A Matlab implementation for predicting the change of ADAS scores using brain MRI with single and multi-task learning. In the present work, we compared single- and multi-task learning approaches to predict the changes in ADAS-Cog scores based on T1-weighted anatomical magnetic resonance imaging (MRI). In contrast to most machine learning-based methods to predict the changes in ADAS-Cog, we stratified the subjects based on their baseline diagnoses and evaluated the prediction performances in each group.
Maintainer: Vandad Imani, (Email:vandad.imani@uef.fi)
Alzheimer’s Disease (AD) is a chronic neurodegenerative disorder that occurs among the elderly. AD's pathophysiological changes begin many years before clinical manifestations of disease and the spectrum of AD spans from clinically asymptomatic to severely impaired. Because of this, there is an appreciation that AD should not only be viewed with discrete and defined clinical stages but as a multifaceted process moving along a continuum. Therefore, early prediction of disease progression would be a crucial step towards designing proper therapeutic, unburden the health care system, and preventing adverse events caused by AD. Due to this reason predicting ADAS-Cog scores with machine learning in order to monitor and quantify patient conditions has been gaining research interest. These progression models come into two categories:
(1) Single-task learning.
(2) Multi-task learning.
Single-task learning (STL), the disease progression of each individual group is predicted at different time-points independently (Figure A). Multi-task learning (MTL), utilizes the essential similarities among various related tasks to predict disease progression. It improves the generalization performance by solving multiple learning tasks simultaneously while exploiting commonalities and differences across tasks (Figure B).
The pipelines depend on various third-party packages, which should be installed before trying these codes. Examples of how to utilize our pipelines appear in the main.m. The directory paths should be corrected in the main.m before running it. The data used in our study were obtained from Alzheimer's Disease Neuroimaging Initiative. It can be accessed through https://www.adni-info.org (which requires registration and approval by ADNI). The subject RIDs used in our study are available as a CSV file in the RIDs directory.
Install Lasso and elastic-net regularized generalized linear models (Glmnet) from here: Download
Install library for support vector machines (This used in cascade ensemble learning method) from here: Download
Install ComBat harmonization in Matlab from here: Download
Install the MALSAR (Multi-tAsk Learning via StructurAl Regularization) package which includes the MTL learning algorithms from here: Download
The directory containing data can be added to the Matlab path as follows:
Datapath = '/path/to/the/data/folder';In addition, the directory containing the mentioned packages can be added to the Matlab path as follows:
addpath('../glmnet_matlab/');
addpath('../libsvm-3.22/matlab/');
addpath('../ComBatHarmonization/scripts/');
% Multi-Task Learning
addpath('../MALSAR/functions/');
addpath('../MALSAR/utils/');-
We used Elastic-net penalized linear regression (EN) in the learning stage.
- Without Harmonization
% OUTPUT : % Results : 'results','results_NC','results_MCI','results_AD','cor','MAE','cor_s','CorNC','CorMCI','CorAD','MAE_NC','MAE_MCI','MAE_AD' % INPUT : % Datapath : The directory containing data % Harmonization : 0- Without data monization, 1- ComBat harmonization, 2- PLS-based domain adaptation % Covariate : 0- Without considering Age as covariate, 1-With considering Age as covariate %-------------------------------------------------------------------------------------------------------------------------------------------- %-------------------------------------------------------------------------------------------------------------------------------------------- Datapath = '/path/to/the/data/folder'; Harmonization = 0; Covariate = 0; ALL_EN(Datapath,Harmonization,Covariate)
-
Datapath = '/path/to/the/data/folder'; Harmonization = 1; Covariate = 0; % Or 1 ALL_EN(Datapath,Harmonization,Covariate)
-
PLS-Based Domain adaptation+ALL-EN
Datapath = '/path/to/the/data/folder'; Harmonization = 2; Covariate = 0; % Or 1 ALL_EN(Datapath,Harmonization,Covariate)
-
We used Elastic-net penalized linear regression along with SVR (EN+SVR) in the learning stage. Where in the first stage, a PLS-based domain adaptation is performed for each brain region separately ,and then, for each brain region, the prediction task is performed. These predictions are then combined in the stacking framework. We applied AAL- atlas to decompose the gray matter density values into 122 distinct regions. For the prediction of the change of ADAS for each brain region, we utilized support vector regression (with a radial basis function kernel). Finally, the per-region predictions were combined using the elastic-net penalized linear regression.
