Matching data fragments with imperfect identifiers from disparate sources

The Multiparameter Intelligent Monitoring in Intensive Care (MIMIC-II) Database includes waveforms and derived parameters from bedside monitors, clinical data from an ICU information system, and data from other hospital laboratories and archives, for thousands of patients. These data come from devices under separate domains that often do not retain detailed information regarding relationships between parameters. We developed software for matching data fragments with incomplete and sometimes incorrect identifiers. We found that names, medical record numbers, waveform times and durations, and ICU admission and discharge records were most helpful when available; however, physiological data can also be used in some circumstances. Rule-based normalization and text edit-distance metrics are used in addition to a visual verification tool for patients whose records cannot be assembled automatically. Thus, a majority of the available waveform recordings are matched to patients in the clinical database.