Hemispheric asymmetry in Alzheimer's Disease patients as feature for pathogenesis and prediction

293 

Abstract Submission 

Laura Broderius¹, Shammi More², Kaustubh Patil¹, Kathrin Reetz³, Patrick Friedrich⁴, Felix Hoffstaedter¹

¹Institute of Neuroscience and Medicine (INM-7: Brain and Behaviour), Research Centre Jülich, Jülich, NRW, ²Juelich Research Center, Juelich, Germany, ³Department of Neurology, University Hospital RWTH Aachen, Aachen, NRW, ⁴Forschungszentrum Jülich, Jülich, Germany

Laura Broderius  
Institute of Neuroscience and Medicine (INM-7: Brain and Behaviour), Research Centre Jülich
Jülich, NRW

Shammi More  
Juelich Research Center
Juelich, Germany

Kaustubh Patil  
Institute of Neuroscience and Medicine (INM-7: Brain and Behaviour), Research Centre Jülich
Jülich, NRW

Kathrin Reetz  
Department of Neurology, University Hospital RWTH Aachen
Aachen, NRW

Patrick Friedrich  
Forschungszentrum Jülich
Jülich, Germany

Felix Hoffstaedter  
Institute of Neuroscience and Medicine (INM-7: Brain and Behaviour), Research Centre Jülich
Jülich, NRW

Hemispheric asymmetries in age-related atrophy are well documented in healthy aging and in Alzheimer's Disease (AD) with asymmetry seeming to develop differently in Alzheimer's cases than in normally aging brains (1-4). Our aim was to investigate the potential of those differences as biomarkers for the detection of AD and mild cognitive impairment (MCI).

For the asymmetry analysis, 3T structural MRI scans from the Alzheimer's Disease Neuroimaging Initiative (ADNI) (5) were analysed in five diagnostic groups: patients with AD (N=116, mean age =62,1 (SD=6,7)), early MCI (EMCI) (N=205, mean age=60,8 (SD=5,1)), unsubdivided MCI (N=74, mean age=63,2 (SD=6,8)), late MCI (LMCI, N=105, mean age=62,9 (SD=5,3)) and cognitive normal (CN) subjects (N=198, mean age=63,4 (SD=5,43)). First, a symmetrical Shooting template was created from the IXI dataset (https://brain-development.org/ixi-dataset/) using the CAT12 Toolbox (6) in SPM12. Subsequently, the ADNI T1w images were preprocessed with default settings using the symmetrical IXI template. To investigate gray matter volume (GMV) based hemispheric asymmetry, we calculated the asymmetry index (7) and tested for univariate differences between groups using GLMs and TFCE with non-parametric FWE correction. Additionally, hemipheres were classified as left or right using the following workflow (8): 1. split images into hemispheres and alignment to the right side, 2. perform supervised RandomForest classification per group using GMV voxels as features and 3. use the Boruta algorithm to identify relevant features for hemisperic classifications6. To test those Boruta features fordetection of AD and MCI, we performed binary classifications using Julearn (9), a Python based machine learning library based on scikit-learn (10). We used a support vector machine (SVM) in a nested 5-fold-cross-validation with hyperparameter tuning (kernel = linear, c = [0.0001, 0.001, 0.01, 0.1]) with age and sex linearly modelled as confounds without data leakage.

In the univariate asymmetry analysis no significant differences were found between diagnostic groups. The Boruta feature selection analysis for hemisphere classification identified thalamus, amygdala, insula, parietal operculum and putamen as biggest clusters in MCI and AD. In CN and early MCI, also thalamus and amygdala contributed relevantly to hemispheric classification, alongside the entorhinal cortex and the hippocampus. The entorhinal cortex and the motor cortex nearly disappear in the clusters of MCI, LMCI and AD, possibly becoming more symmetrical with disease progression, while the parietal operculum appears to become more asymmetrical. In general, hemisphere identifying clusters seem to become smaller and more scattered in AD. Successful classification of AD vs. CN was possible with a similar performance for the whole brain as well as only using the sparse Boruta regions with a test score > 80%. Classification of MCI vs. CN for the whole brain and Boruta regions showed a test score ~80% (Fig. 2). Of note, using the asymmetry indices as features performed considerably worse with test scores <65%.

Hemispheric classifications using Boruta identified clusters as relevant for the decision between left vs. right and are therefore related to hemispheric asymmetry, which differs between diagnostic groups of the AD continuum. In AD Boruta regions are scattered more globally, which might be related to the decreasing structural integrity of brain tissue in AD. The performance of those regions in disease classifications was very similar to the whole brain even though they only contained <1.5% of the features relative to the whole brain. This shows strong potential for hemisphere-determining regions in the prediction of AD as well as MCI in the ADNI dataset.

Neurodegenerative/ Late Life (eg. Parkinson’s, Alzheimer’s) ¹

Classification and Predictive Modeling

Image Registration and Computational Anatomy

Multivariate Approaches ²

Segmentation and Parcellation

Computational Neuroscience

Computing

Data analysis

Degenerative Disease

Hemispheric Specialization

Machine Learning

MRI

Multivariate

Open-Source Software

Univariate