DeepMind’s data use
DeepMind’s work with kidney patients at London hospitals is perhaps the best-known example of deep learning or artificial intelligence used in medicine here in the UK – not because of the project itself, but because of the criticism it sparked on data management.
Owned by Google’s parent firm Alphabet, DeepMind is a British deep-learning startup, which in 2015 announced it was working with the Royal Free Trust in London on a kidney disease monitoring app called Streams; the app would analyse data and suggest interventions to avoid kidney failure.
Shortly after the trial kicked off, a report in New Scientist revealed that the hospitals had shared a wider range of patient data than expected, across 1.6 million patients. The NHS and DeepMind responded by tweaking the data-sharing agreement, but an academic report from the New Scientist reporter, Hal Hodson, and a Cambridge researcher, Julia Powles, said the “inexcusable” use of data should offer lessons for future projects.
What could DeepMind have done better to avoid such criticism? “At the foundation level, no patient was contacted and no effort was made to address the pivotal consideration in use of health data: that patients understand when and why their data is used, and they have realistic options for effective choice,” said Powles.
She called for other NHS health projects to consider consent and value before starting work. “[Our] paper talks about here-and-now issues, such as patient autonomy, confidentiality and legal compliance,” she said. “We also talk about the essential keys to innovative healthcare advances that work for everyone, such as ensuring that the public gets value for data, that there is effective oversight and accountability, and that egalitarian, public interests are preserved over purely commercial ones.” Failures to get data sharing right only delay the use of the technologies – problematic when they could be helping improve people’s health. Powles pointed to the Connecting Care project in Bristol, Somerset and Gloucestershire hospitals in 2015 as a better example. That saw thirdparty providers linked to an electronic patient record, but first mass-mailed all patients to explain why the data processing was happening, allowing people to opt out if they so wished. That case shows deep learning and data protection don’t need to be in opposition.