FDA permits marketing of IDx-DR, the first AI-driven autonomous retinopathy detection device – April 12, 2018

Clinical study (n=900) finds IDx-DR to have 87% sensitivity and 90% specificity; Cloud-based software analyzes images from ~$19k Topcon NW400 retinal non-mydriatic camera

Yesterday, the FDA permitted marketing of IDx-DR, the first medical device to use artificial intelligence to detect diabetes-related retinopathy. The system was reviewed under the De Novo premarket review pathway and granted Breakthrough Device designation in February – the De Novo summary is not yet up on FDA’s database, so we’ll have to wait for more details on what it means for other companies pursuing similar devices (e.g, IBM Watson, Verily). US launch plans and pricing have not been shared publicly, to our knowledge.

The IDx-DR software analyzes retinal images captured specifically by the Topcon NW400 non-mydriatic retinal camera (~$19,000 on The software runs in the cloud and  generates one of two results within 30 seconds: (i) “more than mild diabetic retinopathy detected: refer to an eye care professional”; or (ii) “negative for more than mild diabetic retinopathy: rescreen in 12 months.”

FDA assessed data from a clinical study (n= 900 patients with diabetes) conducted across 10 primary care sites. IDx-DR correctly identified the presence of “more than mild diabetic retinopathy” 87% of the time (sensitivity), and the absence of “more than mild retinopathy” 90% of the time (specificity). Impressive, especially seeing the study in the PCP setting!

Importantly, IDx-DR can be operated by trained laypeople and is intended for use in the primary care clinic – a huge win given the limited number of ophthalmologists and devastating rates of diabetes-related vision loss (~24,000 Americans lose vision each year from retinopathy alone). According to IDx founder Dr. Michael Abràmoff, the software’s autonomy (ease of operation) is also a key differentiator from other retinopathy detection algorithms in development (e.g. IBM Watson, Verily).IDx-DR is currently CE marked as a class IIa medical device and is being distributed throughout the EU in partnership with IBM; we’re not sure what kind of scale it has achieved. Dr. Abràmoff shared in the past that he would welcome the opportunity to partner with IBM for US distribution and noted that the company is in discussions with IBM and others. We’re looking forward to seeing this software on the market, especially in rural areas with low access to optometry care and screening. Indeed, FDA’s release suggests that ~50% of patients with diabetes do not see their eye doctor annually.

For comparison, IBM’s algorithm classifies retinopathy as none, mild, moderate, severe, and proliferative, and Verily’s automated retinal disease assessment (ARDA) conveys more granular scores of damage intensity and score. This additional information wouldn’t necessarily help a PCP make the call of whether to refer to a specialist or not, but the more-detailed report could be very valuable to eye specialists on the receiving end (e.g., tracking progression over time). It’s also possible IDx-DR is capable of more granular classifications, but perhaps that functionality was left out for this initial FDA submission. IBM’s algorithm demonstrates 86% accuracy at classifying images into the five categories, and Verily’s boasts 90+% sensitivity and specificity – we’re not sure about regulatory submission plans for either, though we learned at DTM 2017 that Verily is conducting trials in India.

This regulatory win also suggests some FDA comfort with AI, which could bode very positively for diabetes technology as a whole. We expect to see many more AI-driven software algorithms in the coming years for analyzing CGM and insulin data, providing decision support, and automating insulin delivery (e.g., Medtronic/IBM Sugar.IQ, Bigfoot Biomedical, etc.).


--by Maeve Serino, Brian Levine, Adam Brown, and Kelly Close