It has been tested in 11 hospitals in Thailand.
Artificial intelligence raises a lot of hopes in the medical field. . In September 2019, researchers claimed that AI can make a diagnosis more accurately than a human. Google Health wanted to get out of the labs and test its AI in real condition.
AN AI trained to detect a specific disease
Google decided to wipe the plasters and get out of the controlled space of the search center. The AI of the company’s health division has been awarded the CE label of the European Union, which allows it to be tested in real-world situations
The opportunity presented itself in Thailand, where the European label is recognized. The Thai Ministry of Health has put in place a plan to screen for diabetic retinopathy in 60% of people with diabetes. A disease that can cause partial blindness if not managed in time.
Problem,the country has 200 specialists for 4.5 million patients. Nurses must take en pictures of patients and send the images to specialists. A process that can take up to 10 weeks.
This is where Google and its deep learning system, which is trained to detect signs of eye disease, come in. AI looks for signs of diabetic retinopathy (suchas blocked or leaky blood vessels) from an eye scanner. . Performing in 90% of cases, the success rate of screening is equal to that estimated by “human specialists”.
11 Thai hospitals were equipped and for several months, the team of Emma Beede, researcher in UX (specialized in the study of the user application) at Google Health observed and exchanged with the nurses.
Without being catastrophic, the AI did not live up to the expectations placed in it. Once in five, she proved unable to give a result and sent the patient back to a specialist.. Frustrating for nurses, the cases being regularly obvious.
The problem has been identified, the AI has been trained using very high quality en scanners, and below a certain image threshold it is no longer able to function. Nurses had to waste time trying to get better image quality.
The other difficulty was, in some hospitals, the use of the cloud. A bad connection and the consequences are immediate,delays accumulate.
When everything goes well en on the other hand, the AI works wonders. Emma Beede noted that “a nurse examined 1,000 patients on her own,with this tool, she is unstoppable.” On the plus side, patients were not particularly afraid that an AI would be diagnosed.
Google Health is a step forward in proving that laboratory results and quickly trumpeted are only valuable in the field. In Thailand, researchers continue to work, several promising solutions are being studied: giving nurses the responsibility to judge borderline cases that AI has not been able to deal with, and for the latter, a system modification to manage lower quality images.