Prediction of Chronic Graft vs. Host Disease Using Machine Learning

Abstract

This paper attempts to predict the onset of chronic Graft vs. Host Disease (GVHD) in children with blood cancers who have received a bone marrow or stem cell transplant using machine learning models. It analyzes and compares the results of three different models in terms of how accurate they each are in predicting chronic GVHD. These models are Logistic Regression, J48 algorithm using decision trees, and Multilayer Perceptron. The models are formed using a dataset containing 36 attributes, excluding chronic GVHD itself. Through data preprocessing and analysis in Weka, these 36 attributes are narrowed down for each model to figure out which combination of attributes leads to the best predictive accuracy. The study uses 10-fold cross validation for each model and uses the Receiver Operating Characteristic (ROC) Area as a measure of the accuracy for each model. The study found that Multilayer Perceptron is the best predictor of chronic GVHD. In comparison, Logistic Regression was the worst predictor of chronic GVHD. The J48 algorithm used the least number of attributes to make its prediction.