Frequently asked questions

We're immensely proud of the work we are doing with our partners across the NHS, and welcome discussion and debate about the best way to improve patient care while protecting patient privacy. 

Given the complexity of the NHS and the sensitivity of patient data, we know that we have an important responsibility to build understanding and trust in our work. Below, we've answered some of the most common questions we hear, covering DeepMind Health, Google, our NHS partners, our use of data and much more.

And if you have a question that is not answered here, please get in touch.


About DeepMind Health

  • What is DeepMind, and what is DeepMind Health?

    We founded DeepMind in London in 2010, with the aim of building AI technologies and proving that they could have positive social impact. 

    DeepMind Health is central to this social mission. We work with the NHS to put the UK's most advanced technology at the service of patients, nurses and doctors.

    Most people get world-class care, but 1 in 10 in-patients suffer some kind of avoidable harm because they don't get the treatment they need. Every single day patients in the UK are dying from curable conditions, because their test results aren't interpreted and acted on in time.

    To address this, we research and build AI and mobile tools that help hospitals get patients from test to treatment, as quickly and accurately as possible. Our priorities are shaped by NHS patients and clinicians, unlike many previous top-down IT projects that have proved costly and ineffective. 

  • Is Google involved in DeepMind's work with the NHS?

    Google acquired DeepMind in 2014, because they were excited about the potential for our technology. As part of this acquisition, we agreed that DeepMind would continue to operate independently, and that we would put our technology at the service of other groups like the NHS in line with our social mission. DeepMind Health is our first effort to achieve this.

    Most importantly,  NHS patient data will only ever be used to help improve hospital care. That means data will never be connected to Google accounts or services, or used for any commercial purposes like advertising or insurance. Doing so would be completely impossible under our NHS contracts and the law, and would go against the ethical code that underpins our culture and work. 

  • What qualifies DeepMind to work in healthcare?

    Our team brings together doctors and clinical academics with world-leading experts in technology and security.

    We work in close partnership with NHS hospitals and patients, spending significant time understanding the challenges they face, the systems they use and the outcomes they need first hand.

    To learn more about our team visit our ‘Meet the team’ page

  • What is DeepMind Health's business model?

    We're lucky to have significant financial resources, and so we're able to carry out research and development without needing to generate immediate income. 

    However, in the long term we'd like to make DeepMind Health a self-sustaining initiative. We aren't looking to maximise profit, but rather to achieve sustainability so we can continue to grow our team, work with more hospitals and help more patients. 

    For Streams, our secure clinical app, we don't have an established business model yet. First, we need to prove that our technologies improve care and reduce costs over a period of time. In the meantime, our partner hospitals for Streams may pay us a very limited amount in support costs.

    Once this is proven, we hope that hospitals will want to pay to use our mobile technology, like they pay for other software that supports care. We'd like to explore some kind of outcomes-based element within this, so that costs are related to the benefits we deliver, but that will be up to our partner hospitals. 

    For our AI health tools, which remain at the research stage, we're further away from a firm business model. There are early efforts underway in the NHS to define the right charging models for AI and algorithmic tools, to ensure that they bring improvements in care while also protecting the British taxpayer. 

    This is an important and emerging area, and we look forward to the NHS community deciding the right way forward. 


  • Will DeepMind commercialise NHS patient data?

    Absolutely not. We only use patient data to help improve care, under the instructions of our NHS partners. Each of our NHS partnerships have strict rules about how data can be used. We will never use patient data outside of these rules. That means data will never be connected to Google accounts or services, or used for any commercial purposes like advertising or insurance.

  • Will AI technology replace nurses and doctors one day?

    We don't think so, no. We're developing much-needed tools to support doctors and nurses provide even better care. They are no substitute for a qualified medical professional’s diagnosis.

  • I have an idea - how can I get involved?

    Whether you're a patient, clinician, NHS Trust - or whether you'd like to join our team - we'd love to hear from you! Please head over to this page for more details. 


We founded DeepMind in London in 2010, with the aim of building AI technologies and proving that they could have positive social impact. 

DeepMind Health is central to this social mission. We work with the NHS to put the UK's most advanced technology at the service of patients, nurses and doctors.

Most people get world-class care, but 1 in 10 in-patients suffer some kind of avoidable harm because they don't get the treatment they need. Every single day patients in the UK are dying from curable conditions, because their test results aren't interpreted and acted on in time.

