Abstract
PURPOSE: To develop a general algorithm employing structured and unstructured electronic health record (EHR) data to identify laterality-specific treatment-warranted disease more accurately at the longitudinal eye level. DESIGN: A retrospective treatment-warranted diabetic macular edema (TW-DME) cohort study. SUBJECTS: Patients with diabetic retinopathy (DR) identified from a health safety net system and a university hospital in Los Angeles, California, employing diagnosis and procedure codes from 2013 to 2023. METHODS: We investigated the completeness and accuracy of laterality-specific TW-DME status based on the following 5 categories of data: Tier 1-Physician Procedure Documentation, Tier 2-Charge Codes (Professional and Facility), Tier 3-Medication Orders, Tier 4-Crosswalked Procedure Codes, and Tier 5-Diagnosis Code associated with Procedure. Laterality data completeness was evaluated for each category, independently and in a tiered hierarchical order. Data accuracy was verified by manual chart review for a subset of validation patients. MAIN OUTCOME MEASURES: Algorithm performance in ascertaining cross-sectional and longitudinal TW-DME status. RESULTS: From 2013 to 2023, 7784 patients with DR had 68 465 visits, with 4809 (61.8%) patients identified as having TW-DME. Notably, 67.9% of health safety net patients had visits with missing diagnosis laterality. The proposed algorithm improved laterality completeness in the treatment-warranted DR cohort to 93.6% for the safety net and 99.0% for the university sites. Validation by chart review demonstrated an increase in positive predictive value (safety net 47.0%-93.2%, university 85.3%-98.8%), negative predictive value (safety net 23.2%-33.3%, university 46.9%-72.6%), sensitivity (safety net 35.9%-76.0%, university 79.2%-96.0%), specificity (safety net 60.4%-76.6%, university 38.8%-90.4%), agreement (safety net 38.5%-76.1%, university 74.8%-95.4%), and F1 score (safety net 40.7%-83.7%, university 82.1%-97.4%) at the longitudinal eye level. CONCLUSIONS: Our algorithm employing structured and unstructured data lays out a general and reproducible approach to more accurately identify and extract laterality-specific data from EHRs. This method was valid across sites with disparate documentation and coding practices. Application of this algorithm could improve the utility of clinical data generated as part of routine care for future investigations of ocular disease prevalence, sequelae, treatment patterns, and costs. FINANCIAL DISCLOSURES: Proprietary or commercial disclosure may be found in the Footnotes and Disclosures at the end of this article.