Assessing Data Augmentation-Induced Bias in Training and Testing of Machine Learning Models