Almost all AI-powered medical devices have been trained in only one country. Almost half in China. This is shown by a new study and warns of risks for patients.
Artificial intelligence – AI – has also found its way into medicine. Strongest in radiology. Here, artificial intelligence helps, for example, to make an exact diagnosis more quickly based on CT scans.
Thomas Frauenfelder, director of the Institute of Radiology at the University Hospital of Zurich, says that AI primarily helps with routine tasks in areas with a lot of data. He demonstrates this using images of the lungs of a tumor patient.
These are all aids that enable me to concentrate on the essentials and the complex tasks.
The AI software processed the 380 images from the computer tomograph within five minutes and eliminated everything that was irrelevant to the analysis. The radiologist can therefore see at a glance that there is a nodule in the lung, a so-called pulmonary nodule – a clear indication of metastases in the lung.
Frauenfelder compares AI in radiology with the many assistance systems in modern cars: “These are all aids that enable me to concentrate on the essentials and the complex tasks.”
Half of the studies come from China
AI is booming in medicine. Kerstin Noëlle Vokinger, professor of law and medicine at the University of Zurich, examined this boom. In a study that will soon appear in the renowned New England Journal of Medicine – and has already been published online – she examined all clinical studies worldwide that have been published since 2010 on AI-supported medical devices.
She finds that almost half of all studies come from China (45 percent) and that almost all studies (97 percent) are only carried out in a single country. That could be tricky, says Vokinger to the “Tagesschau”.
If, for example, an AI software has only been trained with Chinese patients, it could potentially lead to incorrect results if it is used on non-Chinese patients: “Population groups can have different risk factors and characteristics for certain diseases. A specific example is the diagnosis of skin tumors. We have different skin types. And if such an AI medical product is used on patients trained with one skin type but then applied to patients with a different skin type, this can lead to different and even incorrect results.
AI regulation is a balancing act
AI in medicine not only has advantages, but also risks. Vokinger therefore demands that Switzerland must require the manufacturers of such AI products to transparently show how the products were trained. Vokinger, however, is critical of the idea of a separate certification for medical AI products.
On the one hand, we have to ensure patient safety. On the other hand, we want and must promote innovation in our country.
AI regulation is a balancing act: “On the one hand, we have to ensure patient safety. On the other hand, we want and must promote innovation in our country. In any case, it is important that Switzerland addresses the question of how it wants and should regulate AI. In our study we also see that a large number of such products are likely to come onto the market, especially from other countries.”
So far, Switzerland has no specific regulation of AI medical devices and is waiting to see what the EU and the USA do in this regard. “We already have the opportunities today, but also the risks. “In this respect, it would be good if Switzerland made progress here,” warns Vokinger.