| Popis: |
Loan R van Hoeven,1,2 Aukje L Kreuger,3,4 Kit CB Roes,1 Peter F Kemper,2,4 Hendrik Koffijberg,5 Floris J Kranenburg,3,4,6 Jan MM Rondeel,7 Mart P Janssen1,2 1Julius Center for Health Sciences and Primary Care, University Medical Center Utrecht, Utrecht, the Netherlands; 2Transfusion Technology Assessment Department, Sanquin Research, Amsterdam, the Netherlands; 3Department of Clinical Epidemiology, Leiden University Medical Center, Leiden, the Netherlands; 4Center for Clinical Transfusion Research, Sanquin Research, Leiden, the Netherlands; 5Department of Health Technology & Services Research, MIRA Institute for Biomedical Technology and Technical Medicine, University of Twente, Enschede, the Netherlands; 6Department of Intensive Care, Leiden University Medical Center, Leiden, the Netherlands; 7Department of Clinical Chemistry, Isala, Zwolle, the Netherlands Background: To enhance the utility of transfusion data for research, ideally every transfusion should be linked to a primary clinical indication. In electronic patient records, many diagnostic and procedural codes are registered, but unfortunately, it is usually not specified which one is the reason for transfusion. Therefore, a method is needed to determine the most likely indication for transfusion in an automated way.Study design and methods: An algorithm to identify the most likely transfusion indication was developed and evaluated against a gold standard based on the review of medical records for 234 cases by 2 experts. In a second step, information on misclassification was used to fine-tune the initial algorithm. The adapted algorithm predicts, out of all data available, the most likely indication for transfusion using information on medical specialism, surgical procedures, and diagnosis and procedure dates relative to the transfusion date.Results: The adapted algorithm was able to predict 74.4% of indications in the sample correctly (extrapolated to the full data set 75.5%). A kappa score, which corrects for the number of options to choose from, was found of 0.63. This indicates that the algorithm performs substantially better than chance level.Conclusion: It is possible to use an automated algorithm to predict the indication for transfusion in terms of procedures and/or diagnoses. Before implementation of the algorithm in other data sets, the obtained results should be externally validated in an independent hospital data set. Keywords: indication for transfusion, selection algorithm, electronic health record data |
