1742

Abstract Submission

Dimensionality reduction techniques are increasingly utilized to characterize meaningful organizational principles within high-dimensional brain connectivity data. The dimensions of such low-dimensional representations - connectivity gradients - capture topographical patterns of intrinsic brain organization (Margulies et al., 2016; Bernhardt et al., 2022). Their calculation usually relies on a n-by-n affinity matrix that is constructed by pairwise connectivity between n nodes. The computational cost increases exponentially with the number of nodes, and for high-resolution data spaces such as the 90,000 grayordinates of the Human Connectome Project, more than 100GB of memory is required for the calculation. This renders voxel-wise calculation of gradients often intractable on consumer hardware, typically requiring users to downsample the data, e.g. through a parcellation strategy. While parcellation and the entailed data averaging can increase signal-to-noise ratio, concerns about the loss of detail and the appropriate choice of parcellation remain.

Here, we propose Fast Connectivity Gradient Approximation (FCGA), a computationally efficient approach to establish high-resolution connectivity gradients, leveraging consumer-grade hardware. At its core, the approach uses a subset of connectivity targets to approximate the underlying connectivity structure at full scale (Figure 1a). We evaluated our approach on the group and individual level, using two different datasets: the Human Connectome Project (HCP; Glasser et al., 2013) and the Enhanced Nathan Kline Institute - Rockland Sample (NKI-RS; Nooner et al., 2010). We evaluated the performance of different connectivity targets based on parcellations and individual vertices, randomly and uniformly distributed across the cortex. We quantified the spatial similarity (Spearman's ⍴) between approximated gradients (G_fcga) and gradients based on the full connectivity matrix (G_fullfc). Furthermore, we studied the practical implications of gradients based on parcellated data by comparing the predictive performance (age and intelligence) to parcellated fine-scale gradients.

The spatial similarity between G_fcga and G_fullfc increased with the number of connectivity targets used to calculate the approximated gradients (Figure 1b). Remarkably, when using 1000 Schaefer parcels as connectivity targets (~1.7% of the full connectivity matrix), the average spatial similarity across 25 connectivity gradients was ⍴ > 0.85. Increasing the number of targets further to 3000 uniformly sampled vertices (~5% of the full matrix), an average spatial similarity ⍴ > 0.98 was achieved with <10% computational time and memory usage, compared to the calculation of G_fullfc (Figure 1c). On the individual level (HCP, n=100), reliability and discriminability analysis confirmed the repeatability and the preservation of individual features for G_fcga (Figure 2ab). Importantly, in brain-behavior prediction using a lifespan cohort (NKI-RS, n=313, age 6-86y), averaged fine-scale gradients G_fcga with parcels outperformed gradients calculated from parcellated time series (Figure 2c). This was observed for both age and intelligence across various parcellations.

Overall, approximation of full-scale connectivity gradients is computationally efficient, feasible on commodity hardware, showing a ⍴ > 0.98 spatial similarity with the full gradient results at a fraction (~10%) of the computational costs. Importantly, calculating large-scale gradients preserves more relevant information for predicting age and intelligence as gradients calculated from parcellated data. The high fidelity with gradients based on the full connectivity matrix paired with its ability to run on consumer hardware can both democratize this powerful approach and advance new insights across a range of applications.

Classification and Predictive Modeling ^{2}

fMRI Connectivity and Network Modeling ^{1}

Workflows

Other - connectivity gradients

Glasser, M.F. (2013), ‘The minimal preprocessing pipelines for the Human Connectome Project’, Neuroimage, vol. 80, pp.105-124.

Margulies, D.S. (2016), ‘Situating the default-mode network along a principal gradient of macroscale cortical organization’, Proceedings of the National Academy of Sciences, vol. 113, no. 44, pp. 12574-12579.

Nooner, K.B. (2012). ‘The NKI-Rockland Sample: A Model for Accelerating the Pace of Discovery Science in Psychiatry’, Frontiers in Neuroscience, vol. 6, 152.