Theodore Leng, MD, MS

home

authors

theodore-leng-md-ms


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.

In some cases, AI relies on natural language processing (NLP) to gather and interpret language-based data. 

Deep learning algorithm proven accurate for AMD classification

In ophthalmology, NLP can process electronic health record (EHR) information from the 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.


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.


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.


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. 

Linking patient imaging data to AAO IRIS Registry

Clinicians use various qualitative (eg, 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.

 Looking at real-world trial data replication outcomes of IRIS Registry data

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.

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 

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.

Applying impact of real-world data on day-to-day practice 


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.

Blockchain technology aims to drive big data to 'train' AI

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.

 

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.


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. 

--




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


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. 

2. Poplin R, Varadarajan AV, Blumer K, et al. Prediction of cardiovascular risk factors from retinal fundus photographs via deep learning. 

2018;2(3):158-164. doi:10.1038/s41551-018-0195-0

Articles

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

Natural language processing pairs with big data curation 

The most rigorous data in evidence-based medicine is considered to be that generated by prospective, randomized, controlled trials (RCTs). 

RCTs are essential for evaluating the therapeutic impact of a drug or device while eliminating potential confounding factors. 

However, there are many situations in which these gold standard clinical trials are not practical or cost effective.



Clinical trial outcomes for macular edema don’t translate to ‘real-world’ settings

RCTs evaluate outcomes for very tightly defined groups of patients with highly standardized care and optimal adherence and follow-up. 

In the real world, however, patients present with a variety of underlying systemic conditions, prior surgery, and adherence patterns that muddy the picture. This is where retrospective, observational studies can help to fill the knowledge gap with real-world evidence (RWE).

RWE can help to answer key clinical questions, such as how a treatment works in patients who are already on other medications, and whether switching to or adding a new pharmaceutical agent would be more beneficial. 

Real-world studies can demonstrate how a new 

 therapy compares with the prostaglandin analogues (PGAs) that most clinicians rely on for first-line therapy, rather than with timolol, the typical control used in FDA studies. 

In cardiology, a major ongoing RWE study, the ADAPTABLE trial (NCT02697916), is trying to settle the debate over whether patients with coronary artery disease do better on low- or high-dose aspirin.

Developing a new therapy: Where do we go from here?


Increasingly, RWE researchers are looking at big data for answers. Very large data registries permit analysis of records from representative patient populations that are orders of magnitude larger than the number of subjects typically found in a phase 3 clinical trial.

Verana Health, through our partnership with the American Academy of Ophthalmology, is actively engaged in curating and analyzing data from the IRIS Registry (Intelligent Research in Sight) to better inform clinical practice in ophthalmology. 

The IRIS Registry is the largest specialty clinical database in medicine, with 343 million patient visits and billions of data points reported by more than 15,000 ophthalmologists through their electronic health records (EHR) systems. 

Large registries like this offer the opportunity to gather RWE about clinical outcomes. They facilitate research on questions that are critical to patient care but may be less likely to garner industry support such as head-to-head comparisons of new and long-established treatments.



IRIS Registry study: Overall anti-VEGF injection intervals for wet AMD similar at 1, 2 years of treatment


Large registry studies are relatively new but have already led to changes in ophthalmic practice. 

For example, a retrospective analysis of all procedures in the Swedish national cataract registry over a multi-year period revealed that intracameral antibiotics were an important factor in reducing the rate of endophthalmitis.

 

This led to a prospective European study of intracameral antibiotics

 and widespread adoption of this method of infection prophylaxis in Europe. A large health maintenance organization (HMO) in the United States also reached similar conclusion about the value of injected antibiotics in cataract surgery after reviewing a large data set from its own EHR system.

Another national registry helped to establish corneal collagen cross-linking as the standard of care for keratoconus. 





In this case, analysis of healthcare resource utilization in the Netherlands before and after the introduction of crosslinking showed that the procedure was associated with a 25% reduction in corneal transplantation over three years.

We have been able to use the IRIS Registry data to evaluate real-world anti-VEGF injection intervals in 56,672 treatment-naïve patients with 

 

In this study, the authors found that the mean number of injections per eye was approximately five per year.

By the end of year 1, approximately 40% of eyes continually treated with anti-VEGF medications needed injections more frequently than every 8 weeks, including 17%-18% who still needed monthly injections.

In another retrospective review of IRIS Registry data for 12.5 million patients seen by an eye care provider in 2017, researchers evaluated the relationship between smoking and IOP.

 

After adjusting for many factors, including age and glaucoma diagnosis, smokers (both current and those with a history of smoking) had higher mean IOP than those who had never smoked. 

In fact, smoking was the most important predictor of IOP after the glaucoma diagnosis (

IOP homeostasis in glaucoma: Looking below the surface


The ability to access large amounts of information to improve medicine is very exciting. But large data sets can be quite challenging to interpret. Imagine a spreadsheet with billions of rows and columns. 

