Quality of Electronic Health Records: Encountering Misclassification

OBJECTIVES: Electronic health record (EHR) encounter documentation describes the patient-provider interaction. Changes in documentation across time and healthcare system can lead to misclassification and potentially to information bias in observational studies. This study uses EHR data to identify and describe misclassification related to encounter class.

METHODS: Encounter class and type were examined using the Oracle EHR Real-World Data, a cloud-based, de-identified, and HIPAA-compliant dataset. Encounter-level class and type over a ten-year period (2013-2022) were assessed for each unique encounter ID. Encounter classes were evaluated by healthcare system and relative to the HL7-FHIR 2020 value set: inpatient, ambulatory, observation, emergency, virtual, and home health. Descriptive analyses were performed.

RESULTS: Over a 10-year period, 1.2 billion unique encounters were reported for 71.7 million patients across 136 healthcare systems. Potential for encounter data misclassification was identified in four areas. First, temporal changes in coding practice within a healthcare system were detected. For example, in a single healthcare system, encounter_type=recurring patient was coded as encounter_class=outpatient during 2013-2014 and as encounter_class=recurring patient from 2015-2022. Second, differential coding practices and related mis-mapping issues were observed between healthcare systems. For instance, encounter_type=urgent care (encounter or facility) was coded by 18 healthcare systems for ≥10 unique encounters/healthcare system as an encounter_class of emergency (38.9%), outpatient (55.6%), or null (5.6%). Third, transition of care during an encounter was inadequately captured in the EHR. For example, of unique encounters with the first encounter_class=emergency, 11.1% have a last encounter_class=inpatient and 3.9% have a last encounter_class=admitted for observation. Fourth, healthcare systems had non-FHIR standard coding encounter class concepts, such as attending clinic or history taking.

CONCLUSIONS: Large, multisystem databases are at risk for misclassifying encounters during data collection, management, and analysis and can result in information bias. Understanding and managing these risks are critical for generating unbiased real-world evidence.