Predicting Mental and Neurological Illnesses Using Cerebellar Heterogeneity

1398 

Abstract Submission 

Milin Kim¹, Nitin Sharma², Esten Leonardsen³, Saige Rutherford⁴, Geir Selbæk⁵, Karin Persson⁵, Nils Eiel Steen⁶, Olav Smeland⁷, Torill Ueland⁸, Geneviève Richard⁷, Aikaterina Manoli⁹, Sofie Valk⁹, Christian Beckmann¹⁰, Andre Marquand¹¹, Ole Andreassen¹², Lars Westlye¹³, Thomas Wolfers¹⁴, Torgeir Moberget¹⁵

¹Norwegian Centre for Mental Disorders Research (NORMENT),Oslo University Hospital & Institute of Cli, Oslo, Oslo, ²University of Tübingen;German Center for Mental Health (DZPG), Tübingen, Germany, ³NORMENT; University of Oslo, Oslo, Norway, ⁴Radboud University, Nijmegen, Netherlands, ⁵Oslo University Hospital; The Norwegian National Centre for Ageing and Health, Oslo, Norway, ⁶NORMENT, Division of Mental Health and Addiction, Oslo University Hospital & University of Oslo, Oslo, Norway, ⁷Norwegian Centre for Mental Disorders Research (NORMENT), Oslo, Norway, ⁸Norwegian Centre for Mental Disorders Research (NORMENT);University of Oslo, Oslo, Norway, ⁹Max Planck Institute for Human Cognitive and Brain Sciences, Leipzig, Germany, ¹⁰Donders Institute for Brain, Cognition, and Behavior, Radboud University Nijmegen, Nijmegen, NL, Nijmegen, Netherlands, ¹¹Radboud University Nijmegen, Nijmegen, Gelderland, ¹²NORMENT, Oslo, Norway, ¹³Norwegian Centre for Mental Disorders Research (NORMENT), Oslo University Hospital, Oslo, Norway, ¹⁴Laboratory for Mental Health Mapping, University of Tübingen, Tübingen, BW, ¹⁵Norwegian Centre for Mental Disorders Research (NORMENT); OsloMet, Oslo, Norway

Milin Kim  
Norwegian Centre for Mental Disorders Research (NORMENT),Oslo University Hospital & Institute of Cli
Oslo, Oslo

Nitin Sharma  
University of Tübingen;German Center for Mental Health (DZPG)
Tübingen, Germany

Esten Leonardsen  
NORMENT; University of Oslo
Oslo, Norway

Saige Rutherford  
Radboud University
Nijmegen, Netherlands

Geir Selbæk  
Oslo University Hospital; The Norwegian National Centre for Ageing and Health
Oslo, Norway

Karin Persson  
Oslo University Hospital; The Norwegian National Centre for Ageing and Health
Oslo, Norway

Nils Eiel Steen  
NORMENT, Division of Mental Health and Addiction, Oslo University Hospital & University of Oslo
Oslo, Norway

Olav Smeland  
Norwegian Centre for Mental Disorders Research (NORMENT)
Oslo, Norway

Torill Ueland  
Norwegian Centre for Mental Disorders Research (NORMENT);University of Oslo
Oslo, Norway

Geneviève Richard  
Norwegian Centre for Mental Disorders Research (NORMENT)
Oslo, Norway

Aikaterina Manoli  
Max Planck Institute for Human Cognitive and Brain Sciences
Leipzig, Germany

Sofie Valk  
Max Planck Institute for Human Cognitive and Brain Sciences
Leipzig, Germany

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

Andre Marquand  
Radboud University Nijmegen
Nijmegen, Gelderland

Ole Andreassen  
NORMENT
Oslo, Norway

Lars Westlye  
Norwegian Centre for Mental Disorders Research (NORMENT), Oslo University Hospital
Oslo, Norway

Thomas Wolfers  
Laboratory for Mental Health Mapping, University of Tübingen
Tübingen, BW

Torgeir Moberget  
Norwegian Centre for Mental Disorders Research (NORMENT); OsloMet
Oslo, Norway

The cerebellum has been associated not only with motor coordination but also with cognitive and emotional processing, extending its relevance to a broad spectrum of clinical conditions. As the case-control model overlooks individual-level variability, normative models have been used to map normative development and aging across the lifespan. The normative model is analogous to a paediatric growth chart used to examine how individuals develop and age. Therefore, we predicted mental and neurological illnesses, including autism spectrum disorder (ASD), mild cognitive impairment (MCI), Alzheimer's disease (AD), bipolar disorder (BD), and schizophrenia (SZ), using the cerebellar lobular and voxel-wise normative modelling features. By assessing the predictive performance of the machine learning models, we investigated the impact of various feature categories.

We employed cerebellar volume and voxel-wise normative models that were introduced in (Kim et al., 2023) which trained on more than 27k individuals across 132 scanning sites, with an age range spanning from 3 to 85 years. The test set comprised more than 26k individuals without a diagnosis, along with clinical sets of 1,757 (Figure 1A). The cerebellar volume normative model utilised ACAPULCO (Han et al., 2020) algorithm, a convolutional neural network-based algorithm that divides the cerebellum into 28 lobules and voxel-wise model used SUIT (Spatially Unbiased Infratentorial Toolbox) (Diedrichsen et al., 2009) toolbox (Figure 1B). The spatial precision normative models enabled integration with existing cerebellar atlases, including 28 anatomical cerebellar regions, 10 regions of interest from the multi-domain task battery (MDTB) (King et al., 2019), and 17 regions of interest from resting-state connectivity (Buckner et al., 2011; Yeo et al., 2011). Here, we developed different machine learning models using raw scores, median and percentage of extreme deviation (threshold at |z| > 1.96) from the normative model mapped onto existing atlases as features (Figure 1C). Deviation score assess how an individuals deviate from individuals without diagnosis at each lobule or voxel in the cerebellum. We assessed the diagnostic accuracy using the area under the receiver operating characteristic curve (ROAUC) with matched samples of the sites and age, addressing the imbalanced dataset with the synthetic minority oversampling technique (SMOTE) (Chawla et al., 2002) which were compared using 1000 permutations (Figure 1D). The models were explained by employing Shapley Additive Explanations to underscore the influence of each individual feature.

In our investigation, the voxel-wise models yielded slightly higher ROAUC performance compared to volumetric models (Figure 2A). The ROAUC values that surpassed chance levels are displayed in the figure. In the voxel-wise models, methods that used extreme negative percentage of deviations as features outperformed extreme positive percentage of deviations in the all three atlases. Notably, ASD, MCI and SCZ model predictions were better than BD and AD. Shapley Additive Explanation (SHAP) showed that regions important for predictions, where Figure 2B displays the extreme deviation of SHAP values.

We discovered that voxel-wise models that utilize percentage of extreme negative deviation exhibited superior performance in classifying mental and neurological disorders, while across atlases showed similar performance. We demonstrated the advantages of implementing normative modelling features in the classification. This underscores the constraints posed by anatomical boundaries of the cerebellum and highlights the importance of employing a functional map of the cerebellum.

Neurodegenerative/ Late Life (eg. Parkinson’s, Alzheimer’s)

Neurodevelopmental/ Early Life (eg. ADHD, autism)

Psychiatric (eg. Depression, Anxiety, Schizophrenia) ²

Early life, Adolescence, Aging

Classification and Predictive Modeling ¹

Aging

Autism

Cerebellum

Computational Neuroscience

Degenerative Disease

Machine Learning

Psychiatric Disorders

Schizophrenia

STRUCTURAL MRI

Other - Normative modelling