Skip to main content

Research shows the hidden biases in cardiovascular disease data

BigData@Heart findings underscore importance of accounting for sex-based differences in medical research

28 September 2023
A heart within a wall of data. Image by ImageFlow via Shutterstock
A heart within a wall of data. Image by ImageFlow via Shutterstock

Decades of medical research has in the past failed to account for the different types of risk factors among different population groups. For example, heart disease remains under-diagnosed and undertreated in women because they often have different or less severe symptoms, and this has not been taken into account in past research.

This can have tragic consequences; research funded by the British Heart Foundation found that between 2002 and 2013 around 8 200 women died in England and Wales because they did not receive the same standard of care as men.

Accounting for sex-based differences in research: attitudes are changing

Now research attitudes are changing.

“Looking at funding schemes from charities like the Dutch and British Heart Foundations, it is clearly stated to make sure to take into account sex differences, for example,” says cardiologist Prof. Folkert W. Asselbergs from the Amsterdam University Medical Center. “I think we have to step up to be a bit more aggressive looking at other parameters as well.”

Fortunately new research is beginning to dig into medical data to identify what differences exist, as a first step to correct this problem. Two recent papers were able to compare sex-based differences between heart failure patients as part of the recently-completed Innovative Medicines Initiative (IMI) project BigData@Heart.

“We have to go to a more novel approach and do individualised risk prediction risk prediction. It's not just ‘one size fits all,’” says Prof. Asselbergs, who was part of the project. BigData@Heart brought together representatives from research, industry, legal and ethical bodies, and patient groups to develop new definitions of diseases and outcomes from existing data.

Comparing people who do participate in trials with those who don’t

In the first paper, published in the journal European Journal of Heart Failure, the researchers turned their attention to randomised clinical trials (RCTs) for patients with heart failure and a reduced ejection fraction (where a weakened heartbeat pumps out less blood).

Although such RCTs are supposed to show how experimental heart failure treatments will work for the wider population, the researchers posited that sex differences likely affected real-life outcomes.

In their research, they compared data from two randomised clinical trials and five heart failure observational registries, which track heart failure characteristics, management, and outcomes in patients. This allowed them to classify patients into three ‘subpopulations’; an RCT population had almost 17 000 patient records (21.7% of which were female); a population of 26 000 registry patients who qualified to be included in RCTs (31.8% female); and a population of 28 000 registry patients who did not qualify to be in RCTs (30.2% female).

Researchers then followed what happened within these subpopulations after one year. This included what proportion died from a heart attack or other causes, or were hospitalised with heart failure.

Apart from sex differences, these deaths could have been caused by multiple other factors acting together, so the researchers adjusted for 11 variables including whether patients had diabetes, their age, body mass index (BMI), and others.

Their results showed that the generalising the results of RCTs towards all patients with heart failure and reduced ejection fraction is likely unwise, since females were under-represented in clinical trials. The females who did enrol in trials also had a lower than expected mortality rate (5.6%) compared to females in the registries who were eligible to take part in trials (14%) and those who were not eligible (28.6%).

There could be many reasons behind this, says Prof. Asselbergs. One is that cardiovascular symptoms show up later in women, which gives time for other health issues to also play a role in their health, which in turn might cause them to be excluded from some trials. Another is social factors. “Just to give an example in Utrecht, we demonstrated that people with lower social status didn't provide as much informed consent as others,” he said. “So that means there's a bias already at the front door of research.”

Protein probe uncovers sex-based differences

BigData@Heart investigators also noticed similar skewed results when looking at the profiles of sex-specific cardiovascular proteins that can affect heart failure. In a second paper published in BMC Biology of Sex Differences the researchers explain how they wanted to get a clearer idea of how sex-specific cardiovascular protein profiles, and their associated risk of adverse outcomes, could help understand the processes leading to heart failure and a reduced ejection fraction.

In the study, they tracked data collected from a Dutch cohort study of stable patients with chronic heart failure. The study collected blood samples from 382 patients (104 women and 278 men) every three months. They wanted to compare samples from two points; first a baseline sample from the beginning of the testing phase, and then two samples closest to the PEP (the ‘primary endpoint,’ for example death, a heart transplant, being hospitalised with heart failure, etc).

They then used an assay to identify 1 105 proteins in the patients that have been previously identified with cardiovascular disease, and then used different models and analytical methods to find any sex-based differences in the baseline samples.

After 30 months, 25% of the group’s women had a PEP, compared to 35% of the group’s men.

They also found that two proteins play a part in sex-specific cardiovascular conditions. The protein endothelin-1 (which helps regulate blood pressure by constricting blood vessels), and somatostatin (which inhibits secretions of other hormones) were more strongly associated with a PEP in men, independent of clinical characteristics.

But an even starker conclusion came from analysing the different baseline levels of these risky proteins in men and women; 34 proteins showed higher mean levels in women and another 21 proteins in men. Although these represent only 5% of the total 1 105 proteins analysed, they could also indicate sex differences in specific causes of heart disease.

“If you look at all the biomarkers that we use, we don't use sex-specific specific thresholds, except for kidney function. It should be easy now since we have electronic health records. While we know it’s different, we still don’t do it,” said Prof. Asselbergs.

A tool to avoid the pitfalls of artificial intelligence

Although the project finished in February, he says that he is planning further research in this area, particularly as researchers begin to experiment with artificial intelligence (AI) to discover trends in cardiovascular patient data.

“The stuff that we're now reporting on sex differences will become less visible in the future, it will be in that ‘black box’ of AI,” said Prof. Asselbergs. “I think that as a society, there’s a risk. The research community, together with tech, should be aware of this and make it visible in experiments.”

One way to overcome this is through another key result of the BigData@Heart project – a set of standards so that researchers, hospitals and the medical industry can share and use patient data in a responsible way. It’s called the CODE-EHR framework and has been published simultaneously in three major medical journals (BMJ, the European Heart Journal, The Lancet), and is listed by The Equator Network.

An advantage of this standardisation and harmonisation is possibility of ‘federated learning,’ says Asselbergs. Instead of having an AI algorithm work on patient data gathered into a single source, the model is instead sent to the data-owners and researchers later aggregate their results. This method preserves privacy and enables models to be trained across multiple institutes, increasing the diversity and thereby generalisability of results.

“There is a clear need to build a network of hospitals with harmonized protocols to enable such an infrastructure to really develop and evaluate AI in a responsible way,” says Prof. Asselbergs. “I think that's where we go next.”

BigData@Heart was supported by the Innovative Medicines Initiative, a partnership between the European Union and the European pharmaceutical industry.