In 2017, DeepMind started trialling a new app with the Royal Free hospital in London. Called Streams, it was intended to help clinicians identify and monitor acute kidney injury (AKI) – a condition linked to 100,000 deaths in the UK every year. But unlike most of DeepMind’s headline-grabbing work, Streams doesn’t contain a jot of artificial intelligence.
Instead, the app brings together medical information, such as blood test results and vital signs, and notifies clinicians when a patient’s kidney health deteriorates, using a well-established formula for evaluating kidney function. Now DeepMind has provided the first hints that using artificial intelligence might be a much better way of assessing whether someone is at risk of AKI.
In a paper published in the scientific journal Nature, DeepMind researchers showed that they had created a machine learning algorithm capable of predicting AKI up to 48 hours before it happened. Using a vast database from the US Department of Veterans Affairs the team at DeepMind trained an algorithm to predict whether a patient would suffer AKI, and in 90 per cent of the worst cases, when patients ended up requiring dialysis, the algorithm’s predictions were accurate.
DeepMind’s clinical lead, Dominic King, hopes that using AI to predict patient deterioration would allow clinicians to intervene earlier. In the case of AKI, rehydration, antibiotics or altering medications can help restore patients’ kidney function fairly easily. “Currently we pick these things up too late and harm is caused to patients, and we think there's a real opportunity for these AI systems to be able to predict and prevent rather than just what currently happens, which is clinicians almost firefighting and running around problems that have already developed,” King says.
The AI system was trained on over 620,000 distinct data points, with it eventually identifying 3,600 of them that were good predictors of AKI. “Where the power of deep learning lies is that it allows you to extract a lot of these signals automatically if you have enough data to provide to it,” says Nenad Tomašev, a senior research engineer at DeepMind and a co-author on the Nature paper.
But actually deploying this kind of AI system will first require training and testing it on much more diverse datasets. This latest study was conducted using historical data, taken between 2011 and 2015, and so wasn’t being used to monitor patients in real-time. And although the total dataset contained over 703,000 patients, only 6.32 per cent of them were women, which meant that the AI system was less effective at predicting AKI when it was tested on female patients.
“While this algorithm would work well for the [Veterans Affairs] population, we wouldn't be proposing that without extra training and validation that you would use it elsewhere,” says King. “There are a lot of conversations and work that still needs to happen for us to deploy this in a real world setting,” he says, estimating that DeepMind is still at least a year away from starting to test the AI algorithm in real clinical settings.
Read more: Why Google consuming DeepMind Health is scaring privacy experts
And that’s where Streams come in. Ultimately, King says that the point of Streams is that it offers a place to surface these kinds of AI-powered predictions – not just for AKI, but also for other conditions including sepsis, acute liver failure and diabetes complications.
That’s why DeepMind has written three more papers – published the same day as the Nature paper – evaluating how useful Streams has been at London’s Royal Free Hospital. An earlier deal between DeepMind and the Royal Free NHS Foundation Trust was found in 2017 to break data protection law. The deal between the two has since been revised.
The studies paint a mixed picture of Streams’ success so far. The app couldn’t be linked to any improvement in kidney function or a number of other metrics used to determine kidney health. But this, King points out, is not totally surprising. This version of Streams still uses the NHS-endorsed algorithm that 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 spike many hours after AKI has already set in. DeepMind’s eventual goal is to replace or supplement this algorithm with a variation of the system demonstrated in the Veterans Affairs study.
And although the app didn’t deliver an improvement in kidney health, it does seem to make it quicker and easier to treat patients at risk of kidney injury. Kidney specialists using the app reviewed urgent cases in fewer than 15 minus compared with four hours for those not using the app – meaning they missed only three per cent of all cases of AKI.
DeepMind also estimates that Streams reduces the cost of admitting a patient with AKI by an average of £2,123, although this figure doesn’t take into account the cost of providing Streams or the cost of long-term dialysis in patients with AKI that goes untreated. King says that he’s confident that – if it is ever implemented in the NHS – Streams would end up being more cost-effective than the current approach to handling AKI.
This will be key to whether the NHS and other healthcare systems are keen to expand their use of Streams. In getting healthcare providers on board with its technology, DeepMind is trying to leverage three key arguments: that its technology improves patient outcomes, that clinicians and patients actually find it useful and that doesn’t increase per-patient healthcare costs.
Streams already seems to be ticking boxes on the last two points and while its progress on the first is yet to be seen, the Veterans Affairs experiment suggests that an AI-powered version of Streams could make a significant impact on patient health. And while DeepMind’s focus until now has remained decidedly uncommercial, it looks like the firm is slowing ramping up to start selling its products to healthcare providers.
This might have to do with the announcement in November 2018 that DeepMind Health is set to become part of Google, under the oversight of Google Health vice president David Feinberg, who took up the role in January 2019. Until now DeepMind parent company Alphabet has kept the London-based firm at a distance, but DeepMind Health’s integration into Google suggests that the team will start to place greater emphasis on commercialising its creations.
Working more closely with Google might also provide some added benefits to Streams, says King. If there is one thing Google knows what to do, it’s ordering large amounts of information in a way that is easily searchable – something that will be critical to Streams if it is eventually expanded to cover other conditions too. “Bringing this together into streams allows us to deliver a much more comprehensive and exciting project,” he says.
This article was originally published by WIRED UK
