Statistical quantification of confounding bias in machine learning models