Friday 6 May 2016
The true scope of the data-sharing deal between Google’s DeepMind project and the Royal Free NHS Trust, comprising of three hospitals within London, has been revealed. The agreement allows Google’s DeepMind complete access to all patient records across these three hospitals rather than a more targeted subset of data specifically aimed at Acute Kidney Injury (AKI) – the original reason for the data-sharing agreement.
Naturally this has raised various concerns from the public and from national media across the UK. Normally, access to patient data is so often heavily restricted to a specific repository and when access is allowed, especially to private sector companies, the information is made transparent immediately.
This was not the case for DeepMind. Instead they obtained open access to sensitive medical data pertaining the records of up to 1.6 million patients in a somewhat secretive manner. Using what many see as a loophole, consent from patients to share their private medical data with DeepMind is “implied” via the NHS’ Caldicott Information Governance Review regime. Under this regime the Royal Free NHS Trust did not require to contact patients beforehand asking if they would like to opt-out. This has provoked many on-lookers to question the legitimacy of the whole acquisition.
The problem stems from DeepMind’s original announcement on the project, back in February, stating that they were in the process of developing software that can uses algorithms to predict when patients were at risk of AKI and then alerting the relevant clinicians for preventative treatment.
Obviously this was a very targeted project which has now, in the eyes of the public, become a much wider project, prompting questions like, “Why do DeepMind need access to ALL of our raw data?”, “Shouldn’t they be using subsets of data pertaining to previous kidney injury?”
Well, as obvious as that sounds, I can tell you from my experience of working within an analytics company it’s not necessarily accurate.
For DeepMind to fulfil their objective of enabling clinicians to perform preventative treatment on patients at risk of AKI then access to a patient’s complete medical history is required. In order to predict outcomes, all factors of a patient’s medical record should be considered because there are no obvious prior symptoms or conditions before being inflicted with AKI. This explanation can be backed up by a statement coming from a spokesperson for the Royal Free NHS Trust.
In our work within the NHS and other organisations across a variety of sectors, we always stress the importance of being able to completely integrate all data, regardless of whether it is currently stored in silo. The benefits of unifying data provide a much more accurate reflection of the analysis you are able to perform, because it allows you to tap into data sets that, whilst not seeming relevant to the analysis you are undertaking, may actually have an underlying impact.
In my simplistic view, DeepMind’s algorithm will look at the complete set of patient records, identify correlating symptoms between those patients who have had AKI in the past and then apply this knowledge to all patients’ medical records to find those who are most at risk.
Whilst I think that DeepMind’s intentions are genuine, the low level of transparency surrounding what information they would actually have access to was poorly communicated and must serve as a learning experience for us all in the future.