Aims: Incorrect classification, diagnosis and coding of the type of diabetes may have implications for patient management and limit our ability to measure quality. The aim of the study was to measure the accuracy of diabetes diagnostic data and explore the scope for identifying errors.
Method: We used two sets of anonymized routinely collected computer data: the pilot used Cutting out Needless Deaths Using Information Technology (CONDUIT) study data (n = 221 958), which we then validated using 100 practices from the Quality Improvement in Chronic Kidney Disease (QICKD) study (n = 760,588). We searched for contradictory diagnostic codes and also compatibility with prescription, demographic and laboratory test data. We classified errors as: misclassified-incorrect type of diabetes; misdiagnosed-where there was no evidence of diabetes; or miscoded-cases where it was difficult to infer the type of diabetes.
Results: The standardized prevalence of diabetes was 5.0 and 4.0% in the CONDUIT and the QICKD data, respectively: 13.1% (n = 930) of CONDUIT and 14.8% (n = 4363) QICKD are incorrectly coded; 10.3% (n = 96) in CONDUIT and 26.2% (n = 1143) in QICKD are misclassified; nearly all of these cases are people classified with Type 1 diabetes who should be classified as Type 2. Approximately 5% of T2DM in both samples have no objective evidence to support a diagnosis of diabetes. Miscoding was present in approximately 7.8% of the CONDUIT and 6.1% of QICKD diabetes records.
Conclusions: The prevalence of miscoding, misclassification and misdiagnosis of diabetes is high and there is substantial scope for further improvement in diagnosis and data quality. Algorithms which identify likely misdiagnosis, misclassification and miscoding could be used to flag cases for review.