LONDON – European scale-up of an artificial intelligence tool for stratifying and personalizing treatment of COVID-19 patients according to the type of complications they are likely to experience will get underway in September, following initial validation.
The tool, developed by researchers at the Hospital Clinic Barcelona, Spain, was ‘trained’ on more than a trillion anonymized data points retrieved from the clinic’s electronic health records system.
The real-time analysis stratifies patients with confirmed a diagnosis of COVID-19 who are hospitalized with respiratory symptoms, according to if they are likely to suffer inflammatory, co-infection and/or thrombotic complications.
The results of an initial study, published in the journal Clinical Infectious Diseases, show that using the algorithm, it was possible to predict the trajectory of the disease in individual patients, allowing for timely and appropriate treatment.
Personalizing treatment according to this stratification led to a 50% reduction in mortality. “We took data from patients’ electronic health records and were able to define different patterns that show a patient will have different complications,” said Carolina Garcia-Vidal of the department of infectious diseases at Hospital Clinic Barcelona, lead author of the paper.
“From that we found which patients were going to have the worst outcomes and how to differentiate patterns of response,” Garcia-Vidal told BioWorld.
Interestingly, predictions made by the AI tool were not confounded by comorbidities, such as diabetes and cardiovascular disease, that have been seen to predispose people to more severe COVID-19 infections. “One of the most important things was that we could improve outcomes in all kinds of patients, independent of comorbidities,” Garcia-Vidal said.
“The AI system that we have built is capable of supporting clinicians in the early diagnosis of patients more prone to develop complications,” Garcia-Vidal said. “Thus we have been able to provide timely and personalized treatments.”
The researchers stress the aim is not to supplant clinical judgement, but to provide an objective tool to aid clinical decision-making.
Development of the AI tool was funded by the EU’s European Institute of Innovation and Technology’s health arm, EIT Health, following an emergency call for research proposals to help deal with the pandemic.
From a standing start in April, the COVID-19 digital control center project has been developed and validated, showing early success in the stratification and personalization of treatment, for a disease that was unknown eight months ago.
“I am very proud of the early results demonstrated by our Digital Control Center for the COVID-19 project, which has been rapidly implemented and is already showing its potential to save lives,” said Jan-Phillip Beck, CEO of EIT Health. “We look forward to further validation and will work to make it available for as many patients as possible.”
The Spanish researchers note that the ability to predict in advance what complications each patient was likely to suffer was particularly important at the height of the pandemic, when clinicians from every speciality were on the front line treating severely affected patients.
Although the project began as a response to the emergency, Garcia-Vidal and colleagues were already working on applying AI to the analysis of data held in electronic health records. As one example, they were interested in real time predictions of the likely progression of multidrug resistant infections, and linking this to clinical decision-making.
When COVID-19 patients started to be admitted, the researchers hypothesized that patients would have distinct analytics patterns that would reflect the three main categories of complications being seen in hospitals around the world.
The project began with an observational study of all patients with a confirmed diagnosis of COVID-19 admitted to the hospital between March 28 and April 1.
Data on patient condition, blood tests and microbiology results were retrieved from the clinic’s electronic health records, as the basis for identifying the defining characteristics of the three dominant clinical patterns of complications.
For example, the inflammatory pattern in patients presumed to have an excessive cytokine response, was defined by C-reactive protein, ferritin, procalcitonin and creatinine levels.
The personalized therapy approaches administered as result were inhibitors of interleukin 1 and/or interleukin 6, in addition to the standard of care.
Comparing the differences between 99 patients who received personalized therapy and 147 who did not, showed there was an improvement at day five in 93.3% in the stratified group, compared to 59.9% on standard of care (P<0.001). At day five, 2% of personalized therapy patients had died vs. 17.7% on standard of care. Twenty eight day mortality was 20% vs. 44.2% (p=0.004).
The researchers acknowledged the numbers were small. However, a requirement of EIT funding is that the results can be replicated. Other hospitals that are members of the EIT network, two in Spain and one each in Netherlands and Belgium, will take part in the initial scale-up.
In addition, Garcia-Vidal said she has had approaches from hospitals elsewhere in Europe that are interested in adopting the AI tool.
“At EIT health we put a lot of focus on scalability,” said Jorge Fernandez-Garcia, director of innovation at EIT Health. “Not only does this benefit patients and citizens in equitable access to care, but it is also [fundamental] to the viability of products and services being developed within the network,” he told BioWorld.
EIT has helped manage the implementation of the validated algorithm developed at Hospital Clinic Barcelona, ensuring clinical decision support systems will be able to retrieve the required data from electronic health records at each hospital.