Natural language processing helps ophthalmologists access data, improves data curation

Knowledge of tools used in data interpretation helps clinicians trust accuracy of findings.

If ophthalmologists can understand how the IRIS Registry is curated and analyzed, they can embrace the results.

Artificial intelligence (AI) is permeating so­ciety, directing everything from the “prod­ucts you may like” portion of an e-com­merce site to a GPS suggesting a faster route to your destination. One of the fastest growing areas for AI is medicine, and oph­thalmology is helping to lead the way in AI evolutions. For example, deep-learning AI programs that interpret fundus photographs of patients with diabetes may be used to improve screening for diabetic retinopathy.1

In some cases, AI relies on natural language process­ing (NLP) to gather and interpret language-based data. In ophthalmology, NLP can process electronic health record (EHR) information from the American Academy of Ophthalmology (AAO) Intelligent Research in Sight (IRIS) Registry, which houses data from 367 million patient encounters with more than 65 million unique patients. Verana Health, the data curation and analytics partner of the academy, organizes data from the registry to prepare it for inter­pretation via NLP.

Related: How ophthalmology is pioneering the field of artificial intelligence

Natural language processing

Although the phrase "natural language processing" may be new to some readers, many individuals frequently (and perhaps unknowingly) interact with it. A common encounter with NLP occurs when in­terfacing with document-scanning technology that converts text into digital data. Optical character rec­ognition, an early NLP method, identifies letters, words, and phrases from static documents and con­verts them to data points.

Imagine a scenario in which you are tasked with entering your passport information into a portal. You have 2 options to enter data such as your given name, surname, and nation of origin. You can manually enter your information in each field or you can use your phone to snap a picture of your passport’s rel­evant pages and allow NLP to populate the fields. The latter option, which extracts language from a photograph and places it in the appropriate areas of the portal, is quicker.

Machine learning applied to NLP has broader uses that suit it for analyzing text-heavy data in the IRIS Registry. If NLP is used to interpret the hundreds of millions of EHR data points in the registry, it may pro­duce information on real-world treatment outcomes, disease prevalence data, and treatment patterns.

Related: International coalition categorizes 25 subtypes of uveitis using machine learning

Clinical findings in ophthalmology

Two examples of how AI could be used to examine IRIS Registry data illustrate the potential of draw­ing insights from large databases via NLP analysis.

Grading severity

NLP could be used to search IRIS Registry data for a series of words or phrases in patient records. Search­ing for prespecified phrases or words may confirm the accuracy of coding data. For the sake of illustra­tion, consider glaucoma.

Clinicians use various qualitative (e.g., types of procedures undergone, medication history, whether cataract surgery has occurred) and quantitative (eg, cup-to-disc ratios, IOP measurements, visual acuity, visual field data) data points to classify a patient’s glaucoma severity. No single data point leads to a diagnosis of mild, moderate, or severe glaucoma, and patients with similar quantitative profiles may be classified differently based on qualitative data points.

ICD-10 provides codes for various degrees of glau­coma severity. In a perfect world, clinicians would accurately code each case during every visit. How­ever, due to many factors, the stage of glaucoma may not be updated in the EHR to reflect the current clinical state of the patient. By using NLP, investiga­tors can confirm that the coded diagnosis reflects the qualitative and quantitative measurements of a patient encounter.

Suppose investigators needed to determine the num­ber of patients categorized as having severe glaucoma. After defining severe glaucoma with a combination of qualitative and quantitative parameters—relying on definitions from the AAO and the American Glau­coma Society designed to reduce subjectivity in stage diagnoses— investigators could perform a custom­ized search of IRIS Registry records using NLP to confirm that the number of severe cases coded in a given time frame match the number of severe cases as defined by details in the encounter.

An NLP data analysis makes the real-world data housed in the IRIS Registry more accurate. Automat­ing these quality checks saves time and increases the quality of the overall body of data at investigators’ disposal.

Related: Integrating AI to manage DR in a primary care setting

Reconciling coding data with real-world prevalence

Understanding the prevalence of particular diseases in ophthalmology may be limited by coding behav­ior after a patient encounter.

For example, a patient presenting to a cataract sur­geon for preoperative evaluation may also have early age-related macular degeneration (AMD) in addition to their cataract. It is possible that this encounter will be coded as a cataract for purposes of reimburse­ment, and the ICD-10 code for AMD not entered. This patient’s AMD would be undetected by investigators leveraging coding data to estimate disease prevalence. However, an NLP-based analysis of IRIS Registry data could detect the presence of underreported or unre­ported disease in patient charts, thereby generating a more robust picture of real-world disease rates.

Related: Screening DR + tele-ophthalmology AI: Less expensive than human assessment

Trusting the algorithm

The more clinicians know about how NLP-based analyses make determinations, and the more trans­parent the models are, the more willing they may be to accept the results of AI reports. All NLP algo­rithms require a degree of explainability, which al­lows investigators to understand to what degree an algorithm values particular pieces of data. If NLP determines that, for example, a certain percentage of patients of a certain age have AMD, then investi­gators can examine the algorithm’s methods to en­sure that a legitimate medical reason exists for this conclusion.

Instances may arise in which AI detects disease that is either imperceptible by human evaluation or is linked to heretofore unknown anatomic manifesta­tions. Machine learning–based algorithms, in which AI platforms learn to detect patterns from massive data sets, have been shown to accurately estimate the age, gender, smoking status, and systolic blood pressure of patients based on fundus photographs alone.2 How or why those algorithms make their deter­minations are not yet understood, but their results nonetheless show the potential of AI to change the landscape of medicine.

Related: Implementing at-home monitoring for exudative AMD

What’s next?

NLP may be one of the most important tools for ex­tracting meaningful insights from real-world data in the IRIS Registry. The better we understand how IRIS Registry data are curated and analyzed, the more we can embrace the results of AI data analyses.

1. Lu L, Ren P, Lu Q, et al. Analyzing fundus images to detect diabetic retinopathy (DR) using deep learning system in the Yangtze River delta region of China. Ann Transl Med. 2021;9(3):226. doi:10.21037/atm-20-3275
2. Poplin R, Varadarajan AV, Blumer K, et al. Prediction of cardiovascular risk factors from retinal fundus photographs via deep learning. Nat Biomed Eng. 2018;2(3):158-164. doi:10.1038/s41551-018-0195-0

Leng is the director of research at the Byers Eye Institute at Stanford University in California and a medical adviser to Verana Health.

Related Content: Ophthalmology | Retina in 2020 | Imaging