Back to Projects

AI-enabled Imaging

Predicting autism spectrum phenotypes from neonatal brain connectivity

Project ID: 2021_011

Student: Ioannis Valasakis

Joint co-1st supervisor: Dafnis Batalle, King’s College London
Joint co-1st supervisor: Maria Deprez, King’s College London
Clinical Champion: Grainne McAlonan, King’s College London

Aim of the PhD Project:

The goal of this project is to improve our understanding of brain development and outcome in babies at risk of autism spectrum disorders (ASD) by:

  • Training machine learning algorithms in adult populations to classify ASD from controls
  • Transferring techniques to state-of-the-art neonatal acquisitions
  • Interpreting algorithms to find altered connectivity patterns associated with outcome

Project Description / Background:

Genetic and environmental risk factors acting from before and shortly after birth are associated with Autism Spectrum Disorders (ASD). However, ASD is highly diverse and not everyone at-risk goes on to develop the condition. Since early interventions work best, understanding underlying mechanisms that lead to ASD and establishing who is most likely to benefit from treatment is currently one of the most important neuroscientific challenges.

Histopathology and neuroimaging studies in children and adults with a diagnosis of ASD have shown disruptions to the organisation of neural systems. Recently, using a simple measure of functional connectivity (degree centrality), reproducible alterations in patients with ASD have been reported in independent datasets. Brain network analysis techniques have been used to show a reduction in global communication capacity in brain networks of three-year-old children with ASD. Despite these recent insights, even in young children, we still cannot easily untangle the causes of ASD from the secondary or compensatory effects of living with the condition. Thus, to really examine what makes a brain vulnerable to ASD, or indeed what might be protective, studies assessing structural and functional connectivity in the neonatal period in infants at risk of ASD together with information about childhood outcomes are needed.

Diffusion MRI (dMRI) and functional MRI (fMRI) have been widely employed to study structural and functional brain connectivity in the healthy and diseased neonatal brain. Recently, artificial Intelligence (AI), and machine learning in particular, has had a large impact on the field of medical image analysis and is opening new avenues of research. This project aims at taking advantage of such algorithms to address a key clinical challenge: link neonatal brain connectivity to typical and atypical neurodevelopmental trajectories and provide a means to subgroup (‘stratify’) the at-risk population earlier than ever before.

We already have a large and growing dataset acquired at KCL of typically-developing infants (N=~800); with state-of-the-art diffusion and functional neonatal MRI and neurocognitive follow-up. Importantly, we also scan neonates who are at greater risk of developing ASD traits than average because they have a strong family history of ASD or have had a significant perinatal exposure linked to ASD, recruited as part of the EU-AIMS Brain Imaging in Babies (BIBS) study. In addition, we will also have access to data from the KCL-lead EU-AIMS LEAP project, the largest multi-centre, multi-disciplinary observational study worldwide that aims to identify and validate stratification biomarkers for ASD. Data available includes 437 children and adults with ASD and 300 individuals with typical development. We will also take advantage of publicly available adult datasets such as ABIDE (including 539 ASD subjects and 573 typically developing controls).

In this project we will assess the predictive power of known reproducible alterations in ASD neuroimaging characteristics such as degree centrality in adults. Then we will transfer trained predictive algorithms to assess the capacity of neonatal brain connectivity markers to predict specific autistic phenotypes measured later in life, such as neurobehavioral testing and eye-tracking paradigms. Using newly developed interpretable machine learning methods we will determine if there are underlying patterns in neonatal brain network organisation and function which relate to childhood outcomes.

Expected academic background: mathematics, computer science, engineering and/or physics.

Figure 1: Resting-state networks of the neonatal brain identified by group independent component analysis


Back to Projects