Fabio Ferreira - James Chapman - CMIC/WEISS joint seminar series

Speaker: Fabio Ferreira
Title: Probabilistic machine learning models to uncover brain-behaviour associations and disease subtypes
Abstract: The heterogeneity of neurological and mental disorders has been a key confound to disease understanding and treatment outcome prediction, as the patient populations typically include multiple subgroups that do not align with the diagnostic categories. In this talk, I will present a probabilistic latent variable model, termed supervised Group Factor Analysis (GFA),  that uncovers associations between subsets of features within subgroups of subjects across multiple data modalities, e.g., brain imaging and behaviour, and assesses whether the latent components can be representative of underlying subtypes. I applied supervised GFA to a genetic frontotemporal dementia cohort to assess whether it could identify genotypes and explore within-genotype variability. Supervised GFA is interpretable, estimates uncertainty and can be easily extended to more complex models.

Speaker: James Chapman
Title: Demonstrating alternating optimisation for flexible regularisation in canonical correlation analysis
Abstract: Canonical Correlation Analysis (CCA) is a common method for modelling the relationships in multi-view datasets. It has found applications in computational psychiatry, genomics and neuroscience for both finding associations as well as modality fusion for downstream tasks. In high dimensional data and in many practical datasets, regularisation is desirable to make the problem well posed or to promote structure such as sparsity. In this talk we will demonstrate how alternating optimisation can be used as a scalable and flexible tool to find regularised solutions to the CCA problem.