Scalable and accessible normative modelling for cross-cultural generalizability in neuropsychiatry

1457 

Abstract Submission 

Pieter Barkema¹, Saige Rutherford², Christian Beckmann³, Andre Marquand⁴

¹Queen Square Institute of Neurology, University College London, London, United Kingdom, ²Radboud University, Nijmegen, Netherlands, ³Donders Institute for Brain, Cognition, and Behavior, Radboud University Nijmegen, Nijmegen, NL, Nijmegen, Netherlands, ⁴Radboud University Nijmegen, Nijmegen, Gelderland

Pieter Barkema  
Queen Square Institute of Neurology, University College London
London, United Kingdom

Saige Rutherford  
Radboud University
Nijmegen, Netherlands

Christian Beckmann  
Donders Institute for Brain, Cognition, and Behavior, Radboud University Nijmegen, Nijmegen, NL
Nijmegen, Netherlands

Andre Marquand  
Radboud University Nijmegen
Nijmegen, Gelderland

The study of individual variability through normative modelling has recently shown to have great potential for understanding the mysterious neurobiology of mental disorders (A. F. Marquand et al., 2019; Rutherford et al., 2023). Accurate modeling of variability across the lifespan, however, requires processing of neuroimaging data from tens of thousands of participants, large amounts of compute power to estimate models, and labour and expertise to implement, which provides a barrier for many users of these models.

To address these issues, we recently introduced PCNportal (https://pcnportal.dccn.nl/), allowing free and online access to pre-estimated normative models, requiring no compute power, code or domain expertise (Barkema et al., 2023) – using the well-validated PCNtoolkit library as a modelling backend (A. Marquand et al., 2021). We demonstrate this framework and provide a showcase application to evaluate the global usefulness of PCNportal in context of cross-ethnical generalizability.

A key feature of PCNportal is how easily models can be added and accessed across the world. On PCNportal's launch we provided access to normative models for cortical thickness and brain volume, and now expand our repertoire to include models trained on resting-state fMRI data using multiple parcellations, and additional models for cortical thickness, surface area and cerebellar volume. Can these models, however, generalize to cohorts of different ethnic backgrounds? The largest neuroimaging cohorts available are strongly biased towards white ethnicity, such as UK Biobank, ABCD and HCP (Fry et al., 2017; Ricard et al., 2023), and several studies suggest that bias may confound studies (Chee et al., 2011; Tang et al., 2018). We test cross-ethnical generalizability of one ethnically biased model by applying it to an East Asian cohort.

Data
We use the SRPBS-database (Tanaka et al., 2021), containing structural MRI data of 150 Image-Derived Phenotypes (IDPs) for 993 patients and 1421 controls from fourteen collection sites in Japan. We split the data into an adaptation and test set (50/50) ensuring an equal split for each collection site.

Model
We use a Bayesian Linear Regression (BLR) model trained on average cortical thickness measures from 58,836 data points from 82 sites (Rutherford et al., 2022) – a model with white ethnical bias. Importantly, this model is hosted on PCNportal and was similarly used in Barkema et al (2023).

Analysis
We analyze the deviations in cortical thickness and compare between the schizophrenia group (N=68) and the control group (N=468). We apply our biased model to obtain individualized deviation scores, modelling age, sex and sites as covariates.

We investigated cross-ethnical generalizability of PCNportal by applying an ethnically biased model to a predominantly East Asian cohort, with a good model fit (Figure 1). We tested whether absolute deviation scores are greater in the schizophrenia group than the control group by using a one-sided Wilcoxon rank sums test and found a significant effect (statistic=2.383, pvalue=0.009). This analysis replicates a well-established finding and shows that models have the potential to generalize to other ethnic groups, despite having a dangerous ethnic training bias.

We showcase the PCNportal tool for online normative modelling of neuroimaging data. We release four extra normative models, trained on fMRI resting-state data, cerebellar volume or average thickness. We also bring attention to an important issue of easy global access by illustrating cross-ethnical generalizability. We replicate a well-established effect of cortical thickness abnormalities in schizophrenia, in a cohort of primarily East Asian ethnicity. We hope this analysis inspires other cross-ethnical generalizability initiatives. PCNportal can tackle ethnical neuroimaging bias through facilitating global model and data contributions, aiming to build a more inclusive future in neuropsychiatric precision medicine.

Psychiatric (eg. Depression, Anxiety, Schizophrenia)

Classification and Predictive Modeling ¹

Informatics Other ²

Data analysis

Informatics

Open-Source Software

Psychiatric Disorders

Statistical Methods

Other - Normative modelling