Artificial intelligence system predicts who might die during future pandemics

Disease and treatment 15. mar 2021 4 min Associate Professor Martin Hylleholt Sillesen Written by Morten Busch

The COVID-19 pandemic led to great confusion about who has higher risk of adverse outcomes. This caused considerable uncertainty among the general public and posed many questions for an already stressed healthcare system. Researchers have now developed an artificial intelligence system that can identify with 90% accuracy the people who will die if they develop COVID-19 and almost as accurately identify who will need a ventilator. The system can help to determine priorities for who should be protected and vaccinated during future pandemics.

Imagine typing some of your data into a computer. How old are you? Male or female? What is your body mass index (BMI)? Press enter and then the computer tells you your risk of severe adverse effects if you develop COVID-19. And what about your father’s risk? And what about your friend who has multiple sclerosis? A new artificial intelligence system can provide part of the answer. Entering health data from almost 4,000 people with COVID-19 in Denmark has enabled the system to provide answers to both individuals and the healthcare system, enabling appropriate action.

“The system can predict with 82% accuracy whether a person with COVID-19 will be hospitalized and can tell healthcare professionals with 72% accuracy which inpatients will need intensive care. Finally, the system can determine with 90% accuracy whether a person will die from COVID-19. This means that we will be able to precisely identify the people at higher risk during pandemics and thus create greater certainty and enable the healthcare system to focus on who should be given priority for treatment and who should be vaccinated first,” explains Martin Hylleholt Sillesen, Associate Professor of Surgery, Center for Surgical Translational and Artificial Intelligence Research (CSTAR), Department of Surgical Gastroenterology, Center for Cancer and Organ Diseases, Rigshospitalet, Copenhagen.

A few data will do

The project began in a completely different direction as a system to predict who would have the most severe complications after surgery, but just as the researchers from the University of Copenhagen, Rigshospitalet and Bispebjerg & Frederiksberg Hospital got ready to test the system, the COVID-19 pandemic struck. They therefore decided to change tack.

“We started working on the models to assist the hospitals, since they were especially concerned during the first wave about lacking ventilators in intensive care. Predicting how many people and especially who are at higher risk could help in setting priorities,” says Martin Hylleholt Sillesen.

Initially, the researchers did not know which factors determine whether a person is at higher risk. They therefore began to feed the system various data on previous medical history and health data plus the outcomes after COVID-19, and the system got better at determining the outcome for each individual. In May 2020, the system began to be able to predict the trajectories of disease severity of people with COVID-19.

“As the system got better and better at prediction, we increasingly realized that a few factors such as BMI, blood pressure, age and sex are sufficient to quite accurately predict the trajectory. Adding knowledge about nervous system disorders, chronic obstructive pulmonary disease, asthma, diabetes and heart disease even more finely tunes the model. As we get data for more and more people, the system will approach 100% accuracy, although we probably will never achieve this,” explains Martin Hylleholt Sillesen.

People’s health data is not the only thing that determines outcome: treatment does too. For example, the hospitals were so overwhelmed for a while in Italy and other countries that this affected when people could be hospitalized and placed on the limited number of ventilators and especially how many patients survived.

“This artificial intelligence model cannot just be transferred between countries but must be trained in the specific country and its healthcare system,” says Martin Hylleholt Sillesen.

Cannot replace doctors

Today, the artificial intelligence system contains health data from 32,536 people with COVID-19 in Denmark, which enables the model to accurately recognize patterns and connections to both an individual’s previous illness history and the trajectory of COVID-19. Age and BMI typically determine how severely COVID-19 affects a person.

“The probability of dying or needing a ventilator is also higher if you are a man, have high blood pressure or have a nervous system disorder. We could not implement the system in a hospital because it cannot directly access current health data and because ethical issues need to be clarified. So for now, the system is primarily being used to support the difficult decisions doctors may have to make about treatment for individuals,” explains an initiator of the project, Mads Nielsen, professor and head, Department of Computer Science, University of Copenhagen.

The researchers do not think that computers will ever be able to replace doctors’ assessment, but they can help doctors and hospitals in monitoring many people with COVID-19 simultaneously and help to decide ongoing priorities. By using a model that can predict with 80% accuracy, artificial intelligence may soon help Denmark’s hospitals in predicting the need for ventilators on an ongoing basis.

“We are working towards being able to predict the need for ventilators five days in advance by giving the computer access to health data on everyone testing positive for COVID-19 in the Capital Region of Denmark,” says Mads Nielsen.

The researchers think that the system may be even more useful in relation to vaccination, predicting with 90% accuracy whether an uninfected person will die from COVID-19 if they become infected.

“The system can be used to fine tune the process of identifying people who need to be vaccinated. It may be sensible to give people priority if they fit one or more of the criteria,” adds Mads Nielsen.

Future pandemics

Although the new artificial intelligence system would have been even more useful a year ago when the pandemic broke out and will potentially continue to be useful during this pandemic, researchers see huge future potential. One major challenge during the first wave was uncertainty about the risk and the great pressure this placed on an already strained healthcare system.

“For example, there were reports that people with chronic diseases had higher risk, but this does not apply to people with chronic neck pain and many other chronic conditions, but many people called their doctor and put huge pressure on the healthcare system. In the next pandemic, we can analyse the new disease from day 1 and identify the most important risk factors. Maybe you could even be notified electronically about your risk and what to do,” explains Martin Hylleholt Sillesen.

This can save considerable resources and weeks if not months of valuable time and thus be more prepared for future pandemics. However, technical work is still needed to make health data from Denmark’s administrative regions available to the computer so that it can calculate the risk for people who have developed COVID-19. This will enable the system to process current data rather than data that is several weeks old.

“Today, only healthcare professionals have this access, not researchers. If this challenge can be overcome, we can create a system that easily and quickly screens healthy people for risk and inpatients for the treatment they need and discovers who needs to be vaccinated first. COVID-19 has shown that people vary tremendously in how ill they become when they develop COVID-19. And no one knows for sure how they will be affected. With artificial intelligence, we can provide the answer – or at least part of it,” concludes Martin Hylleholt Sillesen.

Developing and validating COVID-19 adverse outcome risk prediction models from a bi-national European cohort of 5594 patients” has been published in Scientific Reports. In 2019, the Novo Nordisk Foundation awarded a Research Leader Programme grant to Martin Hylleholt Sillesen for the project Towards Precision Medicine in Surgery – an Integrative Approach Combining Deep Learning, Genetics and Epigenetics and grants in 2020 for the project Advanced Surgical Risk Prediction through Artificial Intelligence and for the project Applied Artificial Intelligence for Real-time Risk Assessment of Patients with COVID-19.

© All rights reserved, Sciencenews 2020