Lung Cancer Prediction

Tina Lin • 12/2018

 

Data Source

The dataset that I use is a National Lung Screening Trail (NLST) Dataset that has 138 columns and 1,659 rows. 1,659 rows stand for 1,659 patients. There is a “class” column that stands for with lung cancer or without lung cancer. The other columns are features of the patients, such as “age”, “height”, “education”, etc. The dataset is provided by a professor at the State University of Arkansas and I am a remote volunteer for his lung cancer research project.

 

Motivation

Lung cancer causes more deaths than any other cancer. The odds for men is 1 in 13 while that for women is 1 in 16. Therefore, I want to create a model which can find the best features for lung cancer prediction. It can be used to aid the doctors in the decision making process and improve the disease identification process.

 

Methods

Feature Selection

1. RandomForest embedded pagckages in H2O
2. Weight of Evidence and Information Value (https://github.com/h2oai/h2o-meetups/blob/master/2017_11_29_Feature_Engineering/Feature%20Engineering.pdf) (better performance)

Models

1. DecisionTree Classifier
2. GradientBoosting Classifier
3. LogisticRegression Classifier
4. XGBoost Classifier