Using Manifold Learning to Uncover the Embedded Brain and Implications for Mental Health in Youth

1482 

Abstract Submission 

Erica Busch¹, May Conley¹, Arielle Baskin-Sommers¹

¹Yale University, New Haven, CT

Erica Busch  
Yale University
New Haven, CT

May Conley  
Yale University
New Haven, CT

Arielle Baskin-Sommers  
Yale University
New Haven, CT

Advancing the study of biosocial transactions, which considers the interplay between youth and environment, is essential to improving our understanding of how information processing and neural differences contribute to the formation and maintenance of mental health problems. To move the study of biosocial transactions beyond what we have achieved so far (e.g., a complex account of contexts and behaviors without understanding the neurobiological embedding within these transactions), we need to find ways of studying the complex, transactional relationships between neurocognitive functioning and the social world (Viding et al. 2023). Manifold learning techniques can discover and highlight latent structure from high-dimensional, noisy biomedical data, such as fMRI. Here, we develop and apply a novel manifold learning technique to capture the interplay between various environmental factors, neuroimaging data, and mental health outcomes in youth.

We used data from 5,245 9–10 year old youth enrolled in the Adolescent Brain and Cognitive Development Study^SM (ABCD Study®) Data Release 4.0 to ask how brain activation during an emotional n-back task and measures of environmental adversity interact and reflect emotional and behavioral problems. We extracted activation patterns from parcels including the amygdala and the executive control network during cognitive processing (i.e., the contrast of 2-back vs. 0-back and of emotional and neutral task blocks) (Conley et al. 2023). fMRI activation maps contain high levels of noise and intersubject variance. Recent work shows that manifold learning unveils meaningful structure from biomedical data (Moon et al. 2019) and fMRI activity (Huang & Busch et al. 2022). Explicitly modeling additional signals like temporal dynamics in the manifold calculation improves insight into the geometry of fMRI activity and its relation to behavior (Busch et al. 2023). We modified the manifold learning method T-PHATE, designed for fMRI timeseries, to combine multivariate environmental features with neural manifold geometry to embed youth brain data in a latent space via approach we call "Feature PHATE (F-PHATE)" (Fig. 1). We then used the coordinates embedded coordinates of each participants' data to predict their Child Behavior Checklist (CBCL) total problem, externalizing, and internalizing scores. We benchmarked prediction using F-PHATE embeddings against embeddings in a standard embedding model without the environmental features (PHATE) and from the full voxel-resolution data.

First, we validated that embeddings with standard PHATE captured cognitive processing by predicting n-back performance. Linear regression trained on PHATE embeddings significantly outperformed those trained on the full voxel resolution data (amygdala and control network, Fig 2A). Next, we predicted CBCL scores from PHATE and F-PHATE embeddings, and the voxel resolution data for comparison (Fig. 2B). In the amygdala, neither PHATE embeddings nor voxel resolution data predict CBCL externalizing (Fig 2C) or internalizing scores (Fig 2D), though the PHATE embedding does predict total problem scores. PHATE embeddings of the executive control network predict both total problem and externalizing scores (but not internalizing scores) better than the voxel data. With the addition of environment information, prediction performance for F-PHATE exceeds that of either standard PHATE or the voxel data.

We present a manifold-learning approach to incorporating environmental or other information in a low-dimensional embedding space. Embedding participants into a joint neural-and-environmental manifold uncovered latent structure predictive of broad emotional and social problems and externalizing behavior. This work holds important implications for understanding the relationship between biosocial transactions and mental health.

Psychiatric (eg. Depression, Anxiety, Schizophrenia)

Early life, Adolescence, Aging ²

Classification and Predictive Modeling ¹

Methods Development

BOLD fMRI

Cognition

Computational Neuroscience

Data analysis

Development

Machine Learning

Open Data