An American study has developed a machine learning model that can identify attention deficit hyperactivity disorder (ADHD) to a 99% accuracy in adults who have received a childhood diagnosis.
Potentially having implications for detecting ADHD, which is notoriously difficult to diagnose due to the ever-changing presentations of the neurodiversity. But also the findings of this study may be relevant for future treatment plans, through understanding where a person can be placed on the ADHD spectrum.
While the behavioural presentations of ADHD may change daily, the brain communication signatures of the neurodiversity do not. So, researchers were able to identify communications between different specific brain regions that are suggestive of ADHD, and when these snapshots of the communications were computer analysed, they could individually predict ADHD to 91% and collectively to 99%.
The study's authors explained that the improved accuracy beyond a routine diagnosis with a specialist or just examining the different brain regions is that machine learning is well suited to analyse non-linear conditional relationships.
- See also: 'ADHD diagnosis and medication gives MSP Daniel Johnson a new lease of life'
- See also: 'Call to end stigmatisation of people living with ADHD'
- See also: 'Autism, OCD and ADHD may share a brain signature'
How does machine learning detect ADHD? And what could its future use be?
Lead author Chris McNorgan, assistant professor of psychology at the University at Buffalo New York, commented: “We measured how interconnected different brain regions were while they were doing a response inhibition task… A response inhibition task is a task where the subject has to refrain from doing a routine action under a particular condition. If you’re familiar with the schoolyard game ‘Simon Says’, that’s a response inhibition task. Brain regions with activity that go up and down together are highly connected, whereas those with relatively independent activity are weakly connected.”
“[These] patterns of connectivity were fed into a deep learning neural network that had to learn which patterns belonged to ADHD vs control, and also which patterns belonged to high vs low scorers on the Iowa Gambling Task.”
The Iowa Gambling Task is a diagnostic tool used to analyse and simulate real-life decision making and has been used to indicate and study ADHD. Prof McNorgan further explained that computer analysis makes predictions by calculating probabilities and uses trial and error to learn how each of the brain connections contributes to the overall probability of ADHD. And once the models are trained, they are tested on brain patterns that it hasn’t come across before and are used to classify it as either high or low on the Iowa Gambling Task.
While Prof McNorgan said that it would be somewhat impractical to use this technology for regular diagnosis, as it is too expensive, however, it may lead to a better understanding of the neurodiversity by clinicians.
Prof McNorgan explained: “For example, ADHD is typically associated with a low score on the Iowa Gambling Task, but some of our ADHD participants scored atypically high, and their brain networks looked different than those who scored low. This could be extended to other diagnostic measures used by clinicians, and give insight into which brain networks might be over- or under-active for a particular diagnostic profile, leading to better-targeted treatment.”