قالب وردپرس درنا توس
Home / Health / DeepMind's new AI predicts kidney injury two days before she enters

DeepMind's new AI predicts kidney injury two days before she enters




SEBASTIAN KAULITZKI / SCIENCE PHOTO LIBRARY / WIRED

In 2017, DeepMind began experimenting with a new app at the Royal Free Hospital in London. Called streams, it was designed to help clinicians identify and monitor acute kidney injury (AKI), a condition that is associated with 100,000 deaths in Britain each year. However, unlike most of DeepMind's headlines, Streams does not contain a lot of artificial intelligence.

Instead, medical information such as blood test results and vital signs are brought together in the app and doctors are informed about the kidney health of a patient worsening, using a well-established formula to assess kidney function. Now DeepMind has provided the first evidence that the use of artificial intelligence is a much better way to assess the risk of AKI.

In an article published in the journal Nature DeepMind researchers have shown that they had developed a machine learning algorithm that can predict AKI up to 48 hours. Using an extensive US Veterans Ministry database, the DeepMind team trained an algorithm to predict whether a patient had AKI. In 90 percent of the worst cases where patients needed dialysis, the predictions of the algorithm were correct.

Dominic King, clinical director of DeepMind, hopes that the use of AI to predict patient deterioration would allow clinicians to intervene earlier. In the case of AKI, rehydration, antibiotics, or altered medications can help to restore the patient's kidney function relatively easily. "At the moment, these things are treated too late and the patient is harmed. We believe that these AI systems have a real chance to predict and prevent rather than just what is happening right now. These are doctors who are almost fighting fire and have problems that has already developed, "says King.

The AI ​​system was trained at over 620,000 different data points and eventually identified 3,600 of them who were good predictors of AKI. "The power of deep learning is that you can automatically extract many of these signals if you have enough data to provide them," says Nenad Tomašev, senior research engineer at DeepMind and co-author of Nature . paper.

In order to implement such an AI system, it first has to be trained and tested on many more different data sets. This latest study was performed on historical data taken between 2011 and 2015 and was therefore not used to monitor patients in real time. And although the total dataset comprised over 703,000 patients, only 6.32 percent of them were women, which meant that the AI ​​system was less effective in predicting AKI in tests on female patients.

"While this algorithm would work well for the [Veterans Affairs] population, we would not suggest that you would use it elsewhere without additional training and validation," says King. "There is still a lot of talk and work required for deployment in the real world," he estimates that DeepMind is still at least a year away to test the AI ​​algorithm in real-world clinical environments.

Why Google Consuming DeepMind Health scares privacy professionals

And this is where the streams come into play. Ultimately, King states that streams provide a place where such predictions can be made with AI – not just for AKI, but for other conditions such as sepsis and acute liver failure and diabetes complications.

For this reason, DeepMind has authored three more articles – published on the same day as Nature Paper – examining how useful streams were at the Royal Free Hospital in London. A previous contract between DeepMind and the Royal Free NHS Foundation Trust was closed in 2017 to violate the Data Protection Act. The agreement between the two has since been revised.

The studies paint a mixed picture of the recent success of Streams. The app could not be associated with improvement in renal function or a range of other metrics for determining kidney health. However, this is not completely surprising, King said. This stream release also uses the NHS-recommended algorithm, which estimates the risk of AKI based on the level of a waste product called creatinine in the blood. The problem is that creatinine levels can rise many hours after the onset of AKI. DeepMind's ultimate goal is to replace or supplement this algorithm with a variation of the system demonstrated in the Veterans Affairs study.

And although the app did not do so It appears to be improving kidney health and speeding up and simplifying the treatment of patients with kidney injury risk. Kidney specialists using the app reviewed urgent cases in less than 15 minutes, compared to four hours for those who did not use the app – meaning they only missed three percent of AKI's cases.

DeepMind also estimates that Streams lowers the cost of admission to a patient with AKI by an average of £ 2,123, though this figure does not take into account the cost of delivering streams or the cost of long-term dialysis for AKI patients who remain untreated. King is confident that Streams, if ever implemented in the NHS, will be more cost-effective than the current approach to treating AKI.

This will be the key to determining if the NHS and other healthcare systems are interested in expanding their use of streams. DeepMind seeks to empower health care providers by exploiting its three main arguments: that its technology improves patient outcomes, that doctors and patients actually find it useful and does not increase healthcare costs per patient.

Streams Apparently, the last two points are already ticked, and while the progress on the first is not foreseeable, the Veterans Affairs experiment suggests that an AI-based version of streams has a significant impact on patient health could. And although DeepMind's focus has remained decidedly non-commercial to date, it appears that the company is progressing more slowly to sell its products to healthcare providers.

This may be related to the announcement in November 2018 that DeepMind Health will be available in the future, part of Google, under the supervision of David Feinberg, vice president of Google Health, who took up the position in January 2019. So far, deep-parent parent Alphabet has kept the London-based company at bay, but the integration of DeepMind Health into Google suggests that the team will focus more on commercializing its creations.

Closer collaboration with Google could also bring Streams some additional benefits, King said. If Google knows what to do, it will arrange large amounts of information in a way that is easy to search – something that is critical to streams when it is expanded to cover other conditions as well. "If we summarize this in streams, we can deliver a much more comprehensive and exciting project."

More great stories from WIRED

🖼️ How to use Google Photos for your messy images

😡 Heat waves make people more violent, angry and grumpy.

🚬 England has an ambitious plan to eradicate smoking by 2030.

privacy-first web browser

🎉 A vaccine against Alzheimer's is on the verge of reality

📧 Get the best technical offers and gadget messages in your Inbox


Source link