Machine-learning Technology Used to Provide Personalized Care

Kellogg researchers are leading several projects using machine-learning technology, a form of artificial intelligence, to analyze huge amounts of data on patients with ocular diseases. The results of their efforts illustrate the technology’s potential to make personalized predictions of disease stability and help guide clinical management decisions.

Nambi Nallasamy, M.D., Joshua D. Stein, M.D., M.S., and Maria A. Woodward, M.D., M.S. All are working on ways to apply machine learning to ophthalmology
Nambi Nallasamy, M.D., Joshua D. Stein, M.D., M.S., and Maria A. Woodward, M.D., M.S. All are working on ways to apply machine learning to ophthalmology

Tapping Into Big Data

For the past two years, Joshua Stein, M.D., M.S., and his team of biostatisticians, data architects and research assistants have been building the Sight Outcomes Research Collaborative (SOURCE) repository, a transformative resource that contains hundreds of millions of data points from the electronic health records and ocular diagnostic tests of patients receiving eye care at the Kellogg Eye Center. After methodically preparing the data and removing all patient identifiers, the team feeds it into sophisticated machine-learning algorithms to provide clinicians with new tools to better care for patients with ocular diseases.

Based on the success of this initiative at Kellogg, other academic ophthalmology departments nationwide are now sharing their data in a new collaborative arrangement with SOURCE. The database contains more than 500,000 patients with ocular diseases, 1.2 million office visits, 36,000 eye surgeries, 8 million laboratory test results, 17.8 million medication orders and 530,000 images of the retina.

Researchers have been tapping into data from SOURCE to learn about an array of different ocular diseases. Examples of projects using this resource include generating personalized forecasts of whether a patient’s glaucoma will remain stable or experience progression over time, using machine learning to predict the small subset of patients who are at high risk for experiencing poor outcomes following cataract surgery, and predicting which patients with keratoconus will require corneal transplantation.

Targeting Treatment for Corneal Ulcers

Corneal specialist Maria Woodward, M.D., M.S., is using machine-learning algorithms to drive individualized treatments for microbial keratitis (MK).

An infection or ulceration of the cornea, MK is the fourth-leading cause of blindness worldwide. Although the clinical features and severity of MK symptoms vary widely, most patients are treated with nonspecific broad-spectrum antimicrobials, increasing their risk of developing antimicrobial resistance.

Her project uses two algorithms to characterize the full clinical spectrum of MK. The first, developed with Karandeep Singh, M.D., an assistant professor of learning health sciences and internal medicine at Michigan Medicine, will extract and analyze MK data from patient records in the SOURCE database. The second, the result of a partnership with Sina Farsiu, Ph.D., an associate professor of biomedical engineering and ophthalmology at Duke University, automates the process of analyzing slit lamp images of MK.

The information will be combined to build a new evidence-based model to classify and score MK, which physicians can use to assess risks and personalize treatments.

Choosing the Best Lens for Cataract Surgery

Machine learning may also help doctors select the optimal artificial lens to implant during cataract surgery, the most commonly performed surgical procedure in the world.

Currently, doctors choose from a number of formulas that recommend a power of intraocular lens (IOL) to implant. There is no optimal formula; each uses measurements such as eye length or corneal power, and each has the potential to overestimate or underestimate these variables.

Kellogg corneal specialist Nambi Nallasamy, M.D., is bringing doctors more precise tools for this critical decision.

Drs. Nallasamy and Stein with the team of biostatisticians, data architects and research assistants working on the SOURCE database.
Drs. Nallasamy and Stein with the team of biostatisticians, data architects and research assistants working on the SOURCE database.

As a first step, his team developed an algorithm to help doctors choose an IOL formula from six established options. The algorithm can predict which formula produces the smallest error in refraction based on a patient’s preoperative eye measurements. Testing with the SOURCE database showed this improved accuracy by 13.5 percent.

Dr. Nallasamy is now working on a data-driven, patient-specific IOL selection tool to someday replace existing selection formulas.

Non-Invasive Diagnosis of Ocular Surface Tumors

Dr. Nallasamy is also applying machine learning to diagnose tumors on the eye’s surface in a noninvasive, completely data-driven way.

Why? Few doctors can differentiate between benign and cancerous surface lesions using imaging alone, so surgical biopsies are needed to confirm diagnoses.

A notable exception is Kellogg alumnus Carol Karp, M.D. (see page 33) who pioneered the use of ultra-high-resolution optical coherence tomography (OCT) alone as a noninvasive optical biopsy. Dr. Nallasamy, who Dr. Karp mentored, believes artificial intelligence tools can expand this diagnostic approach.

Dr. Nallasamy is collaborating with Kellogg ocular oncologist Hakam Demirci, M.D., to build a comprehensive dataset of pathology and imaging on a variety of ocular surface tumors. Then, they will develop and test an algorithm to distinguish between cancerous and benign lesions, classify cancerous lesions, and differentiate tumor cells from healthy tissue. This automated, virtual biopsy may even help identify subtle features OCT does not recognize.