Autism spectrum analysed by AI to better understand neurodiversity

NCBR IDEAS reported on the research project in a release sent to PAP on the occasion of the World Autism Awareness Day celebrated on April 2.

People who openly identify as neurodiverse (e.g. with ASD, ADHD, dyslexia) run blogs, podcasts, fan pages on social media, and appear in videos and television shows. Psychiatrists also have collections of texts and recordings of people diagnosed with autism or ADHD. These statements (including transcriptions) make up a huge collection of texts that researchers can analyse to better understand the way non-neurotypical people see the world, their needs and challenges.

Instead of examining these texts with traditional methods (e.g. reading them, manually comparing them, searching for characteristic elements), scientists from IDEAS NCBR decided to use artificial intelligence. NLP (Natural Language Processing) algorithms make it possible to automatically identify specific topics, keywords, phrases, and semantic structures that are frequent or important for people on the autism spectrum. Thanks to this, it may be possible to draw attention to subtle patterns that may escape classical analysis.

It may be possible to use AI to find certain relationships in statements that will allow scientists to better understand the individual experiences of neurodivergent people. Thanks to this, it will be possible to better measure, analyse and potentially use these experiences in psychoeducation, for example.

The scientists from IDEAS NCBR want to combine two research approaches in their research: phenomenology and computational psychiatry. 'Phenomenology in the context of mental health focuses on reflecting, as faithfully as possible, the inner experiences of a person with ASD, not judging them, but trying to understand and describe them. It is like directing a film in which we try to show the viewer the world through the eyes of the main character. In turn, computational psychiatry is like a director's assistant who uses the latest technologies to analyse and process huge amounts of information about people's behaviour, thoughts and emotions,’ the scientists said in a press statement. 

The researchers want to focus primarily on first-person narratives of people on the autism spectrum. 'It is like trying to understand and capture the essence of a complex movie, where every frame, word and gesture has meaning, and then using this knowledge to create solutions that can help people better understand themselves and the world around them,’ the scientists add. 

Understanding how people on the autism spectrum experience the world also opens the door to more effective screening methods - earlier start of appropriate therapy is often associated with better educational, social and occupational outcomes.

Another matter is that such research may also contribute to increasing public awareness of the autism spectrum and the associated challenges. This, in turn, potentially leads to greater social acceptance and integration of people with autism, as well as to better adaptation of public institutions, schools and workplaces to the needs of neurodiverse people.

The methods developed by Polish scientists may enable the detection of subtle emotional cues, making it easier to understand how people on the autism spectrum experience different situations or types of social interactions. In turn, analysing the tone and emotional intent of a statement may be particularly useful for people who have difficulty expressing their emotions verbally.

'We plan to further expand the research area, using transcriptions of YouTube interviews and data from blogs. Ultimately, we want to obtain the broadest possible picture of the experiences of people with ASD,’ the research team leader Dr. Marcin Moskalewicz said. 

'The analysis of subjective experiences can help therapists, families and caregivers better imagine how differently their patients perceive various situations, even typical ones. This will translate into improved communication between them and more precise transfer of information. It will also increase their ability to support a person with ASD at home and in other environments,’ Dr. Marcin Rządeczka from IDEAS NCBR added. 

Researchers point out that stereotypes about autism are still embedded in large language models. Although, for example, at the text generation level, GPT-4 has numerous safeguards to protect against the repetition of stereotypes, these stereotypes still emerge if, for example, the model is asked to generate graphics illustrating autism.

For such a task, ChatGPT uses the image of a puzzle or the colour blue, despite the fact that the community of people on the autism spectrum is increasingly opposed to the use of such symbols. In GPT-4 visualizations, autism spectrum is quite strongly associated with disability, although a significant proportion of people on the spectrum do not consider themselves disabled. Additionally, most of the autism spectrum characters that GPT-4 generates are white males. 'The false belief that autism mainly affects men and manifests itself in white populations causes huge problems, e.g. in terms of equal access to diagnostics,’ the scientists said. 

IDEAS NCBR is a research and development centre operating in the field of artificial intelligence, established by the National Centre for Research and Development.

PAP - Science in Poland

lt/ agt/ kap/

tr. RL