Random Under-Sampling Ensemble Methods for Highly Imbalanced Rare Disease Classification

摘要

Classification on imbalanced data presents lots of challenges to researchers. In healthcare settings, rare disease identification is one of the most difficult kinds of imbalanced classification. It is hard to correctly identify true positive rare disease patients out of much larger number of negative patients. The prediction using traditional models tends to bias towards much larger negative class. In order to gain better predictive accuracy, we select and test some modern imbalanced machine learning algorithms on an empirical rare disease dataset. The training data is constructed from the real world patient diagnosis and prescription data. In the end, we compare the performances from various algorithms. We find that the random under-sampling Random Forest algorithm has more than 40% improvement over traditional logistic model in this particular example. We also observe that not all bagging methods are out-performing than traditional methods. For example, the random under-sampling LASSO is inferior to benchmark in our reports. Researchers need to test and select appropriate methods accordingly in real world applications.