Use of Natural Language Processing to Accurately Identify Cataracts and Other Lens Pathology in Electronic Health Record Data – A Study Using the Sight Outcomes Research Collaborative (SOURCE) Repository

Abstract

Purpose : Nearly all published Big Data analyses in ophthalmology have been completely reliant upon ICD billing codes to identify the presence of ocular pathology and to assess for disease stability. However, for a variety of reasons, billing codes can be inaccurate. Here we validate an alternative way to identify and characterize ocular pathology using natural language processing (NLP).

Methods : We developed an NLP algorithm capable of searching free text lens exam data inputted by clinicians into the electronic health record. The algorithm identifies the type(s) of cataract present, cataract density, presence of intraocular lenses (IOLs), location of the IOL, status of the capsule, and presence of other lens pathology (e.g., lens dislocation, phimosis). We applied our algorithm to all 17.5 million lens exam records in the SOURCE repository, which captures eye care for more than 3 million eye care recipients receiving care at 11 large US health care systems. We randomly selected 4314 unique lens exam entries and asked 11 clinicians to review the NLP output to validate whether it correctly identified and categorized the lens pathology present.

Results : Among the 4314 lens exam entries, the NLP algorithm correctly identified and properly categorized all lens pathology present in 4104 (95.1%) of the entries as validated by clinicians. There were only 210 entries (4.9%) identified with errors. Among less common lens pathology, the mentions identified by the algorithm were corroborated by a reviewing clinician for 100% of mentions of pseudoexfoliation material; 99.7% for mentions of phimosis, subluxation, and synechiae; and 96.9% for mentions of phacodonesis. Algorithm errors included incorrectly accounting for negation (e.g., “no evidence of pxf”) and improper assignment of “pigment on the lens capsule” as evidence of posterior capsule opacification.

Conclusions : We developed an algorithm using NLP that accurately identifies and classifies lens abnormalities routinely documented by eye care professionals. Algorithms such as this will help researchers properly identify and classify ocular pathology to augment the sorts of questions that can be answered using electronic health record data.

This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.