Joint representations from multi-view MRI-based learning support cognitive and functional performance domains
Discuss this preprint
Start a discussion What are Sciety discussions?Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
Background
Multimodal MRI (sMRI, dMRI, rsfMRI) encodes complementary aspects of brain structure and function; principled joint representations promise more sensitive and interpretable markers of brain health than single-modality features.
Methods
We evaluate Normative Neurological Health Embedding (NNHEmbed), a flexible multi-view framework that uses constrained cross-modal similarity objectives to learn low-dimensional embeddings. Models were trained and tested on the UK Biobank (n = 21,300) and evaluated for transfer and longitudinal sensitivity in independent cohorts (Normative Neuroimaging Library, Alzheimer’s Disease Neuroimaging Initiative, Parkinson’s Progression Markers Initiative).
Results
NNHEmbed produced compact, biologically interpretable components that (a) maps to established neurocognitive systems (e.g., episodic memory, processing-speed, sensorimotor/basal-ganglia circuits), (b) generalizes across cohorts, and (c) captures within-subject change over time. Best configurations balance reconstruction fidelity and shared covariance, improving interpretability while preserving predictive utility. Case demonstrations illustrate individualized normative profiling across multiple visits.
Conclusions
NNHEmbed yields stable, transferable multimodal embeddings suitable for normative mapping and longitudinal monitoring. Software, NNHEmbed configurations and derived bases are available for reproduction and reuse.
