Integrating multimodal connectivity improves prediction of individual cognitive abilities

Abstract

<jats:title>S<jats:sc>ummary</jats:sc></jats:title><jats:p>How white matter pathway integrity and neural co-activation patterns in the brain relate to complex cognitive functions remains a mystery in neuroscience. Here, we integrate neuroimaging, connectomics, and machine learning approaches to explore how multimodal brain connectivity relates to cognition. Specifically, we evaluate whether integrating functional and structural connectivity improves prediction of individual crystallised and fluid abilities in 415 unrelated healthy young adults from the Human Connectome Project. Our primary results are two-fold. First, we demonstrate that integrating functional and structural information – at both a model input or output level – significantly outperforms functional or structural connectivity alone to predict individual verbal/language skills and fluid reasoning/executive function. Second, we show that distinct pairwise functional and structural connections are important for these predictions. In a secondary analysis, we find that structural connectivity derived from deterministic tractography is significantly better than structural connectivity derived from probabilistic tractography to predict individual cognitive abilities.</jats:p>

Combining Regional and Connectivity Metrics of Functional Magnetic Resonance Imaging and Diffusion Tensor Imaging for Individualized Prediction of Pain Sensitivity

“…Thus, only using one MRI modality or using one type of feature (regional or connectivity) cannot offer a complete characterization of brain patterns related to pain and cannot provide sufficient information to accurately predict the individual pain sensitivity. Accumulated evidence have shown the importance of using multiple MRI modality in the understanding of cognitive functions and the diagnosis of neurological diseases ( Michels et al, 2017 ; Dhamala et al, 2020 ). For example, Xiao et al (2021) built a model to predict visual working memory capacity by using voxel-wise multimodal MRI features (amplitude of low-frequency fluctuations from fMRI, gray matter volume from structural MRI, and FA from DTI).…”

The Structural and Functional Neuroanatomy of Post-Stroke Depression and Executive Dysfunction: A Review of Neuroimaging Findings and Implications for Treatment

“…In stroke patients, structural dysconnectivity between cortical regions is correlated with the magnitude of functional connectivity. 88 Both structural and functional disconnection contribute unique variance to predicting cognitive functioning 89 and in cerebrovascular disease are associated with executive functioning. 90 The possibility that functional connectivity may shift or compensate for structural disconnection over time is also a relevant question to ensure that appropriate treatments are delivered depending on the time period post-stroke, especially as chronic stroke survivors can experience secondary degeneration in the thalamus, amygdala, hippocampus, and cingulum bundle.…”

Network level enrichment provides a framework for biological interpretation of machine learning results

“…Among modern ML techniques, linear regression models have found most success in human neuroimaging research, e.g. Ridge regression and Linear Support Vector Regression (LSVR) with a Ridge penalty (Cui & Gong, 2018; Dhamala et al, 2020; Taxali et al, 2021; Vergun et al, 2013). However, several challenges remain for practical application and interpretation of ML models in neuroscience research.…”

“…Several attempts have been made to employ random sampling with cross-validation to avoid the issue of shared variance among related individuals (Winkler et al, 2015). One such approach is to reduce the total sample size by only including non-related individuals (Dhamala et al 2020; Tian and Zalesky 2021; Gbadeyan et al 2022). For instance, some researchers retain only one member from each family (Li et al 2017; Nostro et al 2018; Elliott et al 2019; Seitzman et al 2020; Lohmann et al 2021).…”

“…Yet, biological interpretation of ML results remains a challenge for myriad reasons. While studies have interpreted the beta weights of features from ML models (Plitt et al 2015; Finn et al 2015; Dhamala et al 2020; Jiang et al 2020; Greene et al 2020; Bellantuono et al 2021; Kardan et al 2022), it has been shown that such interpretation could be seriously misleading because some features that are not related to the target label can still have significant weights in prediction (Chen et al, 2022; Haufe et al, 2014). For example, features with the strongest ML beta weights may reflect non-neuronal or nuisance signal such as head motion or machine noise (Chen et al, 2022; Haufe et al, 2014; Siegel et al, 2017).…”