AI models provide accurate and explainable insights to support the assessment of autism

Scientists have developed and tested deep learning models that can support clinicians by providing accurate results and clear, explainable insights, including model estimated probability scores for autism.

The model is outlined in a study published in eclinicalmedicinewas used to analyze static fMRI data. This is a non-invasive method that indirectly reflects brain activity through altered blood oxygenation.

In doing so, the model achieved cross-validation accuracy of up to 98% of autism spectrum disorder (ASD) and neural type classification, generating clear and explainable maps of brain regions that most affect their decisions.

Diagnosis of ASD has increased significantly over the past 20 years, partially reflecting increased awareness, increased screening, and changes in diagnostic criteria and clinical practice. Early identification and access to evidence-based support improves developmental and adaptive outcomes, and although effective, it may improve quality of life.

However, current diagnosis relies primarily on face-to-face and behavioral assessments, and there is an urgent need to improve evaluation pathways, as waiting for confirmed diagnosis can range from months to years.

Researchers hope that with further verification, their models will benefit people with autism and clinicians who evaluate and support them by providing accurate and explainable insights to inform decisions.

The study was the result of a final year undergraduate project by BSC (Hons) computer science student Suryansh Vidya, overseen by Dr. Amir Aly, and researchers in engineering, computing and mathematics at the University of Plymouth. They were supported by researchers who are part of the Peninsula Medical College by the University of Psychology School and the Cornwall Intellectual Disability Equitable Research (CIDER) group.

Dr. Aly, a university lecturer in artificial intelligence and robotics and academic lead of research and corresponding author, said, “There are over 700,000 people with autism in the UK. Many others are waiting to be evaluated. The journey to a certain decision can take months, as diagnosis still relies on expert behavior assessments.

“Our work shows how AI can help. Rather than replacing clinicians, we will further examine the support of prioritizing and adjusting ratings to support accurate and explainable insights, including model estimated probability scores.”

Using autism Brain Imaging Data Exchange (ABIDE) cohort containing 884 participants ages 7-64 at 17 sites, the team analyzed pre-processed RS-FMRI data and compared side-by-side for explanatory methods. Gradient-based techniques are the most effective, and the resulting maps are broadly consistent across pre-processing approaches, indicating which brain regions most affected the prediction of the model.

This research has already progressed with a doctorate. Researchers at Kush Gupta, co-author of the current study, incorporate a variety of multimodal data and machine learning models with the aim of developing robust, generalizable, AI-driven models that can support clinicians in autism assessment around the world. This complements Dr. Aly's broader research programme, including the use of robots to support people with autism, and the development of AI methods to analyze health sector data.

Professor Rohit Shankar MBE, professor of neuropsychiatry at the university and director of the Cider Group, is a senior author of the current study. He said, “I have shown that artificial intelligence can serve as a catalyst for advancement in early autism detection and diagnosis accuracy. However, some of the words of Robert Frost come to mind. Similarly, these are early prototypes that require further verification and research.”