PatientSpace: A multimodal graph-based latent representation framework for modeling neurodegenerative disease heterogeneity.

Neurodegenerative diseases such as Alzheimer's disease (AD) and frontotemporal dementia (FTD) exhibit substantial biological and clinical heterogeneity, complicating diagnosis, subtype characterization, and prediction of disease progression. We introduce PatientSpace, a multimodal graph-based latent representation framework designed to model neurodegenerative disease heterogeneity using T1-weighted MRI and FDG-PET. PatientSpace is built upon a structured variational autoencoder that integrates multimodal neuroimaging features while organizing patients within a latent space constrained by age, diagnosis, and a consistency regularization term encouraging similarity between neuroimaging phenotypes. This design enables the construction of an interpretable patient graph in which neighborhood relationships reflect biological similarity. Applied to cohorts of cognitively normal individuals, AD, and FTD patients, PatientSpace revealed multiple disease clusters associated with distinct neuroimaging patterns and clinical severity. Diagnostic classification achieved performance comparable to state-of-the-art deep learning models, while graph-based neighborhood inference enabled prediction of structural volumes, metabolic activity, and cognitive severity. Projection of mild cognitive impairment (MCI) subjects from an independent cohort further showed that cluster membership was associated with differential risks of dementia conversion and distinct longitudinal trajectories. Together, these results demonstrate that PatientSpace provides an interpretable framework linking multimodal neuroimaging representations to disease subtypes, patient-level characterization, and progression modeling in neurodegenerative disorders.