Evaluating the Impact of Pulse Oximetry Bias in Machine Learning under Counterfactual Thinking