Joint representations from multi-view MRI-based learning support cognitive and functional performance domains

Multi-view magnetic resonance imaging (M3RI) offers a unique opportunity to simultaneously characterize structural and functional aspects of brain health in vivo. Here, we evaluate integrative dimensionality reduction strategies for brain--behavior mapping using the similarity-driven multi-view linear reconstruction (SiMLR) framework. Training was conducted on a large population sample from the UK Biobank (UKB, n = 21,300), with independent evaluation across three complementary cohorts: the Normative Neuroimaging Library (NNL, n = 164), the Alzheimer's Disease Neuroimaging Initiative (ADNI, n = 308), and the Parkinson's Progression Markers Initiative (PPMI, n = 1,070). We compared multiple objective functions within SiMLR to derive joint M3RI embeddings and assessed their utility in participants aged 18--89 across all datasets. The optimal embeddings captured robust, systems-level representations of brain organization that generalized to diverse behavioral and clinical domains. This work represents the first systematic evaluation of a joint multi-modality learning framework using both objective data-driven metrics and heterogeneous behavioral phenotypes across the adult lifespan. The resulting models and resources establish a foundation for sensitive normative mapping of individual differences in brain health, with open-source methods and data available to qualified investigators.