AI model offers accurate and explainable insights to support autism assessment

(Press-News.org) Scientists have developed and tested a deep-learning model that could support clinicians by providing accurate results and clear, explainable insights – including a model-estimated probability score for autism.

The model, outlined in a study published in eClinicalMedicine (a journal from The Lancet), was used to analyse resting-state fMRI data – a non-invasive method that indirectly reflects brain activity via blood-oxygenation changes.

In doing so, the model achieved up to 98% cross-validated accuracy for Autism Spectrum Disorder (ASD) and neurotypical classification and produced clear, explainable maps of the brain regions most influential to its decisions.

ASD diagnoses have increased substantially over the past two decades, partly reflecting greater awareness, expanded screening, and changes to diagnostic criteria and clinical practice. Early identification and access to evidence-based support can improve developmental and adaptive outcomes and may enhance quality of life, though effects vary.

However, because the current diagnosis primarily relies on in-person and behavioural assessments – and the wait for a confirmed diagnosis can stretch from many months to several years – there is an urgent need to improve assessment pathways.

The researchers hope that, with further validation, their model could benefit autistic people and the clinicians who assess and support them by providing accurate, explainable insights to inform decisions.

The study was the result of a final-year undergraduate project by BSc (Hons) Computer Science student Suryansh Vidya, supervised by Dr Amir Aly, and researchers from the School of Engineering, Computing and Mathematics at the University of Plymouth. They were in turn supported by researchers from the University’s School of Psychology and the Cornwall Intellectual Disability Equitable Research (CIDER) group, part of the Peninsula Medical School. 

Dr Amir Aly, Lecturer in Artificial Intelligence and Robotics at the University and the study’s academic lead and corresponding author, said: “There are more than 700,000 autistic people in the UK, and many others are waiting to be assessed. Because diagnosis still depends on a specialist, in-person behavioural evaluation, the journey to a confirmed decision can take many months – and, in some areas, years. Our work shows how AI can help: not to replace clinicians, but to support them with accurate results and clear, explainable insights, including a model-estimated probability score, to help prioritise assessments and tailor support once further validated.”

Using the Autism Brain Imaging Data Exchange (ABIDE) cohort, which included 884 participants aged 7 to 64 across 17 sites, the team analysed pre-processed rs-fMRI data and ran a side-by-side comparison of explainability methods. Gradient-based techniques performed best, and the resulting maps were broadly consistent across preprocessing approaches, showing which brain regions most influenced the model’s predictions.

The research is already being taken forward by PhD researcher Kush Gupta, a co-author on the current study, incorporating different kinds of multimodal data and machine learning models with the objective of developing a robust and generalisable AI-driven model that could support clinicians in autism assessment all over the world. This complements Dr Aly’s broader research programme, including the use of robots to support autistic people, and developing AI methods for analysing health-sector data.

Professor Rohit Shankar MBE, Professor in Neuropsychiatry at the University and Director of the CIDER group, is the current study’s senior author. He added: “We have shown that artificial intelligence has the potential to act as a catalyst for early autism detection and advancing diagnostic accuracy. However, some of Robert Frost’s words come to mind – ‘the woods are lovely dark and deep, but we have miles to go before we sleep’. In the same way, these are early prototypes which require further validation and research.”

END