Researchers recently tested an Australian artificial intelligence (AI) system to determine how well it detects diabetic retinopathy. Unfortunately, findings indicate that the system did not perform well in a real-world setting. Nearly 200 patients were evaluated, and the system identified 17 as having diabetic retinopathy severe enough to warrant a referral. While two of these patients were correctly identified, the system generated 15 false positives.
False positives were driven by a variety of factors, including drusen that looked similar to exudates, uneven light exposure, and dirty lens reflections. However, researchers feel hopeful that training the system to better differentiate these issues can improve AI efficiency at detecting diabetic retinopathy.
These findings were published in JAMA Network Open on September 28, 2018.