Leave no patient behind: lessons from genetics on the risks of biased data

In the last two months, my Dad has been shuttling between various GP surgeries and hospitals, trying to get the dosage right for his warfarin treatment. This has required 15 appointments so far, with frequent frustration when the test results show the dosage is still wrong. As Dad rearranges yet another day around a visit to the doctor, his experience seems far from the vision of personalised care enabled by new technology, genetic testing and big data, as promised by NHSX and the Topol Review.

Reasons to be hopeful?

Despite Dad’s experiences, as a genetic researcher I’m optimistic about the potential for technological innovation to improve care. In fact, I know that scientists are developing genetic tests to help doctors find the right dose of warfarin for each patient much more quickly. These tests could reduce the heavy burden of so many appointments for people like my Dad, enabling more effective, efficient and safer treatment (as well as better use of NHS resources).   

But this vision has its challenges. Many of these new innovations are underpinned by data collected from patients, either from large research cohorts or electronic health records. But these data often do not capture the real-world diversity of patients. Researchers may choose to focus on certain patient groups to reduce the complexity of the problem, perhaps excluding those with pre-existing conditions, for instance.

Even national databases of electronic health records will contain biases. For example, some groups may be less likely to seek treatment and so their needs are not captured in the data. Reliance on these skewed datasets alone can lead to health care innovations that undermine one of the core principles of the NHS – that it meets the needs of everyone.

Problems with data cause problems with care

Innovators across health care technology are grappling with this issue of unrepresentative data. The field of genetics demonstrates the problems caused for patients if we don’t get this right. Our understanding of genetics is often built from the analyses of large international collaborations. Yet the datasets on which genetic innovation is developed still exclude many populations. People of European ancestry make up 79% of all participants of genetic studies, despite being only 16% of the global population.

This omission has a real impact on patients. Because we know less about genetic variation in non-European populations, genetic tests are more likely to give ambiguous or uninterpretable results for these patients, excluding certain groups from the many benefits that genetic tests promise.

Even worse, patients may also be given incorrect medical information. In the USA, because of biased data, patients of African ancestry were wrongly told that they have a high risk genetic mutation for hypertrophic cardiomyopathy (a serious heart condition). This false positive can lead to unnecessary treatment, as well as anxiety for patients and any relatives erroneously told they may also be at risk. These errors could have been prevented with even small increases in the diversity of the patient ancestry included in the cohorts used to develop these tests.

Redressing the data balance in order to reduce health inequalities

Frustratingly, the risk of genetic tests exacerbating existing health inequalities is sometimes only picked up after testing reaches the clinic, rather than as a fundamental step in developing the technology. I am glad to say that others have recognised this risk too, and there are efforts within genetics to redress the imbalance and expand the base of patient data on which genetic models are built.

The Health Foundation has previously highlighted four key areas for ensuring that technology best serves the needs of the NHS and its patients: clarity about the purpose of the technology; transparency about NHS requirements with industry; obtaining fair value when sharing data; and building internal analytics capability. These areas should each be underpinned by the guiding principle that new technologies must serve all NHS patients, including those who face greater challenges and disadvantages regarding their health.

If we want to use innovation to strengthen the NHS and use technology to help reduce (rather than worsen) existing health inequalities, we must address the impact of biased data. First steps include:

1. Involving patients from a diverse range of backgrounds during the early design stages of any new technology.
2. Creating more diverse datasets on which we can build new technologies.
3. Ensuring transparency from developers about the data they are using and encouraging them to proactively identify hidden biases within that data.
4. Evaluating the impact of any new technology on the health system, and across a diverse range of patients.

Without addressing this issue upfront, the NHS is at risk of leaving certain patient groups behind in the rush towards a technology-focused future. Thinking about people like my Dad, struggling with seemingly endless medical appointments, we must ensure that the promise of more efficient and personalised care is realised not just for some, but for every single patient.  

Karen Hodgson (@KarenHodgePodge) is Senior Data Analyst at the Health Foundation