Abstract
BACKGROUND: During the Coronavirus disease 2019 (COVID-19) pandemic it became apparent that it is difficult to extract standardized Electronic Health Record (EHR) data for secondary purposes like public health decision-making. Accurate recording of, for example, standardized diagnosis codes and test results is required to identify all COVID-19 patients. This study aimed to investigate if specific combinations of routinely collected data items for COVID-19 can be used to identify an accurate set of intensive care unit (ICU)-admitted COVID-19 patients. METHODS: The following routinely collected EHR data items to identify COVID-19 patients were evaluated: positive reverse transcription polymerase chain reaction (RT-PCR) test results; problem list codes for COVID-19 registered by healthcare professionals and COVID-19 infection labels. COVID-19 codes registered by clinical coders retrospectively after discharge were also evaluated. A gold standard dataset was created by evaluating two datasets of suspected and confirmed COVID-19-patients admitted to the ICU at a Dutch university hospital between February 2020 and December 2020, of which one set was manually maintained by intensivists and one set was extracted from the EHR by a research data management department. Patients were labeled 'COVID-19' if their EHR record showed diagnosing COVID-19 during or right before an ICU-admission. Patients were labeled 'non-COVID-19' if the record indicated no COVID-19, exclusion or only suspicion during or right before an ICU-admission or if COVID-19 was diagnosed and cured during non-ICU episodes of the hospitalization in which an ICU-admission took place. Performance was determined for 37 queries including real-time and retrospective data items. We used the F(1) score, which is the harmonic mean between precision and recall. The gold standard dataset was split into one subset including admissions between February and April and one subset including admissions between May and December to determine accuracy differences. RESULTS: The total dataset consisted of 402 patients: 196 'COVID-19' and 206 'non-COVID-19' patients. F(1) scores of search queries including EHR data items that can be extracted real-time ranged between 0.68 and 0.97 and for search queries including the data item that was retrospectively registered by clinical coders F(1) scores ranged between 0.73 and 0.99. F(1) scores showed no clear pattern in variability between the two time periods. CONCLUSIONS: Our study showed that one cannot rely on individual routinely collected data items such as coded COVID-19 on problem lists to identify all COVID-19 patients. If information is not required real-time, medical coding from clinical coders is most reliable. Researchers should be transparent about their methods used to extract data. To maximize the ability to completely identify all COVID-19 cases alerts for inconsistent data and policies for standardized data capture could enable reliable data reuse.