Neuromark v2: Spatiotemporal Templates to Search Reproducible Biomarkers in Hybrid ICA Framework

2215 

Abstract Submission 

Zening Fu¹, Lei Wu¹, Anees Abrol², Ishaan Batta³, Oktay Agcaoglu¹, Mustafa Salman¹, Yuhui Du⁴, Armin Iraji², Sarah Shultz⁵, Jing Sui⁶, Vince Calhoun⁷

¹Tri-Institutional Center for Translational Research in Neuroimaging and Data Science (TReNDS), Atlanta, GA, ²Georgia State University, Atlanta, GA, ³Georgia Institute of Technology, Atlanta, GA, ⁴Shanxi University, Taiyuan City, Shanxi Province, ⁵Emory University School of Medicine, Atlanta, GA, ⁶Beijing Normal University, Beijing, China, ⁷GSU/GATech/Emory, Decatur, GA

Zening Fu  
Tri-Institutional Center for Translational Research in Neuroimaging and Data Science (TReNDS)
Atlanta, GA

Lei Wu  
Tri-Institutional Center for Translational Research in Neuroimaging and Data Science (TReNDS)
Atlanta, GA

Anees Abrol  
Georgia State University
Atlanta, GA

Ishaan Batta  
Georgia Institute of Technology
Atlanta, GA

Oktay Agcaoglu  
Tri-Institutional Center for Translational Research in Neuroimaging and Data Science (TReNDS)
Atlanta, GA

Mustafa Salman  
Tri-Institutional Center for Translational Research in Neuroimaging and Data Science (TReNDS)
Atlanta, GA

Yuhui Du  
Shanxi University
Taiyuan City, Shanxi Province

Armin Iraji  
Georgia State University
Atlanta, GA

Sarah Shultz  
Emory University School of Medicine
Atlanta, GA

Jing Sui  
Beijing Normal University
Beijing, China

Vince Calhoun  
GSU/GATech/Emory
Decatur, GA

The most challenging topic in recent neuroscience is the reproducibility of population-based research (Poldrack et al., 2017). We previously developed a hybrid framework called Neuromark (Du et al., 2020), aiming to address the issue of reproducibility for biomarker development in data-driven methods. Neuromark has been successfully applied to many studies, capturing a bunch of robust brain markers across diseases (Fu et al., 2021; Dhamala, Yeo and Holmes, 2023; Vaidya et al., 2023). However, there is a limitation on the implicit assumption of invariant templates with age, which oversimplifies the changes in brain structure throughout the lifespan (Rieck et al., 2021). Therefore, in this study, we proposed to construct 4D (age × 3D brain) templates using a total of more than 6000 fMRI scans from four datasets. We evaluated the reproducibility of independent components (ICs) across data and investigated the unique and shared patterns across templates. Our study is the first attempt to build spatiotemporal templates compatible with the adaptive ICA framework, which will be beneficial for precisely capturing well-replicated mapping of abilities to brain functions.

We adopted the resting-state dataset from the human connectome project development (HCP-D) (Somerville et al., 2018) including 652 subjects aged 5~21 years old to build the developmental template. We built the aging template using the dataset from the HCP aging (HCP-A) (Harms et al., 2018), including 725 subjects aged 36~100+ years old. To build the infant template, we used two infant datasets, including 143 infants collected at Emory University. Group ICA was performed on each data, resulting in multiple groups of ICs for building the template. We examined the replicability of ICs using a greedy spatial correlation analysis. We considered the reference IC replicable if it had the best-matched ICs with correlations > 0.4. The replicable ICs were labeled as intrinsic connectivity networks (ICNs) to construct the template if their peak activations fell in the meaningful gray matter areas. We also examined the similarities and uniqueness shared across the templates by evaluating their spatial correlations.

Fig. 1A displays the results of ICs of sessional data from the HCP-D cohorts. Here, the first session is the reference, and the other sessions are the replication data. All 100 ICs have replicable ICs in the other sessions with a mean correlation > 0.4. Among the replicated ICs, 67 ICs were characterized as ICNs, which were used to construct the developmental template. For the HCP-A cohorts, 99 ICs from the reference data are replicable across sessions (r > 0.4). 56 of the 99 replicable ICs were identified as ICNs to construct the aging template (Fig. 1B). Fig. 1C displays the results of ICs from two infant datasets. 99 ICs from the reference data were replicated in the replication dataset, and we labeled 72 replicable ICs as ICNs to construct the infant template (Fig. 1C). These three templates share similarities and show unique patterns, whereas the older template tends to be more aggregated (Fig. 1D). Fig. 2 displays the composite maps of 4D templates. ICNs were arranged into 9 domains according to the prior functional information, including subcortical, hippocampal, auditory, sensorimotor, visual, cognitive-control, parietal, default-mode, and cerebellar domains.

The mixture of big data and algorithmic advances has propelled the neuroimaging field forward rapidly. Neuromark, combining priori neuroimaging with the data-driven approach, is a promising tool that provides a way forward here, continuously advancing the field. Neuromark v2 templates, which include spatiotemporal information, will have plenty of applications in future neuroimaging studies. They can capture features adaptable to the datasets and disorders in different age populations, which might boost accuracy in mental health research and advance our understanding of lifespan alterations.

Early life, Adolescence, Aging ²

Connectivity (eg. functional, effective, structural)

Brain Atlases ¹

Workflows

BOLD fMRI

Aging

Data analysis

Other - Neuromark; Hybrid; Spatiotemporal functional template; reproducible biomarkers