Investigating the neurobiological basis of psychopathology using bi-factor models: reliably general

544 

Abstract Submission 

Martin Gell^1,2, Mauricio Hoffmann^3,4, Tyler Moore⁵, Aki Nikolaidis⁶, Ruben Gur⁵, Giovanni Salum^4,7,8, Michael Milham^6,9, Simon Eickhoff^10,2, Robert Langner^10,2, Veronika Müller^10,2, Theodore Satterthwaite⁵

¹Department of Psychiatry, Psychotherapy, and Psychosomatics, RWTH Aachen University Hospital, Aachen, Germany, ²Research Centre Juelich, Institute of Neuroscience and Medicine (INM7), Juelich, Germany, ³Department of Neuropsychiatry, Universidade Federal de Santa Maria, Santa Maria, Brazil, ⁴Universidade Federal do Rio Grande do Sul, Porto Alegre, Brazil, ⁵University of Pennsylvania, Philadelphia, PA, ⁶Child Mind Institute, New York, NY, ⁷National Institute of Developmental Psychiatry for Children and Adolescents (INCT-CNPq), São Paulo, Brazil, ⁸Child Mind Institute, New York, United States, ⁹Nathan S. Kline Institute for Psychiatric Research, Orangeburg, NY, ¹⁰Institute for Systems Neuroscience, Medical Faculty, Heinrich-Heine University Düsseldorf, Düsseldorf, North Rhine-Westphalia

Martin Gell  
Department of Psychiatry, Psychotherapy, and Psychosomatics, RWTH Aachen University Hospital|Research Centre Juelich, Institute of Neuroscience and Medicine (INM7)
Aachen, Germany|Juelich, Germany

Mauricio Hoffmann  
Department of Neuropsychiatry, Universidade Federal de Santa Maria|Universidade Federal do Rio Grande do Sul
Santa Maria, Brazil|Porto Alegre, Brazil

Tyler Moore  
University of Pennsylvania
Philadelphia, PA

Aki Nikolaidis  
Child Mind Institute
New York, NY

Ruben Gur  
University of Pennsylvania
Philadelphia, PA

Giovanni Salum  
Universidade Federal do Rio Grande do Sul|National Institute of Developmental Psychiatry for Children and Adolescents (INCT-CNPq)|Child Mind Institute
Porto Alegre, Brazil|São Paulo, Brazil|New York, United States

Michael Milham  
Child Mind Institute|Nathan S. Kline Institute for Psychiatric Research
New York, NY|Orangeburg, NY

Simon Eickhoff  
Institute for Systems Neuroscience, Medical Faculty, Heinrich-Heine University Düsseldorf|Research Centre Juelich, Institute of Neuroscience and Medicine (INM7)
Düsseldorf, North Rhine-Westphalia|Juelich, Germany

Robert Langner  
Institute for Systems Neuroscience, Medical Faculty, Heinrich-Heine University Düsseldorf|Research Centre Juelich, Institute of Neuroscience and Medicine (INM7)
Düsseldorf, North Rhine-Westphalia|Juelich, Germany

Veronika Müller  
Institute for Systems Neuroscience, Medical Faculty, Heinrich-Heine University Düsseldorf|Research Centre Juelich, Institute of Neuroscience and Medicine (INM7)
Düsseldorf, North Rhine-Westphalia|Juelich, Germany

Theodore Satterthwaite  
University of Pennsylvania
Philadelphia, PA

Despite sustained efforts to uncover the neurobiological basis of mental health disorders, finding links to specific syndromes using neuroimaging has remained challenging. Factors contributing to this include both high comorbidity among disorders and heterogeneity within diagnostic categories1. In response, models that separate the shared (or general, transdiagnostic) and the unique (or specific) dimensions of psychopathology2 have become an important area of research. Such bi-factor models offer the possibility to disentangle the general and unique biological underpinnings of psychopathology. However, to investigate associations between individual differences in neurobiology and psychopathology such dimensions also require sufficient test-retest reliability, as it sets an upper bound on effect size3 and determines signal-to-noise ratio in machine learning analyses4. Here we evaluated the reliability of 11 published bi-factor models in two large developmental datasets from the global north and south and investigated the impact of reliability on their associations with brain structure and function.

We used items from the Child Behaviour Checklist (CBCL) to fit 11 bi-factor models in the ABCD5 (n = 5526, ages 9-10 at baseline, mean retest interval = 12 months) and BHRC6 (n = 772, ages 6-14 at baseline, mean retest interval = 8 months) datasets. CBCL items were configured to load on a single general factor and a varying number of specific factors as defined in all published models7. Test-retest reliability was then assessed using correlation, after regressing out participant age, timepoint and their interaction. Sensitivity analyses were performed using alternative reliability metrics (hierarchical omega and factor determinacy) and shorter retest intervals. Next, we used resting-state functional connectivity (FC) and cortical thickness (CT) to predict general and specific factors as well as summary scores in the ABCD dataset. Predictions were performed with linear ridge regression within nested 2-fold cross-validation using matched discovery (n=3,525) and replication (n=3,447) samples8. Using Mplus, bi-factor models were fit separately in each sample to avoid data leakage and improve replicability. Preprocessed resting-state fMRI data and CT maps were acquired from the ABCD BIDS Community Collection9 and parcellated using the Glasser atlas10. FC was then calculated using Pearson correlation between the time series of all parcels.

For all 11 models, the general "P" factor captured most variance across CBCL items (ABCD = 57-78%; BHRC = 60-79%) and had generally higher reliability (test-retest rABCD = 0.7 - 0.76; rBHRC = 0.55 - 0.59) than specific factors (rABCD = 0.36 - 0.61; rBHRC = 0.14 - 0.47) in both datasets (Fig. 1). Despite the favourable psychometric properties, p-factors could be predicted using FC (Fig. 2A) or CT with a comparable prediction accuracy to externalising and attention factors. Notably, p-factors also showed equivalent prediction accuracy and reliability to many standard CBCL summary scores (esp. total problems, attention, internalising and rule-breaking), which combine all sources of variance and ignore the multidimensional structure parsed by the bi-factor models (Fig. 2B).

Here we demonstrate that while capturing most of the reliable variance in CBCL, general psychopathology factors showed comparable associations with brain structure and function to specific factors and summary scores. Across all bi-factor models, many specific factors displayed low test-retest reliability that will substantially attenuate associations and make comparisons between factors hard to interpret. These results suggest that deeper phenotyping is necessary to better characterise the variance unique to specific dimensions. Finally, while bi-factor models are better suited to address phenotypic complexity and heterogeneity in psychopathology, general factors exhibit predictive utility comparable to that of the CBCL total summary score.

Psychiatric (eg. Depression, Anxiety, Schizophrenia) ¹

Classification and Predictive Modeling ²

Connectivity (eg. functional, effective, structural)

Data analysis

Design and Analysis

Development

FUNCTIONAL MRI

Machine Learning

Psychiatric

Psychiatric Disorders

Statistical Methods

STRUCTURAL MRI

Other - Reliability