To address this, we research and build AI and mobile tools that help hospitals get patients from test to treatment, as quickly and accurately as possible. Our priorities are shaped by NHS patients and clinicians, unlike many previous top-down IT projects that have proved costly and ineffective. 

Data and Security

  • What regulations cover your use of data?

    Data is only ever processed by DeepMind Health under the provisions of our agreements in place with our partners, and in compliance with all parties’ information governance requirements and applicable law. 

    In our Streams partnerships, our agreements ensure that patients' data will always be processed in England and won’t ever be linked or associated with Google accounts, products or services. We have established and will maintain the best information security practices, including technical protections, to safeguard this data.

    You can learn more about our data and security processes here.

  • How are the Information Governance policies enforced?

    All staff part of DeepMind Health must undergo NHS Digital training, as well as our internal training programme which specifically assesses their knowledge and compliance with the policies and procedures we have in place. 

    Staff are subject to regular spot checks, and we carry out incident simulations which ensure staff are confident of how to follow procedures during these event types. 

    We also have an Information Governance board which oversees key aspects including reviewing data security reports, approves policy updates, monitors training and reviewing the risk register.

    DeepMind Health has also appointed a panel of Independent Reviewers who meet regularly to ensure independent oversight and scrutiny of all our health work. 

  • Can your partner hospitals check how DeepMind uses patient data?

    Yes. Every time our systems receive or interact with patient data from our partner hospitals, we create a log that hospital administrators can audit later. Those logs are also regularly reviewed by our Information Governance team to ensure that accesses are legitimate, as well as being open to review by our Independent Reviewers.

    We’re also building on this further to give our partner hospitals an additional real-time and fully proven mechanism to check how we're processing data. That will allow our partners to continuously verify that our systems are working as they should, and that data is only being used as it should be.

    Read more about Verifiable Data Audit.

  • How do you protect against data leaks or cyber attacks?

    The DeepMind Health infrastructure has been designed and built to the highest security standards.

    Our security systems and processes have undergone and passed multiple NHS audits. We have also had external penetration tests carried out by CREST certified consultants. 

    All traffic in and out of the infrastructure is restricted and closely monitored, and there are mechanisms which allow us to verify the presence of unusual or unapproved activity. The data itself is encrypted both in transit and at rest. Code is thoroughly reviewed and audited from a security perspective and we analyse any third party libraries we use for vulnerabilities.

    Throughout the course of every project, DeepMind takes rigorous measures to protect the security of patient data.

    Learn more about data and security here 

Data is only ever processed by DeepMind Health under the provisions of our agreements in place with our partners, and in compliance with all parties’ information governance requirements and applicable law. 

In our Streams partnerships, our agreements ensure that patients' data will always be processed in England and won’t ever be linked or associated with Google accounts, products or services. We have established and will maintain the best information security practices, including technical protections, to safeguard this data.

You can learn more about our data and security processes here.

Streams & the Royal Free

  • What is Streams?

    Hospitals like the Royal Free are able to use our Streams app to automatically review test results for serious issues, such as acute kidney injury. If one is found, the system sends an urgent secure smartphone alert to the right clinician to help, along with information about previous conditions so they can make an immediate diagnosis.

    For more detail, please head to our ‘How we’re helping today’ page

  • Why does Streams process identifiable data?

    The Royal Free is the data controller for patient data, so the amount and type of data DeepMind processes is entirely determined by the trust.

    Streams processes personally identifiable data to support direct patient care by notifying doctors of patients who are at risk of Acute Kidney Injury (AKI).

    When a patient with AKI is identified, it's necessary to point the clinicians to that patient. As required by the NHS Digital interface guidelines we must display their name, date of birth, NHS number and gender within the app.


    Learn more about how we process personal data here

  • Is it unusual for a hospital to use a third party service like Streams to process data?

    No. NHS organisations routinely contract with third parties to process patient data, up to and including full electronic patient records services. 

    The data processed for Streams is comparable to the data many other third party organisations also process.

    There are many recorded third-party organisations who receive some level of NHS Digital approval, with the potential to be processing patient data at present. 


    You can also watch a video on how the Royal Free use patient data.

  • Is DeepMind processing more patient data than is needed to diagnose acute kidney injury?

    No. We can only process the patient data that our partner hospitals tell us is necessary to help them care for their patients. 

    The Royal Free currently uses Streams to detect whether patients are at risk of acute kidney injury (AKI). To make this possible, they ask us to process patient information relevant to an AKI diagnosis.

    The primary way to detect possible AKI is via a blood test, which can show whether the kidneys are operating normally. But because other medical factors affect how the kidneys operate, those results differ from person to person.

