Quantifying surprise in clinical care: Detecting highly informative events in electronic health records with foundation models