The coronavirus pandemic boosted the use of artificial intelligence in clinical practice. Faced with medical staff shortages and overflowing patient volumes, more and more hospitals, such as those in the U.S. and U.K., are turning to automated tools, especially in the emergency department. But as significant as these tools are, their rapid adoption comes with risks.
The Corona pandemic has crushed health systems around the world. In a survey by the World Health Organization (WHO), a large percentage (90 percent of 105 countries) spoke of a disruption of basic health services. Particularly in the emergency department, difficult decisions had to be made regarding triage, allocation and reallocation of medical resources. Vital to this: early identification of patients at risk of decompensation and requiring mechanical ventilation
At the same time, the pandemic drove the development of artificial intelligence models - from testing systems to diagnostics to drug approvals related to Covid-19. Rarely did these applications receive marketing approval, a sign of the novelty of this application - but also of the opacity of regulatory processes.
Algorithm beat MEWS system
However, in an unusual move, against the backdrop of the glaring impact of the Corona crisis on the U.S. healthcare system, the U.S. Food and Drug Administration (FDA) has granted several emergency approvals for AI models. Among them is a machine learning algorithm called the COViage Hemodynamic Instability and Respiratory Decompensation Prediction System from Dascena. It predicts the likelihood that hospitalized COVID-19 patients will require intubation.
In spring, Dascena tested the AI tool in a clinical trial of 197 hospitalized COVID-19 patients at five U.S. hospitals. The study in the journal Science found that the algorithm performed more effectively than the Modified Early Warning Score (MEWS), a rule-based system commonly used by clinicians to assess the needs of COVID-19 patients. The algorithm accurately predicted ventilation needs within 24 hours and managed to identify 16 percent more patients than MEWS while simultaneously minimizing false positives.
Meanwhile, U.S. researchers have taken their first look at some U.S. FDA-approved products that attempt to assess neurological, pulmonary and musculoskeletal trauma indications. In a study in the journal Emergency Radiology (2020), the authors conclude that "the ability to triage patients and attend to acute processes such as intracranial hemorrhage, pneumothorax, and pulmonary embolism can largely benefit the health care system, improve patient care, and reduce costs."
Radiology as a driver
Due to bottlenecks and delays in PCR testing, chest radiographs have also become one of the fastest and most affordable ways for physicians to triage patients. Studies had shown that severe cases of covid-19 on radiological images had distinct lung abnormalities associated with viral pneumonia. Here, in particular, the opportunity for digital applications made its debut. Research institutions began collecting medical images. Researchers at university hospitals across the United States ( among them: Stanford, Ohio State, Pennsylvania, and Emory) collected CT scans, chest X-rays, and ultrasound images of the lungs to create prognostic AI models to aid in the treatment of Covid-19.
Artificial intelligence is seen as having a lot of potential, especially in radiology. Deep learning, in particular, has shown results in analyzing medical images to identify diseases such as breast cancer, lung cancer, and glaucoma at least as accurately as human specialists. Researcher David Major of the Center for Virtual Reality and Visualization (VRVis) recently spoke to "medinlive" about the opportunities and weaknesses of artificial systems in radiology.
The first applications of artificial intelligence are also emerging in the treatment of COVID19 patients. A recent example is the ventilation of COVID19 patients. The Munich-based startup Ebenbuild creates a digital twin of the lungs in advance of therapy. Then, based on a large amount of data from other patients, an AI system suggests certain settings for the ventilator. These can be systematically tested on the digital twin lung. In ideal cases, the currently common, laborious and damaging trial and error process on the patient can be completely eliminated and his or her lungs can be spared.
Opportunities versus many weaknesses
The potential of artificial intelligence in healthcare is tremendous, alone when you considering the amount of data available. Experts estimate that almost a third (around 30 percent) of the data saved worldwide relates to the healthcare sector. And new medical data is being created every second. When you include medical claims, clinical trials, prescriptions, academic research and more, the output is on the scale of 750 quadrillion bytes per day. The volume of data has grown exponentially, overwhelming physicians - from surgery to radiology. As the latest information is neither available in a timely manner nor systematically processed in the course of medical interventions, the healthcare system is working less efficiently than it would actually be possible due to information technologies, experts criticize.
However, the first applications of artificial intelligence are also developing in the treatment of COVID19 patients. A current example is the ventilation of COVID19 patients. The Munich-based startup Ebenbuild creates a digital twin of the lungs in advance of the therapy. An AI then suggests certain settings for the ventilator - based on a large amount of data from other patients. These can be systematically tested on the digital twin lung. In the best case, the currently common laborious and damaging trial and error process on the patient is completely eliminated and his lungs can be spared.
Opportunities still face many weaknesses
The potential of artificial intelligence in healthcare is enormous, just looking at the amount of data. Experts estimate that almost a third (eand 30 percent) of the data stored worldwide relates to healthcare. New data is being created every second. When you include medical claims, clinical trials, prescriptions, academic research and more, the yield is on the order of 750 quadrillion bytes per day. The volume of data has grown exponentially, overwhelming physicians - from surgery to radiology. As the latest information is neither available in time nor systematically processed in the course of medical interventions, the healthcare system is working less efficiently than would actually be possible due to information technologies, experts criticize.
Yet the use of artificial intelligence is linked to vulnerabilities. Issues such as the black box problem, discrimination and data protection are completely unresolved. Researchers find it difficult to understand how an AI model reaches a result because of the complex systems it uses. Very often, AI models are trained with a small data set, which always contains a bias. In face recognition, for example, there is also the problem that there is too much data of faces with light skin color, which leads to bias and discrimination. In Austria, the introduction of the General Data Protection Regulation has made it even more difficult to obtain medical data.
Another unresolved issue is its use in everyday clinical practice. "Medical universities and training programs do not teach medical users how to deal with AI. You need some understanding of how the algorithms work and how the decisions are made, and you need to be prepared to be critical," says Daniel Cabrera, associate professor of emergency medicine at the Mayo Clinic, which uses an AI algorithm in a clinical setting. Moreover, even digital applications make mistakes. "For a certain percentage of patients, we're going to get the wrong recommendations," Cabrera says. Still, according to a recent Johns Hopkins study, more than 250,000 people die each year in the United States due to medical errors, making it the third leading cause of death after heart disease and cancer.