Artificial Intelligence to Enhance Detection of Sudden Cardiac Death

What were the key methodological challenges in developing an artificial intelligence model to accurately ascertain sudden cardiac death from heterogeneous electronic health record data sets?

That’s a very important question because the whole purpose of our project was to find individuals who died from this type of death, sudden cardiac death, in death records using electronic health record (EHR) data. The problem is that these individuals are under ascertained. We know that around half of all cardiac deaths are sudden cardiac death, but in death records there’s only a minority that are actually labeled as sudden cardiac death. So the biggest issue for us was trying to come up with a reliable outcome from which to train the model. We ended up using ICD-10 code I46.1, which is a very specific label, but we know misses most of the cases. And then for the controls, we did a number of filterings based on just common knowledge of sudden cardiac death to try to limit the number of sudden cardiac death cases leaked into the control label. So definitely, definition of the outcome was the key challenge. And from there, the project was pretty much a straightforward artificial intelligence application.

In what ways does an AI-based approach improve upon traditional methods of identifying sudden cardiac death in epidemiological and clinical research?

What’s currently being used to identify cases of sudden cardiac death in epidemiology are “cause of death” information, typically ICD-10 codes that define what did an individual die from. There is a label for sudden cardiac death in that system of ICD-10 codes, but only a minority of actual sudden cardiac death cases are labeled as such. And so there’s this whole consortium that did a great job in defining a way to identify cases of sudden cardiac death using those death records, mostly relying on types of cardiac death. There is a sub-portion of cohorts that might have “time frame of death” or autopsy information that allow for more granular identification of cases, but most cohorts are only going to be able to rely on those ICD-10 codes to define cause of death. What we are doing with AI is pretty much refining that definition that the consortium came up with: we take the definition of the consortium, which has around 50% precision and use this AI tool to ascertain which of those are actual sudden cardiac death cases. So that would be where AI can improve ascertainment compared to how it’s done now.

What are the main challenges clinicians and researchers face when trying to accurately capture sudden cardiac death using routine healthcare data?

So this links to the previous answer. They currently work mostly with “cause of death” information, so these IC10 codes, and ideally also with information on the time frame of death and the cardiac origin of death. And that’s because the definition of sudden cardiac death requires you to determine if that death was actually sudden – so there must be, or the gold standard would be, a bystander witness that can determine that death was sudden, and that information is not available in EHR data sets. I think there’s around 50% of cardiac arrests that are witnessed, and just a portion of that 50% that then ends up recorded in EHRs. So most EHRs are not going to have that information.

And then the other crucial piece of information that you need to define sudden cardiac death is the cardiac origin of death, for which the gold standard is an autopsy report. And again, the same problem – most EHRs don’t have that information. I think autopsies in general are performed in around 7-8% of cases of death in Europe and the US, something like that, I think those are the last numbers – just a portion of those 8% end up recorded in EHR. As I said, most EHRs only have these codes to define the cause of death. So those are the two main limitations that researchers face when trying to identify sudden cardiac death.

How can artificial intelligence help improve the reliability and scalability of identifying sudden cardiac death across large and diverse data sets?

More and more researchers are using these huge EHR databases, and as we said, the crucial pieces of information to define sudden cardiac death, which are a time frame and the cardiac origin of death, are typically missing. So artificial intelligence can harness all of that rich clinical data captured in the EHR to potentially shed some light into who died from sudden cardiac death, ideally, whether that death was sudden and whether the origin of that death was cardiac. And that’s where we focused a lot of our analysis, trying to determine if our model predicts the outcome that we define, which is sudden cardiac death, but also if the model can track with those defining attributes of sudden cardiac death, which are the time frame and the cardiac origin of death.

What potential clinical or research applications could arise from more accurate ascertainment of sudden cardiac death in the real-world data?

So that’s a huge problem. Sudden cardiac death is presumed to be the leading cause of death worldwide; it is estimated to cause around 50% of all cardiac deaths. But there’s a limitation to perform research on sudden cardiac death because there are not enough cases of sudden cardiac death in cohorts or now in EHRs. There’s just a handful of cohorts that manually define cases of sudden cardiac death by having this time frame and cardiac origin of death available, and they are limited around the world. Most of them are in this consortium that I mentioned earlier, and because obviously manual ascertainment is time consuming they are only able to ascertain a certain amount of cases which causes power limitations for research on those cases.

So the main goal of our project was to come up with an outcome that researchers can then use to study sudden cardiac death at large scale in EHR. So potentially this AI model, if it’s applied to large scale EHR database, it can enable research on a number of subjects. Some obvious applications would be development of risk prediction tools, but then also sudden cardiac death genetic analysis, which have been underpowered historically, molecular analysis such as proteomics or metabolomics. There’s a whole number of research applications that can derive from this outcome of sudden cardiac death.