    For example, pregnancy alters kidney function. That means pregnant women have different test results, so a blood test result that’s healthy for one woman could be dangerously high for another. Including that information in the app means that doctors and nurses can take the most accurate decisions.

    Historical information is also vitally important for doctors and nurses when they're deciding the best treatment. AKI can be challenging to treat, because the blood test result that is used to detect it has to be compared to earlier results when the kidneys were operating normally. Without that extra context, it’s impossible for clinicians to tell if a result reflects normal kidney function or is dangerously higher.

    Other historical data is also relevant. Patients who have had kidney problems in the past, for example, are much more likely to develop very serious forms of AKI, so it’s important that clinicians are made aware of this as quickly as possible so they can give patients the most appropriate treatment. The same is true for whether a patient has recently had emergency surgery, or if they’ve had heart disease.

    To deliver fast and reliable alerts to clinicians through mobile devices, this relevant patient information must be stored in advance, ready to be sent to a doctor or nurse at the first sign of a problem.

    You can find our more about how we are working with the NHS here, or you can see our agreements online in our transparency page
.

  • Why does Streams process data for patients who haven’t had blood tests?

    One of the most important benefits of Streams is that it presents the right patient data to the right clinicians quickly, integrating data from a number of different electronic systems in one place. Clinicians need to have quick access to a wide range of information about patients’ previous medical history and laboratory results (including blood tests) to make safe and accurate diagnoses. 

    For example, if blood tests pick up a patient suffering sudden AKI, it’s important for clinicians to know if the patient has had a previous kidney transplant. To deliver fast and reliable alerts to clinicians through mobile devices, that relevant patient information must be stored in advance ready to be sent to a doctor or nurse at the first sign of a problem in the blood tests.

    Find our more about how we use Streams to help patients here.

  • Why can’t Streams just process data from patients who are in the hospital and known to be unwell?

    AKI can affect patients for a wide variety of reasons - including patients who develop AKI as a consequence of another procedure, such as a hip replacement, or because of another medical condition, such as pneumonia or sepsis. 

    That means it’s very difficult to predict exactly which patients will develop AKI. In addition, AKI often develops without patients showing any symptoms, so it’s also incredibly hard to isolate just those patients who have it from those who do not.

  • Does Streams use AI? If not, then why is an AI company like DeepMind building it?

    Streams doesn’t currently use AI. Right now, we’re simply focusing on getting the right test results to the right nurse or doctor via a secure mobile app. This is an essential first step before any more advanced technology like AI can be introduced.

    Of course, this first step can also bring enormous patient benefits. We’re already hearing stories of people whose care has been helped by Streams, and some nurses are saying that the technology is already saving them two hours each day. 

    Learn more about how we are helping patients today.

  • Do patients at the Royal Free give their explicit consent for Streams?

    Hospitals are the data controllers with a direct relationship with their patients, and they are in charge of decisions about patient consent and opt-outs. DeepMind Health, as a data processor, strictly adheres to the instructions we're given by the hospital.

    In general, hospitals don't ask for explicit consent from patients before using a "data processor", because the NHS remains in control of the patient information throughout.

    You can learn more about how we use personal data here

  • Why did you sign a second agreement with the Royal Free in November 2016?

    Following our work on kidney injury, we wanted to explore whether Streams can be used to improve patient care in other ways, to help make the Royal Free one of the safest hospitals in the world. 

    In November 2016 we signed a new agreement with the Trust, which superseded the previous agreement from 2015. We will be expanding our alerting technology to a range of other potentially fatal conditions, including sepsis and organ failure. Alerting doctors and nurses to patients who need their attention in seconds rather than hours could dramatically improve patient safety. 

    We’re also planning to include the task management features of Streams at the Royal Free, allowing them to view and update records and assign clinical tasks to each other from their mobile, which we hope will eliminate the need for doctors to shuffle through paper and receive pager alerts.

    By knitting this system together into an infrastructure based on open standards, other medical innovators will also be able to develop their own technologies for the Royal Free.

    Finally, we’re also building an unprecedented level of data security with our audit infrastructure, which will allow the Royal Free to verify exactly when and by whom patient information is accessed, with no possibility of falsification or tampering. 

    You can discover a lot more about Verifiable Data Audit here.

    You can find full details of our agreement with the Royal Free here

  • What were the conclusions of the Information Commissioner’s Office investigation into the Royal Free?

