Loss of reliability in data extraction from clinical records: Source of flaws and usefulness of training

BACKGROUND: Clinical-epidemiological research often requires using data stored in clinical records. There is a paucity of systematic studies of errors in the data extraction process from clinical records in the medical literature. In order to assess the increment of reliability in data extraction from clinical records due to training, we estimate the degree of agreement in the data extraction process from clinical records of rheumatoid arthritis patients. MATERIAL AND METHODS: Test-retest, quasi-experimental study. A random sample of clinical records was selected from a rheumatoid arthritis patients register. The degree of agreement between the two observers, before and after a specific training, was estimated with kappa and intraclass correlation coefficients. RESULTS: Following standardized ranking of the degree of agreement, we observed that 5 out 19 (26%) studied variables improved significantly after training and 14 (74%) remained with the same degree of agreement or did not change significantly. At the end of the study, only two variables received a degree of agreement less than good whereas six did so before training. The improvement was noted in the clinical variables. CONCLUSIONS: It is possible to have a sistematic approach to the source of errors in the use of data from clinical records. The training of observers has a significant impact on the degree of agreement and therefore improves reliability. The training in the extraction and management of clinical information may contribute to the improvement of validity and reliability of observations in medical practice.