Burkhardt2024QSMPPEMA

Title

Quantifying Similarity between Graph-Theoretic Resting-State fMRI Data Processing Pipelines for Efficient Multiverse Analysis
View PDF | Save PDF

Authors

Micha Burkhardt, Andrea Hildebrandt, Carsten Gießing, Daniel Kristanto

Affiliations

Department of Psychology, Carl von Ossietzky Universitat Oldenburg, Oldenburg, Germany

Abstract

Multiverse analysis aims to enhance the robustness and replicability of scientific findings by testing research hypotheses through multiple, well-justified analysis pipelines. However, the multiverse of pipelines is often large making exhaustive evaluation computationally infeasible. Thus, a key goal is to approximate the multiverse by sampling a manageable number of pipelines for robustness analysis. For such an approximation, it is necessary to quantify the similarity between analysis pipelines and guide pipeline sampling by these similarities. To this end, we first used meta-analytic data from Kristanto et al 2024 on fMRI processing pipelines collected from a representative set of papers. Using this meta-analytic data, we propose a Graph Convolutional Network (GCN)-based approach combined with Deep Graph Infomax (DGI) to assess pipeline similarity. Graph-based embeddings were computed using unsupervised learning and subsequently used to derive pipeline features. Pipeline similarity was then quantified via Euclidean distance. Traditional similarity measures, namely Jaccard, Hamming and Levenshtein distances were also computed based on the meta-analytic data for comparison. Clustering analysis revealed consistency across the GCN, Hamming, and Levenshtein measures. Similarity measures based on Hamming and Levenshtein distances treated all processing steps identically, thus biasing them towards pipelines with identical step lengths. In contrast, the GCN-based measure generated distinct features for each step, allowing each to contribute differently to the pipeline similarity measure. Second, we compared the meta-analytically derived pipeline similarity measures with similarity measures obtained from multiverse analysis conducted on empirical data using resting-state fMRI measures from the Human Connectome Project. The comparison showed satisfactory results for the proposed approach, which aims to replace empirical similarity with meta-analytic similarity estimates for computationally efficient multiverse analysis in graph-theoretic fMRI research. These findings will inform future studies aimed at validating meta-analytic pipeline similarity measures based on empirical similarity estimates, providing a solid basis for the development of computationally feasible and valid multiverse analyses.

Keyphrases

Resting-state fMRI, multiverse analysis, data processing pipeline, graph neural network, similarity metric.

Citation

Brainiacs Journal 2024 Volume 5 Issue 2 Edoc XEE8F298E
DOI: 10.48085/XEE8F298E
PDP: /Nexus/Brainiacs/Burkhardt2024QSMPPEMA
URL: BrainiacsJournal.org/arc/pub/Burkhardt2024QSMPPEMA

Dates

Created 2024-09-09, presented 2024-10-09, updated 2024-12-22, published 2024-12-23.