Integrating Multimodal Neuroimaging Features to Predict Working Memory and Psychiatric Disability

1124 

Abstract Submission 

Catherine Walsh^1,2, Jean-Baptiste Pochon¹, Agatha Lenartowicz¹, Sandra Loo¹, Catherine Sugar¹, Carrie Bearden¹, Robert Bilder¹, Jesse Rissman¹

¹University of California, Los Angeles, Los Angeles, CA, ²National Institute of Mental Health, Bethesda, MD

Catherine Walsh, PhD  
University of California, Los Angeles|National Institute of Mental Health
Los Angeles, CA|Bethesda, MD

Jean-Baptiste Pochon, PhD  
University of California, Los Angeles
Los Angeles, CA

Agatha Lenartowicz, PhD  
University of California, Los Angeles
Los Angeles, CA

Sandra Loo, PhD  
University of California, Los Angeles
Los Angeles, CA

Catherine Sugar, PhD  
University of California, Los Angeles
Los Angeles, CA

Carrie Bearden, PhD  
University of California, Los Angeles
Los Angeles, CA

Robert Bilder, PhD  
University of California, Los Angeles
Los Angeles, CA

Jesse Rissman, PhD  
University of California, Los Angeles
Los Angeles, CA

Individual differences in working memory (WM) capacity have been linked to variation in higher order cognition and psychiatric disability. One goal of the NIH Research Domain Criteria (RDoC) paradigm is to formally characterize this relationship and relate neurocognitive markers of WM to psychopathology across broad diagnostic categories [1].
Although a variety of structural and functional brain measures have been shown to account for variance in WM capacity, recent work integrating distinct modalities of brain structure and function explain more variance than can any single modality on its own [2,3]. The present investigation sought to leverage machine learning to characterize the relative importance of different functional and structural neuroimaging measures for predicting WM task performance, trait-level WM capacity, and overall psychiatric disability.

Data from 169 participants (106 females, age 21-40) were collected as a part of a project recruiting adults with a wide range of mental health concerns, as well as non-care-seeking adults. Participants underwent behavioral testing to measure cognitive ability and self-reported clinical symptomatology, as measured by the WHO Disability Assessment Scale (WHODAS) and the Brief Psychiatric Rating Scale (BPRS). Indices of WM were submitted to an exploratory factor analysis, producing visual and verbal WM capacity factor scores.
All participants underwent whole-brain imaging using a Siemens 3.0T MRI scanner. Functional data collection (multiband EPI, TR=1.5s, TE=34.2ms, voxel resolution = 2 mm3) included a resting state scan, a face localizer [4], and a delayed facial recognition WM task (DFR; [5]). Structural data collection included a high-resolution (1 mm3) T1-weighted scan and a diffusion-weighted scan.
We identified 3 modalities of neuroimaging features to use as predictors in our models: task-based fMRI from the DFR task (univariate BOLD parameter estimates and representational similarity-based indices of face-specific representational content and encoding-to-delay pattern stability), resting state fMRI (mean connectivity across networks; global and local graph theoretical measures), and structural measures (cortical thickness and subcortical volume).
Separate ElasticNet models were built for each modality, in addition to models for each task-fMRI measure and a full model including predictors from all modalities. Model were trained and tested with a 10-fold cross-validation framework. Spearman correlation was used to index explained variance, and permutation testing was used to determine statistical significance for overall models and individual predictors.

Different patterns of modalities were able to predict each outcome measure (Fig 1a). Task fMRI models significantly predicted task performance, visual WM capacity, and BPRS. Resting state fMRI models predicted task performance and WHODAS, while structural MRI models could only predict WHODAS. Decomposing task fMRI models into individual measures revealed a striking dissociation, where task performance was most predicted by univariate BOLD and encoding to delay pattern stability measures, while visual WM capacity was best predicted by the strength of face-specific representational content (Fig 1b). Combined multimodal models were able to predict a significant variance in most measures, especially for WHODAS where the multimodal modal model outperformed the unimodal models (Fig 1c).

Predictive modeling revealed unique contributions of structural and functional MRI in predicting WM task performance, trait-level WM capacity, and psychiatric disability. Better trait visual WM capacity was associated with high-fidelity representation of task-relevant features during maintenance, while WM task performance was predicted by neural persistence of encoding processes. Psychiatric disability was predicted by task-related functional activity measures. These potential neural biomarkers could inspire novel targeted interventions.

Psychiatric (eg. Depression, Anxiety, Schizophrenia)

Working Memory ¹

Classification and Predictive Modeling ²

Cognition

FUNCTIONAL MRI

Machine Learning

Memory

Multivariate

Psychiatric Disorders

Other - Working Memory