Mike: Hello, I'm Mike Clarke, Podcast Editor for the Cochrane Library. Age-related macular degeneration is a progressive, sight-threatening disease affecting the retina at the back of the eye. There are several Cochrane Reviews of its treatment and, in October 2024 we published a new review of the use of artificial intelligence for its diagnosis. Here’s the review’s first author, Chaerim Kang from Brown University in the USA, to tell us more.

Chaerim: Age-related macular degeneration, or AMD for short, is characterized by central retinal, or macular, damage. Approximately 10% to 20% of the less severe, or non-exudative AMD, cases progress to the advanced, or exudative form, which may result in rapid deterioration of central vision. This is called eAMD and accurate diagnosis is important because it allows patients to receive prompt treatment from a retinal specialist.
Conventionally, diagnosis relies on eye providers and diagnostic imaging, which can be resource consuming. Artificial intelligence might help with this by automatically identifying and categorizing pathological features, enabling timely diagnosis and treatment.
Our systematic review compares the accuracy of these algorithm-based triaging tools versus human experts for diagnosing eAMD. We found that AI algorithms may correctly identify most individuals with eAMD, without increasing unnecessary referrals. However, we have low confidence in the findings because of problems with how the studies were done.
We identified 36 studies, involving more than 16,000 participants and 62,000 images, that reported on 40 algorithms. More than half of the studies were carried out in Asia and the average percentage of participants with eAMD across all studies was 33%. The included algorithms used various retinal image types as model input, such as optical coherence tomography images, fundus images, and multi-modal imaging. Most algorithms used deep neural networks as their core methods.
Only three algorithms were externally validated, meaning that the algorithm was evaluated on a dataset independent of the dataset used to develop it. The statistics used to assess diagnostic tests are sensitivity, which relates to the number of cases that are missed and specificity, which shows how likely it is for people to be mis-diagnosed as having the condition, when they don’t have it. Pooling the results for the three external validation studies, which used nearly 28,000 images, gave a sensitivity of 0.94 and specificity of 0.99, when compared to human graders. This means that using the algorithm to detect eAMD in 10,000 people if 100 of them truly had eAMD would miss around 6 of these (false negatives) and incorrectly identify approximately 99 people as having eAMD (false positives). The pooled sensitivity and specificity for the 28 algorithms that were internally validated, were similar, at 0.93 and 0.96, respectively.
In conclusion, based on low to very low certainty evidence, AI algorithms may correctly identify most individuals with eAMD without increasing unnecessary referrals. However, more rigorous research is needed to assess the accuracy of AI algorithms compared to human experts. This should recruit participants reflecting real-world demographics and disease severity and ensure that the algorithms are externally validated to enhance the generalizability of their findings.

Mike: If you would like to learn more about the current findings and watch for updates incorporating future research, the review is available online at CochraneLibrary.com with a search for ‘artificial intelligence and AMD’.