    The Information Commissioner (ICO) has now concluded a year-long investigation that focused on how the Royal Free tested Streams in late 2015 and 2016. This testing was intended to guarantee that the service could be deployed safely at the hospital.


    The ICO wasn’t satisfied that there was a legal basis for this use of patient data in testing (as the National Data Guardian said, too), and raised concerns about how much patients knew about what was happening. The undertaking recognises that many of these issues have already been addressed by the Royal Free, and has asked the Trust to sign a formal undertaking to ensure compliance in future.


    The ICO also recognised that the Royal Free has stayed in control of all patient data, with DeepMind confined to the role of “data processor” and acting on the Trust’s instructions throughout. No issues have been raised in relation to the safety or security of the data.


    We welcome the ICO’s thoughtful resolution of this case, which we hope will guarantee the ongoing safe and legal handling of patient data for Streams.

    You can read more about this on our blog here.

Hospitals like the Royal Free are able to use our Streams app to automatically review test results for serious issues, such as acute kidney injury. If one is found, the system sends an urgent secure smartphone alert to the right clinician to help, along with information about previous conditions so they can make an immediate diagnosis.

For more detail, please head to our ‘How we’re helping today’ page

Streams & Imperial

We haven’t yet started processing data for Imperial College Healthcare NHS Trust, but expect to start in 2017 as our technologies go through clinical safety testing followed by a phased deployment.

You can learn more about our partnership with Imperial here

Streams and Taunton and Somerset

Our partnership began in June 2017, when we signed our five-year partnership agreement with the trust to bring Streams’ technology to nurses and doctors at Musgrove Park Hospital.

You can learn more about our partnership with Taunton and Somerset here.

Streams and Yeovil

We have not received or started processing data for Yeovil, but expect to start in 2018.

You can learn more about our partnership with Yeovil here.

Our research with Moorfields

  • What does the project involve?

    Eye health professionals use scans of patients’ eyes to detect and diagnose serious conditions and diseases. Many thousands of eye scans are performed around the UK every day, both in hospital eye clinics and in the community - with more than 1,000 optical coherence tomography (OCT) scans performed every day in Moorfields Eye Hospital alone.

    OCT scans are highly complex and require specialised training for doctors and other eye health professionals to analyse. As a result, there are often significant delays in how quickly patients can be seen to discuss their diagnosis and treatment. To date, traditional computer analysis tools have been unable to solve this problem

    Our research project is investigating how machine learning technology (a form of AI) could help to analyse these eye scans, in order to help clinicians improve the care of patients with sight threatening conditions.

  • What has the research found?

    Our first research paper, published online in Nature Medicine in August 2018, demonstrated a significant step towards our goal of creating technology which would allow eye-care professionals to prioritise patients with the most serious eye conditions and treat them before lasting damage occurs.

    Firstly, as the paper shows [PDF], we have developed technology which can match the accuracy of expert clinicians when identifying a wide range of sight-threatening eye conditions - such as age-related macular degeneration, diabetic eye disease and severe myopia - and recommending the correct course of referral action a clinician should take.

    Alongside this, we have demonstrated that we’ve been able to develop technology that can be applied to real-world scans and existing healthcare systems.  Not only was our research based on OCT scans taken from routine care and benchmarked against real referral decisions used at Moorfields Eye Hospital, but the technology we’ve developed can also provide information that helps explain to clinicians how it arrives at its decisions, rather than presenting them in isolation in a  “black box”. This information includes visuals of the features of eye disease the system has identified on the OCT scan, so nurses, doctors and other eyecare professionals can get insight into its “thinking”, as well as the level of confidence the system has in its diagnosis and referral suggestions, in the form of a percentage.

    We think this will be critically important for clinical application, since it will allow doctors, nurses and healthcare professionals to scrutinise these recommendations to check they are accurate, ensuring confidence in the technology.

    Most exciting for us, we demonstrated that we have been able to develop AI technology which can be easily applied to different types and models of eye scanner, massively increasing the potential number of people across the world that it could benefit, and future-proofing the technology against new devices and models that could emerge in future.

    It’s still early days, but we believe AI could transform the way eye diseases are diagnosed, treated and managed, creating a system which can enable eye care professionals to quickly prioritise patients with the most serious eye diseases before irreversible damage sets in – taking us one step closer to preventing avoidable sight loss.


  • What are the next steps for this research?

    It’s still early days and this technology would need to go through rigorous clinical trials before it could be deployed in hospitals and other clinical settings, but we’re excited about the potential it has demonstrated.

