Data-driven machine learning effectively flags risk for post-stroke dangers

Comparing 89 patient cases retrospectively, the team found that automated features of existing ICU data were as effective as the transcranial doppler procedure currently used to detect a dangerous constriction of blood vessels in the brain. Transcranial doppler tests require a skilled technician to be available and are often only conducted once a day, and while the test is selective and accurately detects people who are risk, it is not as efficient (sensitivity of 56%) at ruling out which patients are not at greater risk of this serious adverse event .

"There is a great opportunity to utilize abundant existing data to provide guidance and clinical decision support, as this model was as effective and much less resource-intensive," said senior author Soojin Park, MD, assistant professor of Neurology at Penn. "However, while this simple method may be valuable, most ICUs don't have the IT infrastructure to synergize data in such a way."

The team plans to look at prospective cases to compare this method directly with other assessments and clinical decisions.