magdal3na - Fotolia

NHS data not fit for AI, Lords select committee told

House of Lords artificial intelligence committee hears evidence from experts about the challenges of sharing NHS patient data

Artificial intelligence (AI) offers the NHS a huge opportunity to complement clinicians, cross disciplinary boundaries in medicine, improve clinical decision support and provide better patient self-care, the House of Lords select commitee on AI has heard.

Among the biggest benefits of AI discussed by expert witnesses at the committee’s third hearing was image analysis, where an AI would be able to identify cancerous tissue more accurately.

Nicola Perrin, head of understanding patient data at the Wellcome Trust, said: “There are huge opportunity. We are beginning to see a few pockets of usage.”

But Dr Sobia Raza, head of science at health policy think-tank the PHG Foundation, reminded the committee that AI is contingent on using large datasets.

This was an area discussed during the previous AI select committee hearing earlier in November.

There was general consensus among the expert witnesses that data sharing would be a major challenge for the NHS.

Julian Huppert, chair of the independent review panel for DeepMind Health, said: “A huge amount of work is needed to make the NHS more digital. We would get much more value from NHS data if it was in a secure, centrally managed system.”

According to Huppert, NHS trusts have many different systems, and some hospitals run hundreds of databases that “don’t talk to each other”.

Martin Severs, medical director at NHS Digital, said: “Medical data is very chaotic at source. My phone is more powerful than many of the computers in hospitals. There is a lot of focus in the media about the development of algorithms, but very little focus on the preparation of data.”

Concern over data misuse

The committee was concerned that web giants such as Amazon, Google and Facebook might misuse patient data. Perrin said: “There has been very little formal analysis of what people think about the use of AI in healthcare.”

Hugh Harvey, a clinical AI researcher and consultant radiologist at Guy’s and St Thomas’ NHS Foundation Trust, said: “I would be happy to make aggregated data available. Patient data should be limited by access and be de-identified and constrained by a legal contract to balance the benefits of data and the risk to privacy.”

Harvey said there would be a societal benefit if anonymised data was used to develop an AI product for the NHS.

Some members of the committee asked whether it would be viable to charge for the use of NHS data.

Severs said: “All the data the NHS holds is funded by the British taxpayer. Any use of that data should generate benefits back to the taxpayer. While we should open up enough data as possible for specific research and use cases, within those data-sharing agreements there should be a return on investment on that data. There is billions of pounds’ worth of value in this data. We need to encourage innovation and allow failure at low costs, but there needs to be a return on investment of that data back into the NHS.”

The witnesses also discussed how useful an AI made outside the UK would be in the NHS. Severs warned that because AI data is more granular than a generic pharmaceutical, it would be hard to know how well it could work in a generalised way. He said any system should be tested on a wide proportion of the UK population.

“There are many examples of diseases across the world that we just don’t see in the UK,” he said. “We have different genotypes and phenotypes of humans across the globe and if they are represented in algorithms developed abroad, you will be getting some errors when applied to a UK population.”

Read more about NHS data and AI

Given that any healthcare AI will need vast amounts of NHS data, committee members were concerned that data access may be curbed by the EU’s General Data Protection Regulation (GDPR), which come into force in May 2018.

Harvey said: “Under GDPR, you have the right not to be subject to an automatic decision. This means you have to demonstrate to the public that when a decision is taken automatically, it is based on sound evidence and the public is getting a good deal.”

Harvey’s view is similar to that of Dame Fiona Caldicott, national data guardian for health and care, Office of the National Data Guardian, who spoke earlier in the session and said: “We have to improve the systems, and take the public with us.”

Unlike previous hearings, which focused on business opportunities, skills and job losses arising from automation, the main issue raised in this latest hearing was not the benefits of AI to the NHS, but the poor state of NHS IT.

As Computer Weekly has reported previously, the Open data in the health sector report, which was commissioned by NHS England in 2016, found that healthcare data is often fragmented and duplicated.

The witnesses warned the select committee that because the state of the NHS’s data stores is seen as quite poor, it is probably not in a great place to make use of AI. Some of the data may still be paper-based, and some in an unstructured form that cannot be accessed easily. No figures were given for the cost of digitising these records, but the general consensus was that it could be very large.

An internet giant or a Silicon Valley startup could help the NHS structure its patient data in an easy-to-access way, but witnesses raised concerns that the general public may not want to have their medical data shared with such companies. There was a sense that anonymising medical data would not be sufficient, because it could quiet easily be traced back to identify an individual. Instead, the experts suggested a form of data contract under which only explicit use of the data would be granted.

Read more on Artificial intelligence, automation and robotics