AAO 2019 DailyOCT 13, 2019

AI Can Transform Medicine—But, First, Clinicians Have Homework to Do

Add to My To-Do ListComments Views

Artificial intelligence (AI) has the power to transform medicine through early diagnosis and tailored treatments. But harnessing that power depends on consensus guidelines that only human disease experts can create, said Naama Hammel, MD, speaking at Sunday’s Spotlight session on Artificial Intelligence and New Technologies for the Ophthalmologist.

What AI needs from ophthalmologists. Dr. Hammel traced the power of AI back to its origins and gave clinicians a homework assignment: Decide on clear disease definitions and severity scales. In turn, these can be used to generate consistent, accurate labeling for machine-learning algorithms.

How ophthalmologists can use AI. Machine learning can achieve two overarching goals in medicine, said Dr. Hammel, who is an ophthalmologist and clinical research scientist at Google. Algorithms can automate repetitive tasks—things clinicians can do but prefer not to, perhaps because they’re stretched for time or lack manpower. Beyond that, machine learning can detect new signals in data that elude human perception: the subtle beginnings of disease in a symptom-free patient, for example, or anatomic changes that point to an elevated risk of disease.

Deep learning. Dr. Hammel described her work in creating a deep learning algorithm that can analyze fundus photographs and predict whether patients with intermediate age-related macular degeneration will progress to neovascular AMD within one year. The algorithm outperformed manually graded 4-category and 9-step scales in a published dataset. “This suggests there is information in the photo beyond what we see with our eyes,” Dr. Hammel said.

Similarly, Google created a deep learning model that uses color fundus photographs to predict diabetic macular edema with the accuracy of optical coherence tomography (OCT). “This is exciting because it means that maybe we can use fundus photos to give more accurate care in settings where there are no OCTs,” she said.

Wanted: More data and definitions. The bottom line, said Dr. Hammel, is that machine learning algorithms are only as strong as the data used to train them. The promise of AI can only be achieved once clinical experts reach consensus on the definitions and stages of disease.  —Anni Delfaro

Financial disclosures: Dr. Hammel: Google: E.

Disclosure key. C = Consultant/Advisor; E = Employee; L = Speakers bureau; O = Equity owner; P = Patents/Royalty; S = Grant support.