No abstract

Resting-state functional connectivity (RSFC) has been widely used to predict behavioral measures. To predict behavioral measures, there are different approaches for representing RSFC with parcellations and gradients being the two most popular approaches. There is limited comparison between parcellation and gradient approaches in the literature. Here, we compared different parcellation and gradient approaches for RSFC-based prediction of a broad range of behavioral measures in the Human Connectome Project (HCP) and Adolescent Brain Cognitive Development (ABCD) datasets. Among the parcellation approaches, we considered group-average “hard” parcellations (Schaefer et al., 2018), individual-specific “hard” parcellations (Kong et al., 2021a), and an individual-specific “soft” parcellation (spatial independent component analysis with dual regression; Beckmann et al., 2009). For gradient approaches, we considered the well-known principal gradients derived from diffusion embedding (Margulies et al., 2016), and the local gradient approach that detects local changes in RSFC across the cortex (Laumann et al., 2015). Across two regression algorithms (linear ridge regression and kernel ridge regression), we found that individual-specific hard-parcellation performed the best in the HCP dataset, while the principal gradients, spatial independent component analysis and group-average “hard” parcellations exhibited similar performance. On the other hand, principal gradients and all parcellation approaches performed similarly in the ABCD dataset. Across both datasets, local gradients performed the worst. Finally, we found that the principal gradient approach required at least 40 to 60 gradients in order to perform as well as parcellation approaches. While most principal gradient studies utilize a single gradient, our results suggest that incorporating higher order gradients could provide significant behaviorally relevant information.

A primary aim of precision psychiatry is the establishment of predictive models linking individual differences in brain functioning with clinical symptoms. In particular, cognitive impairments are transdiagnostic, treatment resistant, and contribute to poor clinical outcomes. Recent work suggests thousands of participants may be necessary for the accurate and reliable prediction of cognition, calling into question the utility of most patient collection efforts. Here, using a transfer-learning framework, we train a model on functional imaging data from the UK Biobank (n=36,848) to predict cognitive functioning in three transdiagnostic patient samples (n=101-224). The model generalizes across datasets, and brain features driving predictions are consistent between populations, with decreased functional connectivity within transmodal cortex and increased connectivity between unimodal and transmodal regions reflecting a transdiagnostic predictor of cognition. This work establishes that predictive models derived in large population-level datasets can be exploited to boost the prediction of cognitive function across clinical collection efforts.