Learning deep representations of brain sMRI informed by latent psychopathology dimensions

Sammanfattning

Identifying neural correlates of psychiatric disorders is a fundamental step towards precision psychiatry – a paradigm aiming to discover biomarkers for early disease detection or personalised treatment selection to eventually improve patients’ quality of life. However, progress has been historically hindered by the classification of psychiatric disorders as discrete entities, suspected of not aligning with the disorders’ underlying biology. In response, modern frameworks have redefined psychiatric phenotypes along continuous latent psychopathology dimensions (LP-dimensions) aligned with biological and behavioural observations and therefore better suited for uncovering brainbehaviour correlations. Another limitation to this research is the small sample sizes of available clinical datasets. Deep representation learning offers a solution to this data scarcity by leveraging large-scale general datasets to learn transferable features to clinical cohorts. In this work, we investigate a method that combines LP-dimensions with transfer learning to generate clinically relevant representations of structural magnetic resonance imaging (sMRI) of the brain. We introduce psy-Aware, a contrastive learning model derived from y-Aware that incorporates LP-dimensions as weak supervision. We evaluate whether this approach improves upon standard contrastive learning using SimCLR. Our analyses focus on youth psychopathology, exploiting the large-scale, longitudinal ABCD study. This work makes three main contributions: (1) we derive a longitudinal three LP-dimensions model of youth psychopathology using exploratory factor analysis; (2) we benchmark two contrastive representation learning models on diagnosis prediction tasks; (3) we propose novel visualisations of the y-Aware loss and relate its behaviour with data distribution. We find that incorporating LP-dimensions yields comparable performances to standard contrastive learning approaches without significant improvement. We further identify the highly skewed distribution of LP-dimension as a likely limitation, suggesting that current psychiatric assessments may lack sufficient variance to fully support a dimensional approach in a broad population.