What happens when artificial intelligence is embedded into a healthcare device such as an X-ray scanner?

A first-of-its-kind example is Critical Care Suite^[1] on Optima^TM XR240amx from GE Healthcare. The AI application automatically reviews images to detect pneumothorax and, if suspected, flags the exam for prioritized review by the radiologist. After the images have been sent for review, Critical Care Suite then brings awareness to the technologist, on the device user interface^[2], of those cases that were flagged for prioritized review by the radiologist.

The growth of AI-enabled devices seems unstoppable. As researcher IDC noted in a 2018 white paper: “Training and inferencing require a tremendous amount of compute today, but as frameworks and libraries become more established, training will be reused broadly, and inferencing will move closer to the endpoints where the data is being created to determine an outcome in microseconds (units of one-millionth of a second) or in real time.”^[3]

“When an X-ray is taken on a patient, especially a patient who’s suffering from an emergent condition or a potentially life-threatening condition, the time that it takes to process, have someone read that, and have the image actually come into a queue is a really important period where minutes and hours matter,” said Rachael Callcut, MD, associate professor of surgery at the University of California, San Francisco Medical Center and director of data science for the Center for Digital Health Innovation.^[4]

Critical Care Suite was developed with the goal of improving timely patient care through a collaboration between GE Healthcare and the UCSF Center for Digital Health Innovation. Callcut partnered with John Mongan, MD, and Andrew Taylor, MD, radiologists at UCSF, to create the initial use case and data science approach behind the pneumothorax (PTX) detection algorithm.

Critical Care Suite has a positive predictive value of 35%, and a negative predictive value of 99%, assuming a department with a PTX prevalence of 4%. Specifically, it has one true notification for two false notifications on average, assuming a department with a PTX prevalence of 4%.

The PTX algorithm operating point was selected to optimize sensitivity for large pneumothoraxes, which are most likely to need treatment, and overall specificity, which will reduce false notifications.

Embedded intelligence at the point of care is essential for widespread adoption of AI within healthcare, and in this case it can help prioritize radiologist review of critical cases, which is intended to help shorten time to treatment.

Improving data capture

Another powerful outcome of embedding AI features into the device is how it will improve data capture before this data feeds up departmental or hospital-wide EHR systems.

About a year ago, an influential healthcare data scientist at Georgia Tech asked if there was a way of improving the capture of data at the original source of the data. As Travis Frosch, global data strategy leader at GE Healthcare, remembers the meeting he said, “It’s going from a device, to PACS (picture archiving and communication system) software, to a radiology workflow, to the EMR, and we’re trying to normalize data out of the EMR, which is a nightmare.”

Indeed, GE Healthcare had already been working on conducting data inferencing directly on devices, to speed up clinical decision-making. “But I never thought about it from the data side,” said Frosch. Adding intelligence at the edge will also address thorny issues around data governance, he said.

Embedding AI into a device reflects a paradigm of “seamless guidance and integration,” according to Matt Wickesberg, senior product manager of the Edison AI Workbench, a part of the Edison platform, GE Healthcare’s intelligence offering. Unveiled last November, Edison, named after GE’s co-founder, Thomas Edison, comprises intelligent applications and smart devices built using the Edison platform, and offers seamless AI services on device, edge, and cloud.^[5]