No human can realistically check it over for errors; sophisticated algorithms are needed to catch and remove errors in a systematic way. The raw data must be cleaned, harmonized, and structured in such a way that it can be used for clinical research.

To better understand this, consider something as simple as visual acuity (VA)—an important outcome measure in ophthalmic studies. 

Data comes into the IRIS Registry every day from thousands of doctors using more than 50 different EHR platforms, each of which collects VA in slightly different fields or formats.



Apps now online to aid patients with visual impairment

VA may have been measured using a Snellen or ETDRS (Early Treatment Diabetic Retinopathy Study) chart, autorefraction, or a pinhole test. The patient may have been tested with or without correction, monocularly or binocularly. 

All of that VA data needs to be harmonized so that “20/20” has the same meaning across millions of records.

At Verana, we are using machine-learning techniques to refine algorithms that characterize important information from “unstructured” data, such as free text and images. 

Imaging is very important in ophthalmology, so the ability to develop algorithms that can mine images for precalculated measurements such as retinal nerve fiber layer thickness, as well as pathological features (eg, drusen) is highly relevant for future research.



In fact, real-world data from large registries opens up many new avenues for clinical research and the advancement of medicine. 

The sheer volume of patients offers a chance to drill down into specific subgroups that are underrepresented in clinical trials and to better understand how a disease process or treatment response might vary by gender, race or ethnicity.

We are also passionate about the opportunity to use RWE from the IRIS Registry and other large registries to develop more personalized patient care. 



Pearls to educate patients about ocular telemedicine options



Currently, a new glaucoma patient is typically treated first with a prostaglandin analog—the best available treatment, based on the literature. 

But what if we had a robust algorithm that could take into account not just the basic diagnosis, but age, ethnicity, gender, and comorbidities, to identify the best treatment paradigm for that individual patient? Ironically, big data has the power to change how we approach small, daily decisions.

And as large as these registries are, there are opportunities to deepen them further through data linking. If we link records from pharmacy prescription databases or serum testing laboratories to the IRIS Registry data, we could get a much clearer picture of diabetic patients’ glycemic control, and potentially be able to evaluate how glycated hemoglobin levels or metformin usage correlate with diabetic retinopathy changes.

We now have the ability to access large volumes of real-world data from very broad patient populations. The possibilities for using that data to improve patient care and drive scientific insights are limitless. 


Roe is chief medical officer of Verana Health. A cardiologist by training, he spent 20 years at Duke University and as an investigator at the Duke Clinical Research Institute in Durham, North Carolina, where he focused on research initiatives leveraging real-world data. He was a principal investigator in the ADAPTABLE trial.


Leng is a practicing retinal specialist, assistant professor of ophthalmology and director of Clinical and Translational Research at Stanford University Medical School in Stanford, California. He has a research background in image analysis, machine learning algorithms, and clinical trial design and execution. He serves as a medical adviser to Verana Health.


Johnston A, Jones WS, Hernandez AF. The ADAPTABLE trial and aspirin dosing in secondary prevention for patients with coronary artery disease. 

. 2016;18(8):81. doi:10.1007/s11886-016-0749-2

Friling E, Lundström M, Stenevi U, Montan P. Six-year incidence of endophthalmitis after cataract surgery: Swedish national study. 

. 2013;39(1):15-21. doi:10.1016/j.jcrs.2012.10.037

Endophthalmitis Study Group, European Society of Cataract and Refractive Surgeons. Prophylaxis of postoperative endophthalmitis following cataract surgery: results of the ESCRS multicenter study and identification of risk factors. 

. 2007;33(6):978-988. doi:10.1016/j.jcrs.2007.02.032

Herrington LJ, Shorstein NH, Paschal JR, et al. Comparative effectiveness of antibiotic prophylaxis in cataract surgery. 

. 2016;123(2):287-294. doi:10.1016/j.ophtha.2015.08.039

Godefrooij DA, Gans R, Imhof SM, Wisse RP. Nationwide reduction in the number of corneal transplantations for keratoconus following the implementation of cross-linking. 

. 2016;94(7):675-678. doi:10.1111/aos.13095

MacCumber M, Yu, JS, Sagkriotis A, et al. Injection intervals in treatment-naïve neovascular AMD patients who received anti-VEGF agents: an analysis of the IRIS Registry. Paper presented at: American Academy of Ophthalmology Annual Meeting October 12-15, 2019; San Francisco. 2019: PO471.

Lee CS, Owen JP, Yanagihara RT, et al. Smoking is associated with higher intraocular pressure regardless of glaucoma.

. 2020;3(4)253-261. doi: 10.1016/j.ogla.2020.03.008

Evidence-based medicine can lead to personalized care, better results for patients

Theodore Leng, MD, MS

Articles

Natural language processing pairs with big data curation

Big data drives opportunity for real-world evidence