Machine learning helps Geisinger cut time to diagnose intracranial hemorrhages

Register now

Geisinger Health System is leveraging machine learning to speed up the diagnosis of potentially fatal internal head bleeding by training computers to analyze computed tomography scans and flagging the most urgent images for review by radiologists.

The healthcare organization has reduced the time it takes to diagnose intracranial hemorrhages by 96 percent, and as a result the technology has been introduced into its regular clinical workflow. Time is of the essence when it comes to early detection of this life-threatening form of internal bleeding, which impacts about 50,000 U.S. patients annually and kills 47 percent of them within 30 days.

“This is not about replacing doctors with machines,” says Aalpen Patel, MD, chair of Geisinger System Radiology. “This is about the smart use of machine learning technology to aid medical providers in delivering better and faster care, especially in these areas where time is critical.”

Also See: Machine learning automatically identifies brain tumors

Patel and his colleagues conducted a study in which a convolutional neural network was trained on more than 37,000 radiological studies and then evaluated nearly 9,500 unseen studies. The machine learning algorithm was implemented prospectively for three months to re-prioritize “routine” head CT studies as “stat” on real-time radiology worklists if an intracranial hemorrhage was detected.

“An artificial intelligence algorithm can prioritize radiology worklists to reduce time to diagnosis of new outpatient intracranial hemorrhage by 96 percent and may also identify subtle intracranial hemorrhage overlooked by radiologists,” conclude the study’s authors. “This demonstrates the positive impact of advanced machine learning in radiology workflow optimization.”

According to Brandon Fornwalt, MD, associate professor and director of the Geisinger Department of Imaging Science and Innovation, this is the first study of its kind to implement a machine learning algorithm into clinical radiology workflow.

Patel recounts a successful intervention in which an 88-year-old woman who presented with symptoms thought to be related to her medication was rushed to the emergency department after the machine learning algorithm flagged her CT scan, reprioritizing her from routine to stat; she was found to be suffering from an intracranial hemorrhage.

“The quicker you can treat it, the better patients do,” contends Fornwalt, who adds that Geisinger is also applying machine learning to patients with congenital heart disease.

For reprint and licensing requests for this article, click here.