Optimized prediction of cognition based on brain morphometry across the adult life span

Abstract

We mapped out the combined and unique contributions of 5 different biomarkers for 2 cognitive outcomes in cognitively healthy adults. Beside associations of biomarkers with cognition in the full experimental sample, we focused on how well any such associations would persist in held-out data. Three hundred thirty-five cognitively normal participants, 20e80 years older, were included in the study. Z-scores were computed for fluid reasoning and vocabulary. The following imaging data were included: regional brain volume, regional thickness, fractional anisotropy of white-matter tracts, volumes of select deep gray-matter regions, and global white-matter hyperintensity. Volume accounted for most of the variance in both cognitive domains. In out-of-sample data, fluid reasoning was best predicted by volumes, but vocabulary by the combination of all modalities. Although the predictive utility was better overall for older participants, the information gleaned relative to null models was less for older participants. An optimized set of brain biomarkers can thus predict cognition in out-of-sample data, to various degrees, for both fluid and crystallized intelligence.

Differential predictability of cognitive profiles from brain structure in older males and females

“…While results stand in contrast to findings in younger cohorts potentially due to differences in cognitive functions further aggravating only during the aging process (e.g. [ 18 , 29 ]), they fit previous accounts in studies across the lifespan and older cohorts [ 10 , 51 , 52 ]. For instance, Feng, Wang [ 52 ] revealed that language functions could be predicted to a considerably smaller degree than attention and executive functions from SC data in two large older cohorts, i.e.…”

Prediction of cognitive performance differences in older age from multimodal neuroimaging data

“…A multimodal approach, however, may allow for a more complete description of age-related cognitive decline than each single modality as aging has been found to affect the brain at all levels [ 67 ]. Initial encouraging results in different samples have demonstrated that the use of multimodal data may improve prediction performance for different cognitive abilities, e.g., fluid intelligence, global cognitive function, visual working memory, fluid reasoning, vocabulary [ 26 , 31 , 33 , 55 , 68 ]. For example, multimodal models, including information from structural and functional imaging, yielded improved prediction accuracies of up to R 2 = 0.05 compared to R 2 = 0.02–0.04 in unimodal models for fluid intelligence in a large sample from the UK Biobank [ 31 ].…”

“…Language functions, thus, not only appear to differ in aging trajectories (e.g., tend to remain more stable than for example executive and memory functions), but also in their predictability to other cognitive domains, e.g., processing speed, memory and executive functions, in older aged individuals [ 97 ]. A potential explanation may be that factors like education or occupational attainment may be highly relevant for the prediction of language-related cognitive performance overshadowing the predictive utility of brain features [ 26 , 98 ]. This is also supported by the feature importance analyses in the current study.…”

“…For instance, by investigating SC, i.e., nodal global and local efficiency, Li et al could successfully predict attention and executive function in a large sample of healthy older adults ( N = 633, age range: 45–86 years) [ 25 ]. In turn, regional GMV was found to predict fluid reasoning abilities across the adult population ( N = 335, age range: 20–80 years) in a study by Tsapanou et al [ 26 ], while Hilger et al revealed decidedly error-prone prediction of intelligence in a large sample of healthy adults ( N = 308, age range: 18–60 years) [ 27 ]. Moreover, recent results from our group emphasize low classifiability and predictability of RSFC strength measures for both, global and domain-specific cognitive abilities, in a large sample of older adults (age range: 55–85 years) [ 24 ].…”

“…This is due to the fact that they may provide information at the individual level and may find patterns in high-dimensional data that might be difficult to capture with univariate methods [21]. Initial ML approaches investigating either resting-state functional connectivity (RSFC), structural connectivity, or grey matter volume (GMV), revealed mixed prediction performance of cognitive measures [15,18,19,[22][23][24][25][26][27]. For instance, by investigating SC, i.e., nodal global and local efficiency, Li et al could successfully predict attention and executive function in a large sample of healthy older adults (N = 633, age range: 45-86 years) [25].…”