    These results are also only the first phase in a five-year research collaboration, which will explore different challenges with potential AI solutions in this field.


  • Where can I find more detail on the agreement between DeepMind and Moorfields?

    You can request a copy of the research collaboration agreement (with minor redactions for commercial sensitivity) and the ROAD (research on anonymised data) form which was approved by Moorfields Research and Development department by emailing press.office@moorfields.nhs.uk.

    You can also view a copy of a research protocol we've written for this project on the open access website F1000Research.

    You can read the Nature Medicine publication here, and download a PDF of the paper here.

  • How much data has been shared?

    Over the course of the research project, Moorfields Eye Hospital has shared approximately one million de-personalised digital eye scans, used by eye health professionals to detect and diagnose eye conditions.

    Anonymous clinical diagnoses, information on the treatment of eye diseases, the model of the machine used to acquire the images and demographic information on age (shown to be associated with eye disease) will also be shared during the course of the project.

    Learn more about our collaboration with Moorfields.

  • What approvals has DeepMind been given for this research project?

    DeepMind has been given permission for data access via a Research Collaboration Agreement with Moorfields Eye Hospital, and an approval to carry out research from the Moorfields Research & Development team through their Research On Anonymised Data (ROAD) approval pathway.

  • What are Moorfields getting out of this partnership?

    In all of our collaborations with our health partners, we believe it’s critically important that everybody involved shares in the benefits. Our priority is to help our partners at Moorfields and the NHS improve care, reduce strain on clinicians and lower costs, and that is what we are aiming to do.

    While we first need to prove that this technology would lead to a product that is safe and effective in clinical care, we’re incredibly excited at the potential of this research to benefit patients and the NHS. If this technology is validated for general use by clinical trials, Moorfields’ clinicians will be able to use it for free across all 30 of their UK hospitals and community clinics, for an initial period of five years.  These clinics serve 300,000 patients a year and receive over 1,000 OCT scan referrals every day – each of which could benefit from improved accuracy and speed of diagnosis.

    We’re also proud that the work we’ve put into this project will help accelerate many other NHS research efforts. The original dataset held by Moorfields was suitable for clinical use, but not for machine learning research. So we’ve invested significantly in cleaning up, curating and labelling the dataset to create one of the best AI-ready databases for eye research in the world.

    This improved database is owned by Moorfields as a non-commercial public asset, and it’s already been used by hospital researchers for nine separate studies into a wide range of conditions - with many more to come. Moorfields can also use DeepMind’s trained AI model for their future non-commercial research efforts.

  • Will this technology replace doctors and opthalmologists?

    No. The AI technology we are developing is designed to be a tool for doctors, opthamologists and other eye care professionals to help them prioritise the patients in most urgent need of treatment, and free up their time to concentrate on what matters most – treating patients.

    Final responsibility for diagnosis and treatment will always rest with expert clinicians, as it should do.


  • How can patients be sure that no personally identifiable data is being shared with DeepMind?

    De-personalisation procedures are thoroughly validated and formally approved by the Moorfields Eye Hospital information governance team before any data transfer to DeepMind takes place.

  • What processes are in place to ensure the data transferred to DeepMind is only ever seen by the research team?

    A data custodian has been appointed by DeepMind to control access to the data. Only those who require access to conduct the research work will be granted access. All researchers who are involved in the study are required to complete information governance training before beginning research work.

Eye health professionals use scans of patients’ eyes to detect and diagnose serious conditions and diseases. Many thousands of eye scans are performed around the UK every day, both in hospital eye clinics and in the community - with more than 1,000 optical coherence tomography (OCT) scans performed every day in Moorfields Eye Hospital alone.

OCT scans are highly complex and require specialised training for doctors and other eye health professionals to analyse. As a result, there are often significant delays in how quickly patients can be seen to discuss their diagnosis and treatment. To date, traditional computer analysis tools have been unable to solve this problem

Our research project is investigating how machine learning technology (a form of AI) could help to analyse these eye scans, in order to help clinicians improve the care of patients with sight threatening conditions.

Our research with UCLH

  • What does UCLH’s research agreement with DeepMind Health involve?

    Under the agreement, UCLH provided DeepMind Health secure access to de-personalised CT scans of approximately 500 head and neck cancer patients. All patients have consented to their data being used for research purposes

    Learn more about our collaboration with UCLH here

  • What is the purpose of the research?

    This is still in its early stages, but the purpose of the research project is to develop technology which can automatically identify and differentiate between cancerous and healthy tissues on CT scans of head and neck cancer patients to help target radiotherapy treatment.

