Multimodal Representation Learning for Parsing Biological Heterogeneity in Psychiatric Neuroimaging

For decades, psychiatric neuroimaging has searched for biomarkers of depression and other disorders, but they remain elusive in clinical practice. While the last 5 years have seen rapid progress, other large-scale correlative studies have found only small, unreliable links between brain measures and clinical symptoms. Growing evidence suggests that such limitations are not just about sample size but depend critically on how models represent data. This review traces a recent shift away from univariate methods to multivariate/multiview approaches that learn more effective representations of biological and symptom measures by flexibly learning multimodal latent representations. First, we review how linear multiview embedding methods have revealed reproducible biological depression subtypes but do not perform well in small samples or samples enriched for mild symptoms. Then, we consider newer work exploring more sophisticated representations for neuroimaging data, including deep-learning and graph-based representations, and multimodal extensions that uncover complex latent patterns that single-modality studies miss. Then, we review recent developments in foundation models, which, once trained on large corpora, can "transfer learn" readily to small clinical cohorts, potentially bringing the advantages of large-scale learning to small, privacy-limited data. Finally, we highlight emerging representation tools that treat the brain as a dynamic, stateful multivariate process. Taken together, these advances point to a future in which the value of neuroimaging will be determined not only by ever-larger sample sizes but also by data quality and by how well our algorithms capture the distributed, multimodal, and evolving nature of psychiatric disorders.