Researchers use artificial intelligence to identify long COVID cases

Centers for Disease Control and Prevention 23 million Americans are affected by long COVID.. The term refers to a variety of conditions that last months or even years after infection, from brain fog and chronic fatigue to neurological problems and blood clots.

But the clear answer to how often this happens, who is most at risk, and why is still elusive.

Meghan McCarty Carino of the Marketplace speaks with Emily Pfaff, an assistant professor of medicine at the University of North Carolina at Chapel Hill. Analyze electronic health records using artificial intelligence, Better identify the syndrome and look for patterns that may treat the patient. According to Pfaff, some of the most common markers that the algorithm detects are fatigue, shortness of breath, and frequent doctor visits.

Below is an edited transcript of their conversation.

Emily Fuff: To get the best understanding of long COVID, the best thing in the toolbox is clinical research. However, without knowing who has a long COVID, it is not possible to conduct a clinical study of a patient with a long COVID. Therefore, we use machine learning models or artificial intelligence to obtain patients who are known to have long COVIDs, or who are convinced that they have long COVIDs, and to large-scale other patients who are similar to those patients. Find it in the dataset. Therefore, machine learning is very useful in that you can find patterns and match them with patterns of people you have never seen before. And that’s exactly how we use it here.

Meghan McCarty Carino: Tell us more about some of the characteristic symptoms identified by the algorithm.

Pfaff: Therefore, age is actually the most important feature of our model, and of course not a symptom. But it does show the fact that you actually affect the likelihood that age will increase your COVID. There is a non-linear relationship with whether the COVID is long. So it’s not easy to say that a 17-year-old is more likely than a 25-year-old. But certainly, age affects the output of the model. The main features used to match people who the model does not know with those who consider the COVID to be likely to be long are shortness of breath, fatigue, and so on. This is certainly a characteristic symptom you have experienced. Probably seen in many video reports on long COVID. That’s a big indicator. The frequency with which you go to the doctor is a very important feature. As a result, people who move between specialists and frequently move in and out of the office in a terrible mood are clearly more likely to be classified as long COVIDs. And new prescriptions such as asthma inhalers (albutamol) and other types of asthma-related medications are just as important as the gender of women.

McCarty Carino: What we often talk about in this show is that algorithms are as smart as trained data. You rely on these electronic health records. What bias is inherent in this data?

Pfaff: Thinking is so important that it spends a lot of time on what we think. What I feel most strongly about electronic medical records is the fact that the data we have is about people who could go to the doctor. Therefore, if you are not insured, cannot take time off from work, or are not raising children, they will not be included in my dataset. Another thing I’ve heard from some of the patient advocates in the long COVID community is that many long COVID patients are careful and frustrated. It’s actually a group of other people who aren’t represented in our dataset just because it’s very difficult to find a cure that works and makes people feel good. And that’s the group I’m really afraid to lose again.

McCarty Carino: Conversely, people who are more likely to continue to seek care may also share certain characteristics.

Pfaff: absolutely. And what we’re talking about when using electronic health records in all use cases, not just long COVIDs, is that sick people are always biased towards sick people because they use care more often. That is. And all this is necessary to stay deep in your mind when working with electronic health record data. I’m not saying that the data shouldn’t be used or that the data is useless. However, if you do not keep the warning in mind, you may draw the wrong conclusion.

McCarty Carino: So how do you think this tool is used in the real world of medicine?

Pfaff: My dream is to run such an AI model across a dataset of healthcare systems, which may be suitable for patients who may be suitable for therapeutic trials, potentially long COVID specialty care. To be able to identify the patient. Therefore, these types of models are not 100% likely to be correct, and probably not close to it. However, they narrow down a set of people who have a long COVID and are likely to be very rich for those who may benefit from receiving calls or recruiting calls for trial. ..

McCarty Carino: And what about other illnesses beyond the long COVID? Maybe other post-infection syndromes we know exist, but have not been studied to the extent that long COVIDs have, or even other chronic or rare illnesses.

Pfaff: absolutely. So I don’t think this kind of model and, by the way, the idea of ​​using machine learning to phenotype the disease of electronic data is not a new idea. And since long COVID is a new illness, I think it applies very well to something like the long COVID innovation here. So this kind of methodology works very well and, as you said, is attractive for rare illnesses that may not have good indicators of characteristic symptoms and new illnesses with a kind of diffuse illness. I think it’s a typical option. Symptoms appear like a long COVID. One lab test does not tell if a patient has X disease. What makes AI really good at compositing is like a set of features. Humans, not so much.

we Fully published article A summary from the National Institutes of Health that supported Fuff’s research by Emily Fuff and her partner.In fact, the electronic health record she used Part of the public dataset made available by NIH To support research that deepens the understanding of long-term COVID.

Electronic health records were also used in the CDC survey I mentioned, finding up to 23 million Americans. About 1 in 5 adult COVID survivors, Have experienced some symptoms of long COVID. It should be noted here that the study did not consider the vaccination status and was conducted during the first 18 months of the pandemic.

As Pfaff pointed out, there are some limitations to the analysis of electronic health records, but there are many opportunities.

Mayo Clinic Recently launched a startup incubator for healthcare AI companies.. They are given access to an anonymized database of the healthcare network of 10 million patients. The first cohort of companies is working to improve the care of patients with chronic illnesses such as diabetes and epilepsy, and to more accurately predict the needs of patients who send messages to healthcare providers through these systems.

I’m not going to lie — the spread of AI in healthcare can be a bit freaky. Recent news on Google’s probably perceptual chatbots.. So the HAL 9000 in 2001: A Space Odyssey had great bedside etiquette at first … but everyone knows how it ended.