    At present, this process, known as segmentation, can take clinicians up to four hours to complete manually, as tumours in head and neck patients are situated in extremely close proximity to healthy structures such as the eyes and nerves.

    The research aims to develop artificial intelligence technology to assist clinicians in the segmentation process so that it can be done more rapidly but just as accurately. Clinicians will remain responsible for deciding radiotherapy treatment plans but it is hoped that the segmentation process could be reduced from up to four hours to an hour.

    Longer term this has the potential to free up clinicians to spend even more time on patient care, education, and research, all of which would likely benefit UCLH clinicians and the populations they serve.

    In addition, given that head and neck cancer is one of the most complex tumour sites to treat, if we can develop technology to assist in planning radiotherapy treatment for these tumours, such a breakthrough may well be transferrable to other types of cancer in other parts of the body.

    Learn more about our collaboration with UCLH here

  • What does segmentation involve?

    Before radiotherapy can be given there is a long process of marking the areas to treat on radiotherapy images. Clinicians want to give the highest radiation dose possible to the tumour while avoiding any damage to surrounding areas like the eyes and nerves. To achieve this, the cancer itself and the areas that must be protected from radiation must be painstakingly outlined on radiotherapy images before treatment can begin. This process, known as segmentation, can take up to four hours to complete and is often spread over a number of days.

  • Why are you only doing this research for head and neck cancers?

    Head and neck cancers were chosen because segmenting radiotherapy images for these cancers is one of the most difficult tasks to develop an algorithm for. There are many delicate structures that sit very close together in a patient’s head, making drawing around each one accurately a challenge. In time, we hope that the technology we develop will be used in the segmentation process for cancers in other parts of the body.

  • What has the research found?

    Early findings have found that our AI system can segment several of the most radiosensitive regions on head and neck scans to a similar standard as experts. We’re hoping that this could make the segmentation process faster, resulting in faster treatment for patients, and freeing up time for clinicians.

  • What are the next steps for this project?

    We’re planning to continue improving our models and are planning to conduct a human evaluation with UCLH to test how the model might perform in practice.

  • Where can I find more detail on the agreement between DeepMind and UCLH?

    A summary of the project is available on the Health Research Authority's website. You can also view a copy of a research protocol we've written for this project on the open access website F1000Research,  and our initial findings on arXiv are here.

  • Does the patient data used in the project relate to former or current patients?

    The research involves anonymised scans dating back to 2008 of head and neck cancer patients who have since completed radiotherapy treatment. These patients consented to their de-personalised data being used for research purposes. 

    Under the agreement, UCLH shared approximately 800 de-personalised historic scans dating back to 2008 from 500 historic patients with DeepMind. No scans of patients currently undergoing radiotherapy treatment were shared.

    Learn more about de-personalised data here.

  • What processes are in place to ensure that the de-personalised data transferred to DeepMind Health is only ever seen by the research team?

    A data custodian has been appointed by DeepMind Health to control access to the data. Only those who require access in connection with the research work will be granted access. 

    All researchers who are involved in the study are required to complete Health and Social Care Information Centre (HSCIC) training and internal DeepMind Health information governance training before beginning research work.

  • What approvals have you got for the project?

    The project has been approved by the Health Research Authority (the UK’s national research ethics service) and the UCL/UCLH joint research office.

Under the agreement, UCLH provided DeepMind Health secure access to de-personalised CT scans of approximately 500 head and neck cancer patients. All patients have consented to their data being used for research purposes

Learn more about our collaboration with UCLH here

Our research with the Cancer Research UK Imperial Centre

  • What is the purpose of the research?

    The purpose of this research is to explore whether cutting-edge machine learning (a form of artificial intelligence) could help clinicians detect and diagnose breast cancers more effectively than current techniques allow.

    Early detection and treatment of breast cancer is shown to lead to much higher chances of a full recovery, saving thousands of lives each year (BMJ, 2015), but accurately detecting and diagnosing breast cancer remains highly challenging.

    At present, doctors use mammograms (an X-ray of the breasts) to try and identify cancers early, but breast screening is not perfect. Thousands of cases are not picked up by mammograms each year, including an estimated 30% of interval cancers (NPJ Breast Cancer, 2017), which are cancers that develop in between screenings. At the other end of the spectrum, false alarms and cases of overdiagnosis are also a challenge, often creating unnecessary anxiety for patients and putting increased pressure on health services.

