Artificial intelligence enabled obstetric cardiovascular screening in SPEC-AI Nigeria

I’m Dr Demilade Adedinsewo, I am a non-invasive cardiologist and I practice at Mayo Clinic in Jacksonville, Florida [United States].

So AI [artificial intelligence] has been employed for use as part of cardiovascular screening in obstetric patients. Using the ECG is the one example that is most common. In this study, we employed artificial intelligence to analyze data that is recorded or captured with a digital stethoscope in the form of heart sound and an ECG, as well as a standard clinical ECG. This AI tool is able to predict the likelihood of cardiac dysfunction or heart failure.

To provide a bit of background about SPEC-AI Nigeria, heart disease is the leading cause of death during pregnancy and the postpartum period, which is the period after childbirth up to 12 months, in the United States. Black women have been reported to have a higher risk of dying during pregnancy or following childbirth – actually, a threefold higher risk of death. When we also look at that late postpartum window, which is that period after 6 weeks through to 12 months postpartum, cardiomyopathy or heart failure is actually the number one cause of death in the United States. Now talking about Nigeria, they have the highest reported incidence of pregnancy-related heart failure worldwide, and this is estimated to be about 1 in 100 live births, which is huge.

Unfortunately, the diagnosis of heart failure during this time period can be very challenging because of the overlap between normal pregnancy symptoms and heart failure symptoms. It’s not unusual for a pregnant woman to report low extremity swelling, which is swelling in the legs, shortness of breath with minimal activity or shortness of breath while laying in the bed at night. And this might be attributed to the fact that she has a gravid belly and some of the discomforts that come with pregnancy. This can cause the diagnosis of heart failure to be delayed or completely missed. When this happens, this is associated with adverse outcomes.

With that background in mind, we decided to evaluate this AI tool that was originally developed in 2019. It takes data from a noninvasive test, which is the electrocardiogram (or the ECG), and it can predict the likelihood of cardiac dysfunction or heart failure, but had not been previously tested in pregnant or postpartum women. Our thoughts were, can we use this tool to improve the recognition of heart failure in pregnant women where it’s hard for us to distinguish whether it’s just normal pregnancy symptoms or heart failure symptoms. And if we are successful, can we facilitate earlier treatment for this condition, which can be life threat threatening, but actually treatable?

The primary endpoint of SPEC-AI Nigeria was identification of cardiac dysfunction, which we define as a low ejection fraction on a cardiac ultrasound. And that number we look for is an ejection fraction <50%. Using this AI tool, with a digital stethoscope being used to capture the ECG as well as the heart sounds, we found out that screening supported by AI actually doubles our detection of pregnancy-related heart failure when we compare this here.

Some of the secondary endpoints from the SPEC-AI Nigeria includes the performance of this AI tool. I mentioned that we use a digital stethoscope, which is a portable technology that can be used across different clinical settings, whether in the hospital or in the clinic setting or even potentially in the community. We also evaluated a standard clinical ECG analyzed also with AI. What we found out was using both tools, we demonstrated in the SPEC-AI Nigeria study that they would have excellent performance for detection of cardiomyopathy or pregnancy-related heart failure. The metric that we use to determine how good a diagnostic tool is something called the AUC or the area under the curve, and this was >0.9 for both tools.

The additional thing that we did was we evaluated the performance of this tool at different ejection fraction thresholds. When we measure cardiac function using a cardiac ultrasound, they are different thresholds depending on the pump function of the heart. So <50% is kind of this general broad group of dysfunction, but more severe dysfunction can happen. Usually in pregnancy we look at the threshold of <45%.

In the heart failure space, we look at <40%. And when we start to consider patients for advanced heart failure, therapies like, implantable defibrillator, we look at <35%. Across all of these thresholds, we also saw that these AI tools performed really well and had identified a numerically higher number of cases using the AI tool when we compare it to usual care where they just go and see their obstetrician.

We also perform subgroup analysis, which is included in the secondary endpoints, which is specific age groups (if they are younger than 30 or older than 30) depending on what region of the country they are located in, and whether or not they have hypertensive disorders. We found that the performance was similar across all of the subgroups that were identified.

One key thing I should mention is that looking at the result, we only found statistically significant results for the digital stethoscope when we compared it to the clinical ECG. And this really was due to differences in the sensitivity of the tool in that patient population.

I think the results from this study are very strong. They’re very compelling. We already have convincing evidence that this tool is effective and can potentially influence maternal outcomes. So now one of the things we are looking at is how we get this technology into the hands of providers.

I think that a very important next step will be implementation studies. We really need to understand what are going to be the potential barriers to adoption. Are there going to be issues when we decide to roll this technology out on a large scale? What are the things that we need to consider: context-specific related issues, particularly in a low resource setting and low-to-middle income country like Nigeria? What are the implications for power, obtaining an echocardiogram and connecting these patients to cardiovascular care? So, of course, our next step is to conduct implementation studies to evaluate these questions, and I believe the information we gather from that will guide how best to roll out the technology.