Pitt algorithm predicts recovery from brain injury
It’s difficult for doctors to predict how long a person will take to recover from a traumatic brain injury, if they do at all. While most people regain consciousness in a matter of days or weeks and recover quickly, some may remain in a coma.
That uncertainty can be agonizing for family members, who are faced with tough decisions as to when to withdraw life support from a loved one with a severe brain injury.
Researchers at the University of Pittsburgh and UPMC want to provide more certainty to ICU clinicians and families of brain-injured patients to help them make moreinformed decisions. They developed a machine learning algorithm that can accurately predict survival and recovery at six months after a traumatic brain injury by analyzing brain scans and certain patient data. They published their findings April 26 in the journal Radiology.
“Life-saving and life-altering decisions are made very rapidly in the first few hours after these severe traumatic brain injuries. We as clinicians do the best that we can, but we’re not always right,” said Dr. David Okonkwo, a professor of neurological surgery at Pitt and UPMC and senior author on the paper.
A traumatic brain injury is a sudden injury, such as a blow to the head, that affects how the brain works. These injuries represent a major cause of death and disability in the United States, and in 2019, they caused more than 223,000 U.S. hospitalizations. In 2020, about 176 people died from traumatic brain injuries each day, according to the Centers for Disease Control and Prevention.
Decisions about whether to withhold or withdraw treatment after severe brain injury often occur within 72 hours of admission to the hospital. A predictive algorithm could be used to screen patients shortly after they’re admitted to a hospital to help doctors decide on the best care.
“If we have a way to analyze the mountains of data that get collected on patients right away when they first come into the hospital
and identify the people who would go on to make great recoveries, then clinicians and families would be in a better position to make those very difficult decisions,” said Dr. Okonkwo.
To build their machinelearning algorithm, the researchers used brain scans and other data, such as age, sex and cause of injury, from 537 patients with an average age of 40 who were hospitalized at UPMC for traumatic brain injuries from November 2002 to December 2018.
They then tested the algorithm on a dataset of 107 UPMC patients and a second dataset of 220 patients from other medical centers. On both datasets, they found that their machine-learning algorithm was better at predicting patient outcomes at six months after injury than a group of three human neurosurgeons who were asked to evaluate the cases.
In the UPMC patient group, it was also more accurate than an existing calculator, known as IMPACT, used in clinical trials to predict outcomes from head injuries. However, it didn’t perform better than the IMPACT calculator on the dataset of external patients.
Dr. Christopher Zacko, director of neurocritical care and a neurotrauma surgeon at Penn State Health Milton S. Hershey Medical Center, who was not involved in the Pitt study, said the findings were encouraging.
“If we can find a way to narrow down the prognostic possibilities, I think that would be really, really helpful,” he said.
But to make the algorithm more powerful, he wants to see it trained on a bigger dataset.
Right now, the model includes brain imaging and other data only from when the patient is first admitted to the hospital. But a person’s health status can change over time, sometimes very quickly.
“Over the course of the patient getting care or recovering potentially, there are additional data being generated that we can incorporate,” said Shandong Wu, associate professor of radiology, bioengineering and biomedical informatics at Pitt, who collaborated with Dr. Okonkwo on the paper. “We don’t want to miss any patient who otherwise has a chance at survival.”
The team said their algorithm could be even more predictive if it included data from at least the first 72 hours after a patient is admitted to the hospital. They’re hoping to work with the Department of Defense to further develop the algorithm so that it could be used on soldiers with battlefield head injuries.
“We don’t want to miss any patient who otherwise has a chance at survival.”
—Shandong Wu University of Pittsburgh
They don’t see any algorithm ever replacing a physician’s judgment, but they saidartificial intelligence can help spot patterns that humans might miss. Arthur Caplan, a bioethicist at New York University, said he thinks it’s ethical to use an algorithm like this to help guide patient care, but at the end of the day, it’s still the physician who is responsible formaking patient decisions.
“We’re not giving authority over to machines. We’re trying to use a huge amount of information that probably no single doctor could retain or have available and make that available in a computergenerated formula,” he said. “No one’s saying you have to do what the formula says.”