Google's DeepMind study on its medical AI is singled out for its lack of rigor

Through a study published in the scientific journal Nature on July 31, 2019, the start-up DeepMind claimed to have developed an artificial intelligence technique capable of detecting acute kidney failure 48 hours before they occur. After investigation, it seems that the subsidiary of Google has advanced somewhat ... and that his research has major biases. Explanations.
Last week, the British start-up DeepMind (subsidiary of the Alphabet group) announced with fanfare to have developed a new technique of artificial intelligence able to detect acute kidney failure up to 48 hours before they occur. Like many other media, relayed the news. But here it is: after investigation, it seems that the method used by the little sister of Google as part of his study, published in the - highly respected - scientific journal Nature, has biases ... and that his system is not so promising that announced.
As Julia Powles, Associate Professor of Technology Law at the University of Western Australia, notes on the Medium website, "no prediction has actually been made." DeepMind would have simply dug into a database of kidney problems of veterans of the US military, then cross-checked them with some 9,000 criteria for each of them. This technique allowed him to bring out broad trends in this given population. A trend that, incidentally, would not work for sure. The study only shows an accuracy of 55.8%, and is even lower when one tries to predict the occurrence of the disease well in advance. Result: the system would generate on average two false positive results for a proven.
The population concerned by this research is, moreover, almost exclusively composed of men (93.6%). This data must necessarily be taken into consideration when we know the importance of the quantity and diversity of samples in the development of a prediction model based on machine learning. "This database used in the study is representative of veterans in US and, as with any learning model, it will need to be fed with more data before it can be applied to a wider audience”.