Wall and colleagues recently published a study using artificial intelligence to develop a shortened interview to diagnose autism. They started with data from a group of children with autism who completed a comprehensive autism diagnostic instrument called the Autism Diagnostic Interview-Revised (ADI-R).
The ADI-R is a 93 item interview that queries a parent or other caretaker about the history and development for children two years of age and older. The interview covers early development, developmental milestones, social interaction, play behavior and language.
The ADI-R is considered one of the gold standard interviews for the diagnosis of autism. A key limitation of in the interview is that interviewers need extensive training to learn the interview and that it can take up to two and one half hours to complete for a single child.
The research team in the current study examined the responses ADI-R responses for 891 individuals with a research diagnosis of autism from the Autism Genetic Research Exchange (AGRE) data set. They used a variety of artificial intelligence strategies known as machine language algorithms in an attempt to determine if a smaller subset of items from the ADI-R could be efficient in accurately classifying autism.
After employing 15 distinct learning algorithms to the data set they identified the ADTree algorithm as the best algorithm. Using this algorithm, they were able to accurately categorize individuals as meeting the autism diagnostic category compared to control individuals. A subset of seven ADI-R items that accurately categorized individuals included the following items:
- Simple language comprehension
- Reciprocal conversation
- Imaginative play
- Imaginative play with peers
- Direct gaze
- Group play with peers
- Age abnormalities were first present
Using these seven items from the ADI-R produced a 99.9% accuracy in the AGRE data set. The research team then replicated this high diagnostic accuracy in a set of two independent research samples of nearly 2000 individuals with autism.
The authors note their seven item subset "could be of vaue in early, rapid detection of autism, a potential shared by and possibly combinable with recent work on observation-based detection of autism". The shortened interview may hold the advantage of less training requirements and may extened availability by adapting the items to a web-based or handheld device.
The authors also highlight limitations of this brief interview research. The data sets used for development of the artificial intelligence screen was comprised primarily of children and adolescents between the ages of 5 and 17. Its accuracy in younger children will need further research. An additional limitation is that the screen has not been validated in other autism spectrum diagnostic groups such as Asperger syndrome.
This research does highlight the potential of computer algorithm techniques for improving the clinical and research diagnosis of neurodevelopmental disorders. Look for more research combining complex computer models in the diagnosis of a variety of clinical neuroscience conditions.
Photo of adult flamingo feeding a young flamingo at the San Diego zoo from the author's files.