AbstractBackgroundMore than 6 million Americans suffer from Alzheimer’s disease (AD), the most common age‐associated, neurodegenerative dementia, characterized by progressive cognitive decline and the sporadic emergence of neuropsychiatric symptoms (NPS). We applied linked independent component analysis (LICA), a unsupervised learning approach for data fusion, to identify multi‐modal gray matter patterns of covariation that were predictive of dementia ratings and depression symptoms to understand the brain circuitry underlying both.MethodPublicly available data from 277 participants (177 healthy controls, 61 mild cognitive impairment, and 39 AD (aged 58 – 93) in the Alzheimer’s Disease Neuroimaging Initiative (ADNI) dataset were used to conduct secondary analyses.T1‐weighted MR images were used to compute spatial maps of cortical thickness (CT) and pial surface area (PSA) using Freesurfer and grey matter (GM) density maps using FSLVBM. CT, PSA and GM maps for each modality were concatenated across participants into 4D data files that were analyzed using FSL Linked ICA to estimate 25 multi‐modal spatial components and subject loadings quantifying the strength of each component pattern per participant. Random forest regression with 5‐fold cross validation was used to predict CDR and GDS from the resulting subject loadings (with age, sex and site).ResultTwo LICA components (Fig 1) had highest feature importance in predicting CDR (R2 = 0.41). The same two patterns (Fig1) plus two additional patterns (Fig 2) were most important in predicting GDS (R2 = 0.08).ConclusionGreater CDR was predicted by larger (Fig 1A) decreased GM and CT of widespread temporal lobe regions and (Fig 1B) of reduced GM of entorhinal cortex/mesial temporal lobe with nodes of the default mode network, implicated in disease progression. Greater GDS was also predicted by these two patterns as well as a pattern (Fig 2A) reflecting decreased GM in nucleus accumbens, amygdala and a cerebellar‐thalomocortical (CTC) circuit strongly implicated in cerebellar cognitive affective syndrome (CCAS) and another (Fig 2B) comprised of CT decreases in cortex and temporal regions and GM decreases in frontoparietal regions resembling areas of the frontopariatal functional connectivity network. These multi‐modal brain‐behavior mappings could be translated to better understand dementia and neuropsychiatric symptoms of AD.