The goal of my project was to create a machine learning model that could most accurately predict lung transplant graft failure or success. This is a pressing issue because primary graft dysfunction affects 10 to 25% of lung transplant patients and is a primary cause of post-transplant mortality.

I started with a large lung transplant data set from the United Network for Organ Sharing and my first task was to select prediction variables pertaining to information about the patient’s health and use these prediction variables to predict graft failure.

I then had to select and write code for machine learning models that I could use to predict graft failure. I chose to use logistic regression, a support vector machine (SVM), and a random forest classifier. After many trials and a lot of tweaking of parameters, I achieved the best results using a second-degree polynomial kernel for the SVM and creating a random forest model with 100 decision trees. I also made a neural network that used a sequential model with two input layers and an output layer. I then made an ensemble model combining the logistic regression, the SVM, and the random forest classifier. An ensemble model uses multiple machine learning algorithms to obtain better predictive performance.

Model performance was judged by accuracy score and average precision (area under the precision-recall curve). I found that the random forest classifier had an accuracy score of 0.79 and an average precision of 0.81, the logistic regression had an accuracy of 0.78 and an average precision of 0.79, the SVM had an accuracy of 0.77 and an average precision of 0.82, the neural network had an accuracy of 0.75, and the ensemble model had an accuracy of 0.80 and an average precision of 0.80. I also plotted a Kaplan Meier Curve to show the probability of graft failure from the time of the transplant.

I performed two statistical tests on each prediction variables versus graft failure to study the different relationships. I used the chi-squared test to study categorical variables. From the chi-squared test, I found that there was a significant relationship between graft failure and four variables regarding a patient’s lung preference for transplant. The Wilcoxon Rank-Sum Test indicated that patient age was the only numerical variable that was statistically significant. Statistically significant in this case means that the distribution of age values is significantly different between graft failure and success.

A limitation of the dataset was that there were a lot of missing values. After removing all the NA and null values, the number usable of rows decreased from 178,000 to 941. In order to have more usable data points, in the future I should use imputation to fill in empty values. In conclusion, the random forest classifier and the ensemble model were able to predict graft failure most accurately with an accuracy of 0.79 and 0.80 respectively.