Predicting regional glucose metabolism from structural vs. functional connectivity
Poster No:
2599
Submission Type:
Abstract Submission
Authors:
Yifan Mayr1,2, Aldana Lizarraga1, Arianna Sala3,1, Igor Yakushev1,2
Institutions:
1Technical University of Munich, Department of Nuclear Medicine, Munich, Germany, 2Graduate School of Systemic Neurosciences GSN-LMU, Munich, Germany, 3Université De Liège, Liege, Belgium
First Author:
Yifan Mayr
Technical University of Munich, Department of Nuclear Medicine|Graduate School of Systemic Neurosciences GSN-LMU
Munich, Germany|Munich, Germany
Technical University of Munich, Department of Nuclear Medicine|Graduate School of Systemic Neurosciences GSN-LMU
Munich, Germany|Munich, Germany
Co-Author(s):
Arianna Sala
Université De Liège|Technical University of Munich, Department of Nuclear Medicine
Liege, Belgium|Munich, Germany
Université De Liège|Technical University of Munich, Department of Nuclear Medicine
Liege, Belgium|Munich, Germany
Igor Yakushev
Technical University of Munich, Department of Nuclear Medicine|Graduate School of Systemic Neurosciences GSN-LMU
Munich, Germany|Munich, Germany
Technical University of Munich, Department of Nuclear Medicine|Graduate School of Systemic Neurosciences GSN-LMU
Munich, Germany|Munich, Germany
Introduction:
Operation of any communication system requires energy input. Both structural and functional connectivity of the brain were recently shown to be positively linked to its energy consumption, but the strength of this link remains unclear. Here, we directly compared structural connectivity (SC) and functional connectivity (FC) in their ability to predict glucose metabolism of the brain.
Methods:
We analyzed diffusion tensor imaging (DWI), and fluorodeoxyglucose (FDG) positron emission tomography (PET) data of 55 healthy, middle-aged individuals. Images were spatially parcellated into 106 brain regions. FDG uptake was intensity normalized by the mean uptake in the whole brain gray matter. SC was measured by probabilistic tractography. FC was estimated from fMRI data. FC weights were computed as Pearson correlation between blood-oxygen-level-dependent signals of each pair of regions. In total, we obtained 55x106=5830 observations of FDG uptake. For each FDG uptake observation, 106 connection weights were available. Linear regression explores the linear association between connection weights and FDG uptake, nonlinear regression models allow more complex relationships. We tested 5 types of regression models: linear regression, k-nearest neighbors, support vector, extremely randomized trees, and gradient boosting.
We fitted 3 sets of regression models: one set with SC as predictors, another with FC as predictors, and a third set with both SC and FC as predictors. Each set included the 5 regression models listed above. The complete dataset was split into a training (80%) and a test set (20%), stratified by brain region. All regression models were trained using the same training set, with 5-fold cross validation repeated 5 times. Mean training R2 scores and root mean squared error (RMSE), and standard deviation were reported for each model. Test R2 score and RMSE were evaluated on the test set using the best regression model.
We fitted 3 sets of regression models: one set with SC as predictors, another with FC as predictors, and a third set with both SC and FC as predictors. Each set included the 5 regression models listed above. The complete dataset was split into a training (80%) and a test set (20%), stratified by brain region. All regression models were trained using the same training set, with 5-fold cross validation repeated 5 times. Mean training R2 scores and root mean squared error (RMSE), and standard deviation were reported for each model. Test R2 score and RMSE were evaluated on the test set using the best regression model.
Results:
Up to 75% of the variance in the brain's relative FDG uptake was explained by SC, but only 30% by FC. Across all regressors, the SC models performed substantially better than FC models: the mean R2 of each SC model is more than double that of its counterpart FC model (Figure 1), and the mean RMSE of each SC model is lower than its counterpart FC model (Figure 2). The test R2 score and RMSE were similar to the training scores. Additionally, including FC predictors did not enhance model performance achievable by SC alone.
Conclusions:
SC predicts glucose metabolism more accurately and strongly than FC, in both linear and nonlinear models, regardless of the model complexity. FC does not add to the predictive power when combined with SC. Our results indicate that brain energy demands are closer linked to its white matter infrastructure than to its functional communication.
Modeling and Analysis Methods:
Classification and Predictive Modeling
Connectivity (eg. functional, effective, structural) 2
PET Modeling and Analysis
Physiology, Metabolism and Neurotransmission :
Cerebral Metabolism and Hemodynamics 1
Keywords:
CHEMOARCHITECTURE
Data analysis
NORMAL HUMAN
Positron Emission Tomography (PET)
Tractography
WHITE MATTER IMAGING - DTI, HARDI, DSI, ETC
1|2Indicates the priority used for review