Characterizing Physicians Practice Phenotype from Unstructured Electronic Health Records.

Area of Science:

• Clinical Informatics
• Health Services Research
• Cardiology

Background:

• Clinical practice variations can introduce bias into electronic health record (EHR) data.
• This bias can impact the performance of predictive models, independent of patient health.
• Understanding documentation behavior is crucial for accurate health data analysis.

Purpose of the Study:

• To investigate variations in how primary care physicians (PCPs) document heart failure signs and symptoms (FHFSS) in EHRs.
• To identify distinct subgroups of PCPs based on their documentation behaviors.
• To determine if these documentation differences are associated with patient health factors.

Main Methods:

• Utilized EHR encounter note data from 5,187 primary care patients (aged 50-85).
• Employed a validated text extractor to identify Framingham heart failure signs and symptoms (FHFSS) mentions.
• Conducted hierarchical clustering analyses on PCP encounter note data to find documentation subgroups.

Main Results:

• Identified three distinct subgroups of PCPs with differing FHFSS documentation patterns.
• These subgroups varied in their rates of documenting both assertions and denials of FHFSS mentions.
• Physician subgroup differences were not attributable to patient disease burden, medication use, or other health-related factors.

Conclusions:

• Significant heterogeneity exists in PCP documentation of FHFSS within EHRs.
• These documentation variations represent a potential source of bias in predictive modeling.
• Further research is needed to address and mitigate documentation-based biases in clinical data.