How Reliable Are Networks? A Bayesian Modeling Approach

Céline Provins, Yasser Alemán Gómez, Pedro M Gordaliza, Dominic M. Padova, Matteo Mancini, Violeta Céspedes, Gang Chen, Patric Hagmann, Oscar Esteban

Keywords: Bayesian modeling, Reliability, Reliability Characterization, Uncertainty, Graph, Networks, Neuroscience, PYMC, Probabilistic Programming, Brain, MRI, Magnetic Resonance Imaging, Structural Connectivity

TL;DR: We present and validate a Bayesian modeling framework to assess the reliability of network connections across repeated measures.

Abstract: Network analyses of white matter pathways linking brain regions---noninvasively extracted from diffusion magnetic resonance imaging---hold great clinical application promise. However, these networks display low reliability at the level of single brain connections, severely limiting inference. We present a Bayesian modeling framework to assess the reliability of network connections across repeated measurements. We model connection strength as a mixture of two probabilistic components: one representing the presence of a true connection, and its true absence. Using simulated, repeated-measures data, we estimate the posterior distribution of connection strengths and quantify the reliability by examining the spread of these distributions. The model was sensitive to connections with varying levels of reliability. However, it underestimated the probability that a connection is absent, and failed to recover the parameters after generating data with the same model.

Submission Number: 37