Search

In most electronic health record (EHR) systems, clinicians record diagnoses using interface terminologies, such as Intelligent Medical Objects (IMO). When extracting data from EHRs for collaborative research, local codes are often transformed to standard terminologies for consistent analyses despite the potential for loss of fidelity. EHR diagnosis codes may be standardized directly during the Extract-Transform-Load (ETL) process to the "Meaningful Use" clinical data exchange standard, SNOMED-CT, or to the International Classification of Diseases (ICD) terminologies commonly used for billing. We examined the performance of ETL standardization via the direct IMO mapping to SNOMED-CT, and via IMO mapping to ICD-9-CM or ICD-10-CM followed by UMLS mapping to SNOMED-CT. We found that for both ICD-9-CM and ICD-10-CM, only 24-27% of diagnosis codes map to the same SNOMED-CT code selected by the direct IMO-SNOMED crosswalk. We identified that differences in mapping lead to loss in the granularity and laterality of the initial diagnosis., (©2020 AMIA - All rights reserved.)

Clinical data research networks (CDRNs) invest substantially in identifying and investigating data quality problems. While identification is largely automated, the investigation and resolution are carried out manually at individual institutions. In the PEDSnet CDRN, we found that only approximately 35% of the identified data quality issues are resolvable as they are caused by errors in the extract-transform-load (ETL) code. Nonetheless, with no prior knowledge of issue causes, partner institutions end up spending significant time investigating issues that represent either inherent data characteristics or false alarms. This work investigates whether the causes (ETL, Characteristic, or False alarm) can be predicted before spending time investigating issues. We trained a classifier on the metadata from 10,281 real-world data quality issues, and achieved a cause prediction F1-measure of up to 90%. While initially tested on PEDSnet, the proposed methodology is applicable to other CDRNs facing similar bottlenecks in handling data quality results.