They tested on tons of different external datasets, and at least one of the training datasets was balanced. Same results were obtained.