In 2018, Google Health began a program in Thailand to screen for diabetic retinopathy using artificial intelligence (AI). The AI was designed to analyze photos of diabetic patients’ eyes to detect signs of eye disease. The AI was promising in theory – during testing, it was 90% accurate in detecting diabetic retinopathy in eye scans, according to MIT Technology Review. However, in practice, the AI presented various problems for clinical staff. For example, the process of capturing images varied across clinics, many eye scans were rejected by the algorithm due to picture quality issues, and internet issues in the clinics delayed the image uploading process.
This is a reminder that AI needs to be responsive to the ways clinical staff operate in the real world.
Even though the AI demonstrated accuracy in the lab – the algorithm detected eye disease just as well as an ophthalmologist – field performance can be a different story. Machine learning experts need to have input from healthcare professionals before the algorithm is deployed. For example, what would doctors in the field do with an eye scan that is not clear, or that is a borderline case? Perhaps we should expect AI to do more than just reject any imperfect image, when experts in the field would engage in reasoned debate about the interpretation of the image.
According to CBI Insights, Google – more than any other tech company – is focusing on software and artificial intelligence in the healthcare arena. Google plans to use health software for data generation and lifestyle management tools, with their AI efforts focused on disease detection. Despite the hiccups in the Thailand diabetes study, Google looks to be responsive to the challenges of AI – their researchers assert that “the success of a deep learning model does not rest solely on its accuracy, but also on its ability to improve patient care.” That is the right attitude which will help healthcare AI find greater success in coming years.
Read more AI content here: AI Tag