Research

Heterogeneity in neurodevelopmental disorders

Neuroanatomical variability in Autism

Autism Spectrum Disorder is very heterogenous. Understanding individual differences in neuroanatomy could lead to better predictions of disorder projection and pave the way for targeted healthcare approaches in ASD. I use Contrastive Variational Autoencoders (CVAEs) to identify ASD-specific neuroanatomical features, and generate AI-driven case-control brains ("synthetic TC-twins") to better understand ASD-specific individual differences in neuroanatomy.

Identifying clinical features at the individual level

Contrastive Variational Autoencoders (CVAEs) disentangle neuroanatomical features that are shared between ASD and Typical Control (TC) participants (neuroanatomical features unrelated to ASD) from variability that is ASD-specific.

Better relationships to symptoms

Once disentangled - Shared features (blue) correlate better with age, gender, and scanning site, while ASD-specific features (green) correlate better with clinical scores, such as ADOS and DSM-IV behavioral subtypes.

Replicability

Brain-behavior associations discovered in one dataset (ABIDE I) are replicable in an independent dataset (SFARI VIP). Additionally - SFARI VIP data shows that a CVAE model trained on ABIDE is sensitive to differences between 16p11.2 deletion and duplication differences

AI-generated counterfactual brains

How to identify brain areas affected by ASD in individual subjects, controlling for ASD-unrelated variability? CVAEs can be used to generate "synthetic TC-twins" closely matched on shared neuroanatomical properties. This allows for more precise characterisation of ASD-specific abnormalities, e.g. looking at first two principal components to reveal major axes of neuroanatomical variability in ASD.

Loci of neuroanatomical variability in ASD

1st PC positively correlated with differences in somatosensory, motor and language regions (among others) and correlated with ADOS RRB and ADOS communication symptoms. 2nd PC did so in theory-of-mind regions (among others) and correlated with ADOS communication deficits.

Neural bases of knowledge representation

Neural representation of person-knowledge

The brain is always active. But where are different aspects of knowledge stored in the brain? By pushing the brain towards different conceptual loads, we can see which brain regions are tuned to specific cognitive tasks (recalling names, physical attributes, social knowledge & memories)

Regional co-ordination

Group of regions responding together during retrieval of person knowledge (subnetworks).

Neural "taskonomy"

Which types of person-knowledge knowledge are represented similarly in brain? Which are distinct? In my PhD we identified that the brain uses distinct patterns to encodes certain domains of knowledge (trains, memories and names).

Cognitive "tuning"

Brain regions do not perform 'single' functions. But by looking at activity in context - we can see that some regions show preferential engagement in some task domains.

Methodology

Bayesian nonparametric modeling

If there exist subtypes of ASD - how many of them there are? two? four? a hundred? For some of the most important questions in precision psychiatry we don't always know what to look for. It is desirable to "let the data speak" and infer these answer from the data. Bayesian Nonparametric models are a class of probabilistic analytical tools that can infer this kind of structure from the data. In the example to the right, given an unknown dataset - these models can identify the underlying number of clusters and/or factors in an unsupervised manner.

Page updated

Google Sites

Report abuse