Using the electronic medical record to predict the pharmacological management of acute myocardial infarction

Improving health care quality while reducing costs requires the elimination of unnecessary and unintended variation in the care process. Decision support applications already exist to foster adherence to standards that would accomplish this. The challenge resides in developing based on scientific evidence and yet consistent with local practice norms.; In this study, data routinely collected by a hospital information system have been examined. Tools and techniques from the field of Knowledge Discovery in Databases (KDD) have been applied to induce models that characterize the pharmacological management of acute myocardial infarction in the LDS Hospital Emergency Department.; NevProp3®, a backpropagation neural network simulator, the NeticaTM application for developing Bayesian networks, CN2, a rule induction program, and logistic regression have been utilized to predict the administration of antiarrhythmics, beta blockers, thrombolytics, and other cardiac-related medications. The independent predictors included MI type, Killip class, age, gender, electrocardiogram results, time from onset of chest pain, and bleeding risk.; Five classification and prediction experiments were conducted. Learning tool sensitivity and specificity were calculated. Agreement reliability among tools was assessed on a case by case basis. None of the tools achieved a sensitivity ≥0.80, though agreement among tools was generally strong. NeticaTM's Bayesian algorithm performed best overall.; Though gigabytes of data are collected each day in the clinical setting, the data most descriptive of and pertinent to clinical decision-making seem to be left out. It is most difficult to glean information from data elements that do not exist.