Imputation of Disease Severity for Diabetic Retinopathy Clinical Encounters in the Iris® Registry (INTELLIGENT RESEARCH IN SIGHT)
Author(s)
Chapman R1, Nguyen T1, Hatfield M1, Leng T2, Mbagwu M1, Torres A1, Borkar D3
1Verana Health, San Francisco, CA, USA, 2Stanford University School of Medicine, Palo Alto, CA, USA, 3Duke University School of Medicine, Durham, NC, USA
Presentation Documents
OBJECTIVES:
Diabetic retinopathy (DR) is a chronic, progressive disease and one of the leading causes of irreversible blindness in the U.S. Electronic health record (EHR)-based registries provide an opportunity to understand the progression of this disease through ICD coding. However, our prior work demonstrated that a significant portion of EHR clinical encounters are coded for unspecified disease severity. This study sought to assess whether a predictive model could be developed to impute DR disease severity.METHODS: All DR clinical encounters with known disease severity based on ICD-9 coding and a prior encounter for the patient in the American Academy of Ophthalmology IRIS Registry between January 1, 2014 and September 30, 2015 were included in the study. The dataset was randomized 80%/20% into training and test sets, respectively. A multinomial logistic regression model was developed utilizing structured demographic, clinical, and procedure data from the most recent prior clinical encounter, including known disease severity, to impute DR disease severity as mild, moderate, severe nonproliferative DR, or proliferative DR. All analyses were performed in Databricks (San Francisco, CA).
RESULTS:
Overall, 747,279 clinical encounters of 174,336 patients were included. A training set of 572,943 encounters and a test set of 152,584 encounters were used. The model achieved an accuracy of 96.9% and an aggregate F-1 score of 0.950 in the test set.CONCLUSIONS:
These findings suggest that it may be possible to impute missing DR disease severity in EHR-based records. The ability to include all clinical encounters for a patient, including those with missing disease severity, could help improve the quality of longitudinal DR research. Utilizing data from prior encounters, including those with known disease severity, appears to yield strong predictive power and may also provide a safeguard against data leakage in such research.Conference/Value in Health Info
Value in Health, Volume 26, Issue 6, S2 (June 2023)
Code
RWD134
Topic
Methodological & Statistical Research, Study Approaches
Topic Subcategory
Electronic Medical & Health Records, Missing Data
Disease
Diabetes/Endocrine/Metabolic Disorders (including obesity)