Do estimates of age based on functional connectivity differ as a function of reading ability?

350 

Abstract Submission 

Oliver Lasnick¹, Jie Luo¹, Brianna Kinnie¹, Shaan Kamal², Spencer Low³, Natasza Marrouch⁴, Fumiko Hoeft¹

¹University of Connecticut, Storrs-Mansfield, CT, ²University of Connecticut School of Medicine, Farmington, CT, ³Boston University Chobanian & Avedisian School of Medicine, Boston, MA, ⁴University of Lisbon, Lisbon, Lisbon

Oliver Lasnick, M.S.  
University of Connecticut
Storrs-Mansfield, CT

Jie Luo  
University of Connecticut
Storrs-Mansfield, CT

Brianna Kinnie, B.S.  
University of Connecticut
Storrs-Mansfield, CT

Shaan Kamal, MD, M.S.  
University of Connecticut School of Medicine
Farmington, CT

Spencer Low  
Boston University Chobanian & Avedisian School of Medicine
Boston, MA

Natasza Marrouch, Ph.D  
University of Lisbon
Lisbon, Lisbon

Fumiko Hoeft, MD, Ph.D  
University of Connecticut
Storrs-Mansfield, CT

Previous work has shown that brain growth charts which predict age based on properties of the brain can inform scientists about neurodevelopmental trajectories, and may be used to obtain early markers for atypical development (Dosenbach et al., 2010; Kessler et al., 2016). In this study, we extended brain-age modeling based on functional connectivity (FC) data to developmental dyslexia (reading disorder, or RD). We hypothesized that (1) models trained to predict age are biased for poor and exceptional readers compared to controls, reflecting a developmental delay for the former and acceleration for the latter; (2) a model trained with whole-brain FC data better predicts age compared to a model trained only with FC data from regions of interest (ROIs) in the brain's reading network; (3) models trained with FC data from the reading network have a greater prediction bias than the whole-brain model, i.e. they are more likely to underestimate the age of poor readers.

We used fMRI scans of N=742 participants aged 6-21 years (M=10.7, SD=3.0) from the public, de-identified CMI HBN database; IRB approval was not required. We used the Test of Word Reading Efficiency Total Word Reading Efficiency index score to classify participants as poor (PR, <=90), typical (TR, 91-109), or exceptional readers (ER, >=110). A support vector machine was trained on fMRI-FC data to predict age. An ROI correlation matrix was generated for each participant (each value is the correlation between two fMRI BOLD-signal time series). Principal component analysis (PCA) was applied to the training set of connectivity matrices to reduce the number of features. One thousand training permutations were done: for each, data was first split into the training and test sets; PCA was performed; and 5-fold cross-validation was done on the training set. A literature search for meta-analyses was performed to identify prominent brain regions associated with reading/RD. These were used to repeat the model-training procedure first using whole-brain FC data and then several more times using smaller sets of ROIs (reduced-ROI models).

The ROIs present in the highest-ranked connections for the whole-brain model came from the right dorsal attention and somatosensory motor networks and bilateral visual and temporal regions (Fig 1). Interhemispheric connections were more heavily weighted than intrahemispheric ones. Of the most frequent ROIs in the top 10% of connections, 53.9% were from the right hemisphere; when top ROIs were restricted to a frequency >2 SDs above the median, 81.3% were from the right hemisphere-particularly right dorsal attention and somatosensory motor networks.

The relationship between predicted age and true age was significant in all models (ps < .001). Comparing all models using Akaike's information criteria showed that the whole-brain model performed better than all others as expected. There were no main effects of Group on model predictions, however the interaction of Group x Age was significant in all models (ps < .05), indicating differences in model fit based on group (Fig 2).

The whole-brain model was the best predictor of age. Contrary to hypotheses, when trained with reading/language ROIs, model bias based on group did not increase. Rather, model fit varied by group, being better for the ERs/PRs compared to TRs: this effect was largest in the whole-brain model. One possibility is that variability of FC patterns is larger in TRs, hindering the ability of our model to predict age for the TRs. Finally, the difference in explained variance between the whole-brain model (R² = .343) and the most-reduced model (R² = .155) is small relative to the difference in data used for training. While training with whole-brain data results in more accurate age estimates, accuracy is not linearly proportional to the number of ROIs/connections used to generate features. Markers of age and reading ability are present in diffuse connectivity patterns with sizable redundancy.

Neurodevelopmental/ Early Life (eg. ADHD, autism) ¹

Reading and Writing ²

Connectivity (eg. functional, effective, structural)

fMRI Connectivity and Network Modeling

Computational Neuroscience

DISORDERS

FUNCTIONAL MRI

Machine Learning

Other - dyslexia; reading ability