Abstract

IntroductionThe rising cost of providing healthcare services creates an extreme pressure to know what works best in medicine. Traditional methods of generating clinical evidence are expensive and time consuming. The availability of electronic clinical data generated during routine patient encounters provides an opportunity to use that information to generate new clinical evidence. However, electronic clinical data is frequently marred by inadequate quality that impedes such secondary uses. This study provides a proof-of-concept and tests the classification accuracy of ClinicalTime—a temporal query system—to identify patient cohorts in clinical databases.Methodswe randomly selected a sample of medical records from the MIMIC-II database, an anonymized database of intensive care patients. Records were manually classified as having an acute kidney injury or not according to the AKIN criteria. Those records were then blindly classified using ClinicalTime to represent the AKIN criteria. Classification accuracy was measured.ResultsClinicalTime correctly classified 88% of all patients, with a sensitivity of 0.93 and specificity of 0.84. Its performance was superior to simply using ICD-9 codes, which correctly classified 66% of all patients.ConclusionsClinicalTime, a temporal query system, is a valid method to add to the currently available ones to identify patient phenotypes in patient databases and, thus, improving our ability to re-use routinely collected electronic clinical data for secondary purposes.