of 70 to 90 percent. The effectiveness of specialists ranges from 60 to a random 50 percent, says Dr. Izabela Chojnicka, assistant professor at the Faculty of Psychology, University of Warsaw.

The autism spectrum, related to atypical development of the nervous system, includes developmental challenges related to communication, social interactions, and the occurrence of repetitive, unusual and inflexible behaviours, interests and activities. 

Doctor of medical sciences, assistant professor Chojnicka at the Faculty of Psychology at the University of Warsaw, together with Dr. Aleksander Wawer from the Institute of Computer Science at the Polish Academy of Sciences and Dr. Justyna Sarzyńska-Wawer from the Laboratory of Psycholinguistics and Cognitive Psychology at the Institute of Psychology at the Polish Academy of Sciences conduct interdisciplinary research on the use of computational techniques (artificial intelligence and natural language processing) in autism spectrum research. 

AI is used to analyse spoken and written narratives. Algorithms 'learn' to detect linguistic differences between the statements of people on the autism spectrum and those of neurotypical people.

PAP: Please tell us about the methodology of your research.

Dr. Izabela Chojnicka: We conduct research using various control groups. We compare the statements of people with autism both with the statements of neurotypical people carefully selected in terms of age, gender and verbal and non-verbal intelligence quotients, as well as with people recruited from the population, which also includes people with developmental challenges or mental health problems other than autism. We analyse spoken statements and stories written by participants.

In the research conducted by our team, we analysed, for example, essays written by primary school children as part of the national eighth-grade exam. No one has yet worked with this type of data, despite the fact that they are generated in most countries in the Western world and large amounts of data for neural networks can be obtained from them.

We also obtain narratives using various stimuli, including the ADOS-2 tool. It is a standardized observation protocol for diagnosing autism spectrum disorders. It consists of various tasks and activities. One of them, for example, is telling the story presented in the pictures in a book. There are also personal narratives - the participants talk about their experiences, events from their life - we have prepared various topics of conversation.

PAP: What is the difference between the narrative of an adult person on the autism spectrum and the narrative of a neurotypical person?

I.C.: Let's start with the fact that the term 'autism spectrum' results from the fact that it is a very heterogeneous group and no two people on the autism spectrum are the same. The difficulties and skills of people on the spectrum are very different, they have a different intensity and may manifest themselves in different ways. However, statistical trends that appear in this group can be determined.

At the clinical level, for example, it can be observed that when a neurotypical person tells what they see in a picture, they tell the story contained therein in the form of a story with a plot. In the case of a person on the autism spectrum, we may encounter less cause-and-effect connections between individual images, and a greater focus on the details presented in individual illustrations than on the story as a whole.

PAP: So, for example, a person on the autism spectrum will say: there is a house, a girl is standing here, a dog is running, and a neurotypical person will say: the dog is running away from the house because it was afraid of what it saw there, and the girl is chasing it.

I.C.: You described it very accurately - one description lists directly visible elements, and the other offers insight into the intentions or emotional states of the characters. Moreover, when a drawing contains humorous, magical or unrealistic elements, they sometimes go unnoticed by a person on the spectrum, or are interpreted differently than by the majority of the population.

In order to capture such differences, we checked the level of abstraction of participants' statements using the Linguistic Category Model (LCM).

With this tool, we can calculate the level of linguistic abstraction for a given participant and their responses. Our research showed statistically significant differences in the level of linguistic abstraction between the narratives of people on the autism spectrum and the narratives of people from control groups. This applied to both spoken statements and written stories. In the narratives of neurotypical people, we more often recorded references to mental and emotional states that increased the level of abstraction of statements, and in the narratives of people on the spectrum there were statistically fewer of them, which does not mean that they do not exist.

