Advancing replicability of structural connectivity-based multivariate brain-wide association studies
Poster No:
1967
Submission Type:
Abstract Submission
Authors:
Raviteja Kotikalapudi1, Giuseppe Gallitto1, Balint Kincses1, Robert Englert1, Kevin Hoffschlag1, Jialin Li1, Ulrike Bingel1, Tamas Spisak1
Institutions:
1University Medicine Essen, Essen, NRW
First Author:
Co-Author(s):
Introduction:
Multivariate structural connectome-based predictive models – sCPM – are pivotal in brain-wide association studies (associating brain data with individual behavioral traits), yet the replicability of their predictive strength lacks conclusive evidence. This study aims at investigating the replicability of sCPM across diverse behavioral phenotypes using the HCP1200 dataset. Furthermore, we propose a novel data-driven feature thresholding approach and evaluate the effect of such approaches on replicability.
Methods:
MRI and behavioral phenotype data from HCP1200 young adult cohort was utilized. To generate participant (sample size = 1000) level structural matrices of white matter connectivity, mrtrix3-based fiber tractography was performed by generating 10M streamlines. Raw connectomes (TZ, no thresholding) – serving as baseline – were used to predict 58 different phenotypes. The replicability stream consisted of a) predicting phenotypes using a Ridge regression inside of a nested cross-validation, b) repeating the sCPM on increasing sample sizes (from N = 25 to 500, steps of 25) with 100 shuffles per instance, c) calculating the replication probability (Preplicability) for each sample size as the ratio of number of significant predictions in replication set given significant predictions in the discovery set, and number of significant predictions in the discovery set. We further hypothesize that the replication curve will improve given a thresholded connectome (instead of TZ), due to factors such as noise attenuation and focus on robust connections. To achieve this, the replicability stream was applied, not on TZ, but on thresholded connectomes (Tproportionality), where the threshold (hyper-parameter) was determined as connections present in a proportion of the discovery dataset (ranging from 1% to 100% of participants). Finally, the predictive performance in terms of mean absolute error was compared for un-thresholded and thresholded methods, for the replicable phenotypes, by training-testing a regression model on the entire dataset in a nested cross-validation scheme.
Results:
First, using the baseline model, we found that 9 (15.5%) phenotypes showed a median Preplicability > 0.7 across increasing sample sizes. Furthermore in comparison, while implementing connectomes with Tproportionality in predictive modelling, Preplicability=0.8 required comparatively less samples depending on the phenotype of interest (figure 1). Second, in terms of prediction, the mean absolute error (MAE) for sCPM with Tproportionality were lower in comparison to using Tz(figure 1 – confidence intervals).
·Replicability of structural Connectome-based Predictive Models across phenotypes from HCP1200 datasets.
Conclusions:
Overall, the replicability of phenotypes is dependent on the nature of the target variable. In general, a thresholded connectome guided predictive model (e.g., proportionality thresholding) might result in a better replicability with fewer samples, in comparison with raw connectome-based (un-thresholded) sCPM. These findings suggest that leveraging Tproportionality in sCPM offers practical advantages, potentially enhancing the efficiency of predictive models and refining the accuracy of predictions in the study of various phenotypes.
Modeling and Analysis Methods:
Multivariate Approaches 1
Novel Imaging Acquisition Methods:
Diffusion MRI 2
Keywords:
Data analysis
Modeling
Multivariate
Open Data
STRUCTURAL MRI
WHITE MATTER IMAGING - DTI, HARDI, DSI, ETC
1|2Indicates the priority used for review