Deriving actionable clinical knowledge from large retrospective set of cardiac echo reports: evaluation of machine learning-based discovery of rules for detecting conflicts between report statements