Google AI lead: Machine learning 18 months away from NHS hospitals

However, in an interview with HSJ, Dominic King also said his “optimistic” prediction about the use of machine learning tools in the UK – which he added would still be in very limited circumstances – would depend on overcoming “regulatory hurdles”.

In the US, DeepMind has used a “deep learning approach” on a large number of health records from the Department of Veterans Affairs system to show it could predict the onset of acute kidney injury – a common and serious condition often developed in hospital – up to 48 hours in advance. A paper on the project is published today by Nature.

Speaking to HSJ, Dr King – who has trained as a surgeon and has worked across digital health, health policy, and behavioural science – said this had not been possible in the NHS so far, but “optimistically” he believed this kind of approach using “carefully controlled… early pilots” could be trialled in UK hospitals in around 18 months.

However, he said there would be regulatory barriers that would need to be tackled before it was possible, adding agencies, including the UK’s Medicines and Healthcare products Regulatory Agency, were working on how they can regulate these kinds of tools. He said that, in some ways, the NHS was very well placed.

Dr King said: “Algorithms have existed in medicine for decades, [but] what we are talking about here are systems that are able to constantly learn from experience.

“Ideally, every new patient that comes into the hospital, the system learns a bit more from.”

It will also depend on testing how the US analysis applies in the UK, and establishing how clinical teams will react to receiving alerts much earlier, he said.

Dr King said it required rich and extensive de-identified patient data, which has not been available in the UK until recently. Michael Macdonnell, another director at DeepMind, added that AI may be applied clinically in the NHS more quickly than 18 months for reading diagnostic scans.

Streams

DeepMind has also today published papers evaluating use of its digital health app Streams at the Royal Free London Foundation Trust. This uses a non-AI system to alert doctors when patients are developing signs of AKI.

According to the report, the alert resulted in urgent cases being reviewed and addressed more quickly (in less than 15 minutes), fewer cases of AKI being missed, and an average cost reduction per AKI patient of 17 per cent. There was no firm evidence of a related improvement in outcomes, however.

DeepMind is now working to broaden the app to predicting other aggressive conditions, such as sepsis and pancreatitis, and factors such as likely time to discharge and chance of readmission. It will also talk to more potential NHS partners to establish what they would need to do to be ready to use the new technology.

Dr King said trusts needed “a basic level of digital maturity” to embrace the benefits of these studies, which many do not have. The Royal Free and DeepMind’s other UK partner – Taunton and Somerset FT – are known for having more advanced core IT systems than most.

He said DeepMind’s work so far showed there was high potential for its technologies to improve care and save money, although initially it often required clinical teams to reorganise themselves and could feel like it produced more work.

He said: “The reality is 60 to 70 per cent of all physician clinician time now is spent on non-direct care using these clunky electronic systems, spending 10 minutes logging in to find one blood test.

“Whereas you can do it in seconds with modern secure technology. We haven’t realised the benefits yet in the NHS or indeed in any health system of investment that has been made in technology.

“I think we are on the cusp of having technology like Streams that makes a really positive difference. It can support the delivery of better outcomes, it can improve experience, it can ideally reduce cost too.”

He added: “We have learnt a huge amount from the Royal Free and [other work at Imperial College Healthcare Trust] about how you support staff being on mobile and they are very happy to share those lessons with other trusts too.

“We are laying the ground work to scale this up. We are interested in scaling the routes of Streams but also the breadth of functionality.”

The Streams evaluation highlighted “resource constraints” once a clinician had been alerted to a patient that is at risk of developing AKI. Dr King said he does not want the technology to “generate huge amounts of needs for change”.

In its early stages, the pioneering work between Google DeepMind and the Royal Free became very controversial because of concerns about data sharing. It was criticised by the Information Commissioner’s Office in 2017.