University of Florida presents at U of T AI conference
March 31, 2022
Various hospitals in Canada and around the world have deployed Early Warning Systems – solutions that track the health of critical care patients and seriously ill patient, and sound an alert when patients are in danger of crashing. It’s a way of identifying and treating patients before it’s too late.
However, most of the systems in use only collect data every few hours, while the health of a patient can change in minutes.
That’s why an AI research unit at the University of Florida has devised a system called DeepSOFA that collects data continuously. It also analyzes the data using artificial intelligence and machine learning, to better predict which patients are in danger.
Using the data, the team has not only been able to flag patients in immediate danger, but the AI-based system can also predict outcomes for patients in the future.
“We can predict 30-day, 60-day, 90-day and even a year later mortality,” said Dr. Parisa Rashidi, an associate professor of biomedical engineering at the University of Florida, in Gainsville, and director of the school’s Intelligent Health Lab.
Dr. Rashidi was a speaker at the University of Toronto AI Conference, an online event held in February. She said one of the most popular Early Warning Systems in use today is SOFA – short for Sequential Organ Failure Assessment.
However, it typically makes assessments of a patient’s health by gathering data about six different systems – respiration, coagulation of platelets, liver health, cardiovascular hypotension, central nervous system data, and renal creatinine levels – every few hours.
Dr. Rashidi’s group improved on the predictive abilities of SOFA through continuous monitoring and by building an AI database of over 36,000 ICU encounters. Moreover, it compared its data with another database of 48,000 encounters.
With the AI model and monitoring, the system is able to provide an opportunity for interventions.
On another front, the Intelligent Health Lab created an AI-based system that can predict possible problems for post-surgical patients, such as sepsis, delirium and ventilation issues.
The team has also created a solution that makes use of video monitoring to help spot problems being experienced by ICU patients, some of whom are intubated and can’t express themselves very readily.
The solution makes use of video-based continuous monitoring to pick up changes in the patients’ facial expressions, posture, and other variables, such as the noise levels in the ICU and whether or not the lights are too bright. Something that’s also being tracked are the number of visits that a patient gets – the stimulation can improve the health of patients.
“Noise in some ICUs can be as loud as street traffic,” said Dr. Rashidi, who noted that it can affect patient outcomes and contribute to the onset of delirium.
The system is designed to help ICU nurses, who are often overburdened and able to visit patients only once every four hours. But by using a continuous monitoring system that tracks visual cues, nurses can be notified when a patient is in trouble.
This project is still ramping up, with 190 patients enrolled. Dr. Rashidi said more will be enlisted.
She also described a project to assess patients for dementia using AI analysis of pictures drawn by the patients. In particular, patients are being asked to draw a clock face.
“It’s such a simple test, but you’d be surprised at how much it can tell you about fine motor skills and memory,” she said.
The University of Florida team’s system is analyzing 23,000 clock images drawn by patients. It can make inferences based on the size of clock face and its shape (many patients with cognitive issues will draw the circular clock face as an oval or avocado shape), the numbers on the clock face, and the angle between the hands.
Dr. Rashidi said that obtaining large data sets is a problem in AI research, and that there is a challenge in getting data sets that are ethnically and regionally diverse. Only when such wide-ranging data sets are used can bias in the data be offset.
One of the barriers to creating large and diverse data pools are patient privacy laws, which prevent the data from leaving the encounter site.
A solution, said Dr. Rashidi, is federated learning, in which the data remains at its original site but is connected to a central repository. Even in this case, strong privacy protection must be built into the system.
Dr. Rashidi noted that diagnostic AI systems have made a lot of progress. However, the next step will be especially useful to clinicians. “At some point, we must move to models that recommend actions,” she said.
While systems can be built to make autonomous diagnoses and to suggest therapies, Dr. Rashidi said clinicians will always have the final word. “The critical decisions will still be made by the doctors.”