Comparison of Diagnosis Codes to Clinical Notes in Classifying Neovascular Glaucoma Patients

Abstract

Purpose : The electronic health record (EHR) contains a wealth of clinical data, but improved and automated classification approaches are needed to accurately identify patient cohorts in order to use this data for research. We aim to evaluate if a classification algorithm using clinical notes can improve the accuracies of neovascular glaucoma (NVG) and neovascularization of iris (NVI) diagnoses compared to International Classification of Disease (ICD) codes.

Methods : De-identified EHR data from an academic medical center identified patients aged ≥ 18 years with a documented ophthalmology encounter. 221 patients were included in the study cohort: 100 randomly selected with a history of diabetic retinopathy, and 121 with a diagnosis of NVG or NVI by ICD codes. Chart review by 3 ophthalmologists established the gold standard. ICD-9 or 10 codes were used depending on encounter date. Natural language processing and text mining methods identified positive mention of NVG or NVI from in-person office and procedure machine redacted notes (notes method). Classification by ICD codes and notes were compared with the gold standard on sensitivity and specificity. Confidence intervals (CI) were obtained by bootstrapping.All 712 patients with proliferative diabetic retinopathy (PDR) by ICD code was classified for NVG using the ICD codes or the notes, and compared.

Results : For NVG classification, the notes method had significantly lower sensitivity (79.5% see table 1 for CI) compared to ICD codes[CS1] (95.5%), with higher specificity (89.5% vs. 80.5%, not significant). Analysis revealed the redaction process mislabeled “NVG” as protected health information and removed. ICD-9 codes outperformed ICD-10 codes on sensitivity (94.3 vs 74.6%). Combining notes and ICD codes increased sensitivity to 98.9[CS2] %. For NVI, the notes method significantly outperformed ICD codes on sensitivity (80.0% vs. 47.4%) but was worse for specificity (80.1.0% vs. 92.3%). For the entire PDR cohort, NVG classification by the notes method each identified a subset of patients with poor overlap (Table 2).

Conclusions : Clinical notes increased the sensitivity of NVI classification compared to ICD codes but had decreased sensitivity for NVG due to the redaction of “NVG” in the notes. Clinical notes and ICD codes together increased our detection of NVG patients. Resolving errors in redaction will likely improve the accuracy of our notes algorithm.

This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.