Model use leads to mixed associations because at-risk patients avoid adverse outcomes, and the EHR captures this. Future models trained on data containing these mixed associations perform worse.