Parcel-wise stacking ensemble provides improved age prediction and brain-aging insights

1155 

Abstract Submission 

Georgios Antonopoulos^1,2, Shammi More^1,2, Federico Raimondo^1,2, Simon Eickhoff^1,2, Kaustubh Patil^1,2

¹Institute of Neuroscience and Medicine (INM-7: Brain and Behaviour), Research Centre Jülich, Juelich, Germany, ²Institute for Systems Neuroscience, Medical Faculty, Heinrich-Heine University Düsseldorf, Düsseldorf, Germany

Georgios Antonopoulos  
Institute of Neuroscience and Medicine (INM-7: Brain and Behaviour), Research Centre Jülich|Institute for Systems Neuroscience, Medical Faculty, Heinrich-Heine University Düsseldorf
Juelich, Germany|Düsseldorf, Germany

Shammi More  
Institute of Neuroscience and Medicine (INM-7: Brain and Behaviour), Research Centre Jülich|Institute for Systems Neuroscience, Medical Faculty, Heinrich-Heine University Düsseldorf
Juelich, Germany|Düsseldorf, Germany

Federico Raimondo  
Institute of Neuroscience and Medicine (INM-7: Brain and Behaviour), Research Centre Jülich|Institute for Systems Neuroscience, Medical Faculty, Heinrich-Heine University Düsseldorf
Juelich, Germany|Düsseldorf, Germany

Simon Eickhoff  
Institute of Neuroscience and Medicine (INM-7: Brain and Behaviour), Research Centre Jülich|Institute for Systems Neuroscience, Medical Faculty, Heinrich-Heine University Düsseldorf
Juelich, Germany|Düsseldorf, Germany

Kaustubh Patil  
Institute of Neuroscience and Medicine (INM-7: Brain and Behaviour), Research Centre Jülich|Institute for Systems Neuroscience, Medical Faculty, Heinrich-Heine University Düsseldorf
Juelich, Germany|Düsseldorf, Germany

Predicting chronological age using structural magnetic resonance imaging (MRI) has shown great potential for studying aging in health and disease. A model trained on MRI scans from healthy individuals can provide biological insights into healthy brain aging and it can also potentially assist in detection of abnormal aging in psychiatric, and neurodegenerative disorders [1, 2]. Such a model may lead to novel monitoring and treatment options. However, both accuracy and explainability of age prediction models need to be improved before they can be applied in the real-world. To this end, we propose parcel-wise stacking ensemble models (SEM) [3]. In SEM the voxels in each region are not weighted equally, as in the standard approaches. Regional models evaluate the contribution of each voxel in the process making this way, better use of voxels' information. Moreover, combining predictions in sequential models leads to reduced overall bias and variance, of the final prediction.

We used T1w MRI scans of healthy subjects from 4 open datasets (IXI, eNKI, CamCAN and 1000Gehirne. N>500 each, total N=3103, age range 18-90 years). Voxel based morphometry using CAT12.8 [4] was employed to estimate gray matter volume (GMV) for each subject. Performance was estimated in terms of mean absolute error (MAE) assessed in leave-one-dataset-out (LODO) set up.
Proposed SEM consists of two levels, denoted as L0 and L1 (Figure 1). GLMnet [5] was used for both L0 and L1 models. The L0 uses an 873-parcel atlas and trains a model for each parcel using corresponding voxels-wise GMV. The features for the L1 model are obtained as out-of-sample (OOS) predictions from a 3-fold cross-validation scheme on L0 models. Two types of L0 models were obtained; either by pooling data from different sites (L0p) or by making predictions for each training site separately (L0s). Similarly, in L1 models were obtained by either pooling L0 predictions (L1p) or by training the L1 model for each site separately and averaging the per-site predictions (L1s).
Additionally, to examine the case where enough data is available at an application site, we estimate L0 OOS predictions from that site (L0oos). These are then used to obtain predictions using the L1 models. As a baseline we also trained models using the average GMV in each parcel -by averaging predictions of the site-specific models or training models on pooled data across sites-, while ensuring use of consistent training samples across the set ups.

The highest test performances were observed for the L0oos setups where L0 predictions came from the test site. The best predictions were for the L0oos-L1p (average MAE=4.8), closely followed by the L0oos-L1s (MAE=4.9). Setups using pooled L0 predictions to train L1, independently of how L0 was trained, L0p-L1p and L0s-L1p, had both MAE=5.1. Models using mean parcel-wise GMV had MAE=5.7 when the model was trained using the train sets together with the 2 folds of the test set. Performance was worse for the other two mean parcel-wise GMV models trained in three datasets, with the L1p setup up being slightly better compared to L1s (MAE=6.2 and MAE=6.7 respectively).
L0 models provided robust interpretation of regional aging effects, i.e. the Pearson correlation of real age with OOS predicted-age was higher than with GMV (averaged across all datasets used for training, Figure 2). While there is a considerable overlap in the identified regions between the two methodologies, SEM distinctly emphasizes certain areas, notably the subcortex and cerebellum.

SEM provides improved age prediction performance compared to using parcel-wise average of GMV as well as novel biological insights regarding healthy aging. Further improvements in the SEM design could be achieved by selecting suitable learning algorithms with appropriate hyperparameter tuning for L0 and L1 models.

Aging ¹

Classification and Predictive Modeling

Methods Development ²

Multivariate Approaches

Aging

Machine Learning

MRI

STRUCTURAL MRI