There are 20 patients with acute lymphoblastic leukemia (ALL) and 32 patients with acute myeloid
leukemia (AML), both variants of a blood cancer.
The makeup of the groups as follows:

Each individual has an expression value for each of 10000 different genes. The expression value
for each gene is a continuous value between -1 and 1.
You’ve built your model for discriminating between AML and ALL patients and you find that it
works quite well on your current data. One month later, a collaboration tells you she has fresh
data from 100 new AML/ALL patients. You run the samples through your model, and turns out
your model has very poor predictive accuracy on the new samples; specifically, your model
predicts that all males have ALL. What is the most reliable way to fix this problem?

A.
Change the distance metric

B.
Reduce the number of dimensions

C.
Use a Gibbs sampler on a Bayesian network

D.
Perform matched sampling across other provided variables