We also used sentiment, emotional tone analysis in our research. We checked the number of positively emotionally charged words and negatively emotionally charged words in the narratives. As in the case of the linguistic abstraction analysis, we also noticed statistically significant differences here. In the narratives of people on the autism spectrum, we recorded fewer words with positive emotional content than in the narratives of people from the control groups.

PAP: Does this mean that people on the autism spectrum evaluate the world as worse or suffer more?

I.C.: Our research does not allow us to draw such a conclusion, because there were no statistically significant differences for words with a negative connotation, differences occurred in the case of words with a positive connotation.

PAP: How good are algorithms at selecting people on the autism spectrum based on their narratives?

I.C.: In our research, the effectiveness of artificial neural networks was very high; depending on the data and groups, we achieved sensitivity and specificity coefficients ranging from 0.70 to 0.90. These results are similar to the effectiveness of psychometric tools used in diagnostic processes or autism screening, which gives hope that in the future it will be possible to develop AI tools that will support the clinician and replace or complement the paper and pencil tools used today.

PAP: And what sensitivity do clinicians have?

I.C.: Our research also included the participation of so-called competent judges, who performed the same task as artificial neural network models for text analysis. The judges, i.e. specialist psychologists and psychiatrists experienced in diagnosing autism spectrum disorders, were tasked with assessing whether a given statement came from a person on the autism spectrum or from a person from the control group, without having any additional data about the participant. Their effectiveness was clearly lower, at the level of approximately 0.6 to 0.5, or 50 percent. This is a very low effectiveness, at a random level.

However, I would like to emphasise that the effectiveness of a clinician in the office, when they have much more data, come into contact with the patient, and can observe other, non-verbal means of communication - eye contact, facial expressions, gestures, way of speaking - is fundamentally higher. In addition, there may be more sessions, and the evaluation is made by a team, not one person.

PAP: This linguistic analysis tool at your disposal can identify individuals that should be subjected to deeper examination by a team of clinicians.

I.C.: This is called screening, a study conducted among people from the population to detect the risk of developing a certain disease, for example, or in the case of our conversation - autism spectrum disorder. Currently, screening tests are conducted using various questionnaires as well as other psychometric tools. Based on the completed questionnaire, the risk of autism spectrum disorder is calculated, and the person is referred for a much more accurate and complete diagnostic process. We can imagine that in the future, current paper and pencil tools will be replaced by AI-based computational methods. The frequency of diagnosing autism spectrum disorders is increasing every year, like other neurodevelopmental disorders, access to specialists is either expensive or limited, so such a screening in the form of computer tools would be useful in many areas of the world - it would increase availability and reduce costs. Especially in underdeveloped countries, where it is easier to get a smartphone than to see a good psychiatrist.

PAP: How many people are on the autism spectrum?

I.C.: According to the US Centers for Disease Control and Prevention, currently one in 36 children is diagnosed with autism spectrum disorder. Extensive research has not yet been carried out in Poland to collect such epidemic data. However, it is assumed that this indicator is 1%. The next research we are planning will be devoted to girls and women on the autism spectrum. We want to analyse women's narratives and check whether in their case we will observe similar relationships as in the groups studied so far, i.e. ones in which boys and men predominated.

Rozmawiała: Mira Suchodolska (PAP)

mir/ ral/ js/ kap/

tr. RL

The PAP Foundation allows free reprinting of articles from the Nauka w Polsce portal provided that we are notified once a month by e-mail about the fact of using the portal and that the source of the article is indicated. On the websites and Internet portals, please provide the following address: Source:, while in journals – the annotation: Source: Nauka w Polsce - In case of social networking websites, please provide only the title and the lead of our agency dispatch with the link directing to the article text on our web page, as it is on our Facebook profile.

More on this topic

  • Fot. Adobe Stock

    Mind wandering supports creativity, says scientist

  • Credit: Adobe Stock

    Polish astronaut on ISS to carry out psychological study prepared by University of Silesia

Before adding a comment, please read the Terms and Conditions of the Science in Poland forum.