    Working alongside leading breast cancer experts, clinicians, and academics, in partnership with AI health researchers at Google, we’ll be exploring whether machine learning could help overcome these challenges.

    This collaboration will see us carefully analyse historic, depersonalised mammograms from approximately 7,500 women to explore if machine learning tools can spot signs of cancerous tissue on these X-rays and alert expert radiologists more effectively than current screening techniques allow.

    If successful, this has the potential to significantly improve the breast screening process, allowing clinicians to prioritise images for review, leading to fewer missed cancers, fewer false alarms and, we hope, more lives saved.

  • How much data is being shared for this research?

    As part of this research project , the Cancer Research UK-funded OPTIMAM database will provide DeepMind Health with secure access to historic, HRA-approved, de-identified mammograms from around 7,500 women. These digital images have been stripped of any information which could be used to identify patients and have been available to research groups around the world for a number of years. 

  • Do patients have to give their consent for their data to be used?

    No. The HRA approval and DeepMind’s agreement for access to the OPTIMAM database makes clear that the data used in this research is not personally identifiable, and these digital images have been available to research groups around the world for a number of years. When research is working with such data, which is depersonalised with no way for researchers to identify individual patients, explicit consent from patients for their data to be used in this way is not required. (For more information please refer to the ICO code of conduct.)

  • Where can I find out more information on this research agreement?

    For the contracts between Royal Surrey, Cancer Research Technologies and DeepMind please contact the Cancer Research UK press office on pressoffice@cancer.org.uk

    For the contracts between Imperial College London and DeepMind, please contact the Imperial press office on press.office@imperial.ac.uk

  • Why are Google and DeepMind both involved in this project?

    Our partners in this project wanted researchers at DeepMind and Google involved in this research, so that the project could take advantage of the world-leading AI expertise in both teams, and Google’s supercomputing infrastructure - widely regarded as one of the best in the world and the same global infrastructure that powered DeepMind’s victory over the world champion at the ancient game of Go. Together with leading academics and breast cancer clinicians from across the NHS, we believe this combination of partners will achieve more impactful results for patients, which is everyone’s priority.

    You can read more about this in the FAQs on Imperial's website, here

The purpose of this research is to explore whether cutting-edge machine learning (a form of artificial intelligence) could help clinicians detect and diagnose breast cancers more effectively than current techniques allow.

Early detection and treatment of breast cancer is shown to lead to much higher chances of a full recovery, saving thousands of lives each year (BMJ, 2015), but accurately detecting and diagnosing breast cancer remains highly challenging.

At present, doctors use mammograms (an X-ray of the breasts) to try and identify cancers early, but breast screening is not perfect. Thousands of cases are not picked up by mammograms each year, including an estimated 30% of interval cancers (NPJ Breast Cancer, 2017), which are cancers that develop in between screenings. At the other end of the spectrum, false alarms and cases of overdiagnosis are also a challenge, often creating unnecessary anxiety for patients and putting increased pressure on health services.

Working alongside leading breast cancer experts, clinicians, and academics, in partnership with AI health researchers at Google, we’ll be exploring whether machine learning could help overcome these challenges.

This collaboration will see us carefully analyse historic, depersonalised mammograms from approximately 7,500 women to explore if machine learning tools can spot signs of cancerous tissue on these X-rays and alert expert radiologists more effectively than current screening techniques allow.

If successful, this has the potential to significantly improve the breast screening process, allowing clinicians to prioritise images for review, leading to fewer missed cancers, fewer false alarms and, we hope, more lives saved.

Our research with the Department of Veterans Affairs

  • What does your partnership with the Department of Veterans Affairs involve?

    We’re working together to analyse patterns from historical, depersonalised health records to develop machine learning (a form of artificial intelligence) algorithms that could help predict the deterioration of patients in hospital care.

    Patient deterioration – when a patient’s condition worsens and is not recognised or acted upon quickly enough – is a significant global health problem. Across the globe, studies estimate that 11% of all in-hospital deaths are due to patient deterioration not being recognised or acted on appropriately quickly enough.

    Alongside world-renowned clinicians and researchers at the VA, we’ll be analysing patterns from approximately 700,000 historical, de-personalised medical records in order to identify the risk factors for patient deterioration and correctly predict its onset.

    We’ll starting by focusing on Acute Kidney Injury (AKI), one of the most common conditions associated with patient deterioration, and an area where DeepMind and the VA both have expertise.

