Smart, But Not Smart Enough
AI image interpretation will not solve the diabetic retinopathy epidemic
Steve Charles | | Opinion
Many ophthalmologists and researchers believe that the use of artificial intelligence (AI) to interpret digital retinal images represents a solution to the well documented worldwide diabetic retinopathy epidemic. However, there are both tactical and strategic issues that are seemingly overlooked by those promoting this “solution.” And I would like to explore some of them here.
Diabetic retinopathy is a function of poor serum glucose control as shown by the DCCT and many other worldwide, multi-center clinical trials. The worldwide obesity explosion both in developed and developing countries is a core problem driving the rapid growth of diabetic populations. Diet education starting in early childhood is a crucial part of addressing this epidemic. Access to preventive medicine, diabetes medications, and frequent or real-time blood sugar monitoring is a huge financial burden but less costly than waiting until diabetic retinopathy and other complications develop to initiate care.
Socioeconomic and cultural issues are far more important than a perceived shortage of ophthalmologists to read digital fundus images. Assuming AI was effective – who would pay for technicians to acquire images, who would purchase the imaging devices, who would pay to transport patients to the imaging device or the device to the patient? And even if all these issues were addressed, who would inject anti-VEGF agents or implant sustained delivery devices, and who would pay for the treatment? Anti-VEGF therapy has been shown to be more effective than laser photocoagulation; who would pay for the laser treatment and lasers if we moved back to this outmoded therapy? These socioeconomic and cultural issues far outweigh a perceived image interpretation burden. In reality, there is not a backlog of unread images.
All ophthalmologists frequently see diabetic retinopathy patients that fundus examination and digital fundus images would be interpreted as inactive or stable, but OCT demonstrates diabetic macular edema. Widespread availability of low-cost, portable and reliable OCT devices that can be operated by minimally-trained individuals is mandatory for screening to be effective. Trained graders perform as well or better than ophthalmologists in reviewing fundus images, as has been shown in clinical trials; this approach could be applied to OCT images as well.
And so, while I agree that we do have a serious problem with a diabetic retinopathy epidemic – I do not believe AI is a serious solution.