New Automated Granular Data Segmentation Approach: Context Impacts Categorization by Physicians (Preprint)

BACKGROUND Granular control over medical record sharing empowers patients to manage access to sensitive information, like substance use disorders (SUD), often withheld due to stigma. Although federal and state laws protect this data, manual segmentation processes remain error-prone, complicating care coordination. Integrating contextual EHR data into segmentation technologies could enhance accuracy, better aligning sensitivity classifications with providers’ assessments. OBJECTIVE This is the first study to examine the influence of context, specifically pertinent health data availability, on how physicians categorize potentially sensitive health data items. The goal is to increase the accuracy of automated sensitive data segmentation technologies. METHODS 1,000 data items from deidentified EHRs from patients receiving behavioral health care were distributed across 12 surveys. These data items were randomly assigned to 12 pairs of physicians to categorize into a sensitive data taxonomy. During a follow-up interview, participants categorized 120 survey data items and could use the medical record to support their answers. Survey and interview results were compared. RESULTS Educational materials provided during surveys and interviews for the sensitive data taxonomy were helpful (75.0% and 70.8%, respectively), but the provided data sensitive taxonomy may require further revisions and validations. 62.5% of the participants used the provided patient EHR data to justify their categorizations during the interview. Some participants felt they needed access to more patient data (50%) or speak with the patient (8.3%) to best categorize a data item. When considering the 240 survey and interview responses, accessing additional patient health data led to multiple (72.1%) changes in categorizations. When comparing the survey and interview categorizations of data items, disagreements decreased by 20.1%, partial agreements increased by 26.4%, but agreements decreased by 6.4%. CONCLUSIONS Incorporating context using additional data from a patient’s EHR impacted how physicians categorize potentially sensitive health data. As future work, this study's results will open new research discussions in physician-validated data segmentation methods.