% OUTPUT : % Results : 'results','results_NC','results_MCI','results_AD','cor','MAE','cor_s','CorNC','CorMCI','CorAD','MAE_NC','MAE_MCI','MAE_AD' % INPUT : % Datapath : The directory containing data % Harmonization : 1- ComBat harmonization, 2- PLS-based domain adaptation % Covariate : 0- Without considering Age as covariate, 1-With considering Age as covariate
-
Datapath = '/path/to/the/data/folder'; Harmonization = 1; Covariate = 0; % Or 1 Cas_EN(Datapath,Harmonization,Covariate)
-
Datapath = '/path/to/the/data/folder'; Harmonization = 2; Covariate = 0; % Or 1 Cas_EN(Datapath,Harmonization,Covariate)
-
In the Multi-task learning we used different multi-task learning methods based on least-squares loss function including multi-task Lasso (Least Lasso), Joint Feature Selection (JFS), Dirty Model (Least Dirty), and Trace-Norm Regularization (Least Trace).
-
- Without Harmonization
% OUTPUT : % Results : 'results','results_NC','results_MCI','results_AD','cor','MAE','cor_s','CorNC','CorMCI','CorAD','MAE_NC','MAE_MCI','MAE_AD', ... 'Fold_NC_H','Fold_MCI_H','Fold_AD_H' % INPUT : % Datapath : The directory containing data % Harmonization : 0- Without data monization, 1- ComBat harmonization, 2- PLS-based domain adaptation % Covariate : 0- Without considering Age as covariate, 1-With considering Age as covariate % Formulations : 0- Least Lasso, 1-Joint feature selection (JFS), 2- Dirty Model, 3- Low rank assumption (LRA) %-------------------------------------------------------------------------------------------------------------------------------------------- %-------------------------------------------------------------------------------------------------------------------------------------------- Datapath = '/path/to/the/data/folder'; Harmonization = 0; Covariate = 0; Formulations = 0; % Or 1 Or 2 Or 3 MTL_fun(Datapath,Harmonization,Covariate,Formulations)
-
Datapath = '/path/to/the/data/folder'; Harmonization = 1; Covariate = 0; % Or 1 Formulations = 0; % Or 1 Or 2 Or 3 MTL_fun(Datapath,Harmonization,Covariate,Formulations)
-
PLS-Based Domain adaptation + MTLs
Datapath = '/path/to/the/data/folder'; Harmonization = 2; Covariate = 0; % Or 1 Formulations = 0; % Or 1 Or 2 Or 3 MTL_fun(Datapath,Harmonization,Covariate,Formulations)
-
To correct for differences in features caused by imaging at two MFSs in ADNI1 and ADNI2, we used MTL models as 6-task learning approaches for MFS adaptation.
% OUTPUT : % Results : 'results','results_NC','results_MCI','results_AD','cor','MAE','cor_s','CorNC','CorMCI','CorAD','MAE_NC','MAE_MCI','MAE_AD' % INPUT : % Datapath : The directory containing data % Formulations : 0- Least Lasso, 1-Joint feature selection (JFS), 2- Dirty Model, 3- Low rank assumption (LRA) %-------------------------------------------------------------------------------------------------------------------------------------------- %-------------------------------------------------------------------------------------------------------------------------------------------- Datapath = '/path/to/the/data/folder'; Formulations = 0; % Or 1 Or 2 Or 3 MTL_6tasks(Datapath,Formulations)