Mining Complex Clinical Data for Patient Safety Research Final rept. (Sept. 1, 2001-Aug. 31, 2004)

摘要

The project was carried out in a large, urban academic medical center with a repository of 15 years of data on 2.4 million patients seen in inpatient and outpatient areas. Electronic data included registration data, laboratory data, narrative ancillary reports (radiology, etc.), and notes written by providers (discharge summaries, resident signout notes, visit notes, etc.). Narrative data were structured and coded using the MedLEE NLP system and merged with other coded clinical data. Queries were written by experts or derived from machine learning to detect a broad range of medical events. The accuracy of event detection was estimated (sensitivity, specificity, predictive value). Cognitive analyses and case-based reasoning were also applied. Narrative reports such as discharge summaries and resident signout notes were found to contain useful information for uncovering medical events. NLP successfully identified which of 45 NYPORTS events occurred in 57,452 discharge summaries, achieving a PPV of .44 and sensitivity of .27. NLP was found to complement and improve upon error detection based on coded data. Providers were found to explicitly report medical errors in the electronic health record at rates similar to voluntary error reporting (.3 to 1.9%). Cognitive studies revealed complex causation for errors and flawed decision-making processes. Case-based reasoning on an errors database was accurate. An errors terminology was developed and incorporated into the LOINC national standard.