    Our goal is to find ways to improve the algorithms currently used to detect AKI and allow doctors and nurses to intervene sooner. Eventually, we hope to apply similar approaches to other signs of patient deterioration as well, leading to improved care for many more patients, with fewer people developing serious infections and conditions—ultimately saving lives.

  • How does this differ from the work you are doing with the Royal Free London?

    While both collaborations touch on AKI, the nature of the two projects is completely different.

    Our partnership with the VA is a medical research project. We’re collaborating with researchers to explore the ways artificial intelligence could improve future patient care by examining whether machine learning could be used to predict conditions like, but not limited to, AKI before a patient’s condition deteriorates.

    In contrast, our partnership with the Royal Free involves helping doctors and nurses at the Royal Free care for their patients right now, via our clinical app Streams. Streams processes blood test results to generate instant alerts, which it presents to kidney doctors and nurses alongside other information. By warning them that a patient may be suffering from AKI, and showing them relevant contextual information, clinicians can come to the bedside right away and offer specialist treatment. Streams, and our work with the Royal Free, does not use any artificial intelligence technology, and we’re not conducting any research on this data.

    Patient deterioration is a global healthcare challenge. By collaborating with world-leading researchers in predictive analytics from international healthcare partners, we hope to show how machine learning techniques could benefit patients across the globe.

  • Will the results of this research be integrated into Streams?

    This research project is exploratory in nature, so it’s far too early to say what the results will be and how exactly they might be applied to patient care.

    However, we’ve always believed that machine learning has huge potential to improve outcomes for patients with AKI, and could be key to solving the problem of getting patients from test to treatment as quickly as possible. So, if the research proves as successful as we hope, then – subject to necessary regulatory approvals – integrating the results with Streams is definitely something we’d be interested in exploring in order to further improve patient care in collaboration with our partner trusts.

    To be clear, though, DeepMind will never use machine learning for patient care without all necessary legal, regulatory and contractual approvals, and our partnership with the Royal Free does not currently involve machine learning.

  • How much data is being shared? How much data has DeepMind been given access to?

    Over the course of the project, the Department of Veterans Affairs plans to share the fully depersonalised electronic health records of approximately 700,000 patients with DeepMind for the purposes of this research. The VA has applied all the relevant regulatory standards for processing this data, including, in this case, standard procedures around depersonalisation, to ensure that no patients can be reidentified.

    The data is being stored in Google's secure infrastructure, and kept separate from other Google data. It is encrypted and accessible only to a limited number of employees given explicit permission to do so by the VA. All individual access to the data is also automatically audited and logged in a secure fashion.

  • What approvals has DeepMind been given for this research project?

    This collaboration has been approved by the VA’s local and national Institutional Review Boards, as well as senior officials within the VA. In addition, all data transfers are reviewed and approved by information governance staff at the VA’s data centre, and are subject to a number of US federal data protection laws, including HIPAA regulations.

  • How can patients be sure that no personally identifiable data is being shared with DeepMind?

    The VA used standard industry processes to depersonalise data. No names, addresses, ZIP codes, dates or social security numbers or other unique codes have been sent to DeepMind, and nor has any detail on a patient's age or location, in accordance with federal law.

    In addition, any statistical data, like a blood test result, was adjusted by the VA before transfer to make it impossible to identify any patients from the dataset. This process is known as adding “statistical noise”, which allows researchers to analyse a dataset for patterns without being able to identify any of the original data.

  • What happens to the data at the end of the agreement?

    DeepMind must securely destroy all copies of anonymised data received through the agreement.

We’re working together to analyse patterns from historical, depersonalised health records to develop machine learning (a form of artificial intelligence) algorithms that could help predict the deterioration of patients in hospital care.

Patient deterioration – when a patient’s condition worsens and is not recognised or acted upon quickly enough – is a significant global health problem. Across the globe, studies estimate that 11% of all in-hospital deaths are due to patient deterioration not being recognised or acted on appropriately quickly enough.

Alongside world-renowned clinicians and researchers at the VA, we’ll be analysing patterns from approximately 700,000 historical, de-personalised medical records in order to identify the risk factors for patient deterioration and correctly predict its onset.

We’ll starting by focusing on Acute Kidney Injury (AKI), one of the most common conditions associated with patient deterioration, and an area where DeepMind and the VA both have expertise.

Our goal is to find ways to improve the algorithms currently used to detect AKI and allow doctors and nurses to intervene sooner. Eventually, we hope to apply similar approaches to other signs of patient deterioration as well, leading to improved care for many more patients, with fewer people developing serious infections and conditions—ultimately saving lives.