Example of data mining use to follow indicators in clinical pathways

Objectives: To measure the reliability of data mining for indicators related to patient treatment at hospital discharge. Methods: Design: Retrospective cohort study. Population: Patients discharged alive after an admission for heart failure in a general internal medicine department from 2009 to 2010. Data: Key treatments at patient's discharge extracted from the clinical information system compared with data extracted manually from the medical records. Endpoint: Accuracy of data mining for treatment prescription. Analysis: Sensitivity, specificity, positive and negative predictive values (PPVs and NPVs) of data mining for angiotensin-converting enzyme (ACE) inhibitors and betablockers prescription discharge. The gold standard was manual data extraction. We then investigated causes of discrepancies between the two methods. Results: A total of 724 patients were included. At discharge, 85.2% received an ACE inhibitor and 72.4% a beta-blocker. For ACE inhibitors, data mining yielded a sensitivity of 90%, a specificity of 100%, a PPV of 100% and an NPV of 64%. Corresponding values for beta-blockers were 95%, 100%, 100% and 88%, respectively. Main causes for discrepancy were: omission of some molecules in the electronic query used; non-standard writing of a prescription in the clinical information system; formats incorrectly interpreted by the query. Conclusion: Immediate reliance on data mining for drug prescription is currently unwarranted because this complex process is still prone to errors. Results should be manually checked before they can be used as quality indicators.