Contact: [[email protected]]
This package supports the submission "Learning vector representation of medical objects via EMR-driven nonnegative restricted Boltzmann machines (eNRBM)" to Journal of Biomedical Informatics (JBI) - Elsevier
The package is written in MATLAB and tested on Windows 64bit machines. Please make sure MATLAB are available.
There are two main parts:
1. Synthetic data
2. Code and demo scripts
The details are described as below:
The data are generated by a learnt RBM model. Here we simulate 1000 patients with 5321 features and 3 classes:
1: no risk,
2: moderate risk,
3: high risk.
The features are similar to the EMR features of heart failure cohort described in the manuscript.
We split the data to perform 10-fold cross-validation. The patient data can be loaded from the file gen_data.mat wherein:
X: data matrix,
y: classes,
idx_train: a cell of 10 arrays of the data indices used for training,
idx_test: a cell of 10 arrays of the data indices used for testing/validation.
The file "feat_correl.txt" contains feature graph in which each line "i,j,1" implies the feature #i connects to feature #j.
All codes are in ".m" files with fair detailed comments. Here we provide 2 demo scripts:
demo_lr_emr: run Logistic Regression,
demo_enrbm_lr_emr: run eNRBM followed by Logistic Regression.
@ARTICLE { tran_nguyen_phung_venkatesh_bi15learning,
AUTHOR = { Tran, Truyen and Nguyen, Tu and Phung, Dinh and Venkatesh, Svetha },
TITLE = { Learning vector representation of medical objects via {EMR}-driven nonnegative restricted {B}oltzmann machines },
JOURNAL = { Journal of Biomedical Informatics (JBI) },
YEAR = { 2015 },
VOLUME = { 54 },
PAGES = { 96--105 },
MONTH = { April },
DOI = { 10.1016/j.jbi.2015.01.012 },
URL = { http://www.sciencedirect.com/science/article/pii/S1532046415000143 },
}