After years of pontificating over the advent of big data and what it could mean for healthcare, the medical industry has finally found itself on the verge of a new and smarter era.

But what does artificial intelligence actually mean? How will these sophisticated algorithms be regulated and where are the results going to first show themselves? In some cases, these questions are as complex as the technology itself – but as companies continue to invest in these new capabilities the answers are beginning to reveal themselves.

GE Healthcare drew inspiration from NASA by creating its own Mission Control Center. But instead of managing space flights, it manages a patient’s hospital experience with the help of sophisticated algorithms and predictive analytics. The command center technology continuously analyzes real-time data across multiple sources to detect and prevent risk, manage staff workload and streamline the discharge process.

Licensed providers within the command center are tasked with monitoring and leveraging decision support applications. One of the applications, for example, monitors patients scheduled to be discharged and addresses problems that could result in preventable delays.

When the hospital or health system first launches their command center, they choose between five and 22 applications based on their specific needs. After that, they have the opportunity to add to their suite each year.

GE is currently working on future applications for perinatal quality, elderly care, patient safety, length of stay, computer vision, population health, genomics and home health.

“We worked problem-back with clients to understand real-world challenges facing caregivers in leading hospitals,” said Jeff Terry, CEO of healthcare command centers at GE. “AI-powered command centers evolved through years of work as a way to impact many of those challenges.”

To date, GE has implemented 10 command center programs – eight within the U.S., one in Canada and one in the UK. These 10 command centers support 41 hospitals in total.

AI across the whole health system
One of those command centers was installed at Oregon Health & Science University (OHSU) in July 2017. The health system’s main goal was to reduce capacity-based transfer declines and utilize their community partner hospitals more effectively.

In the year before the command center opened, OHSU declined more than 500 transfer patients from referring hospitals. But a year after it opened, the health system was able to accept 554 more transfer patients.

OHSU was also able to place 519 transfer patients into community partner hospitals, which freed up beds for patients who require the more complex, quaternary inpatient care that is only available at their academic health care center. To top it off, they also saw a seven-fold increase in return on investment in that first year.

“In order to accommodate various, growing demands on our system related to inpatient access and capacity management, we recognized the need to get reliable real-time data feeds and supporting alerts and to change how we run the daily operations in a more system mindset with a patient-centric approach,” Dr. James Heilman, chief medical transfer officer at OHSU, and Dr. Matthias Merkel, chief medical capacity officer at OHSU, wrote in an email response.

In September, GE announced plans to develop a software application for OHSU to better manage sepsis patient care. Sepsis is a major issue in the U.S. At least 1.7 million American adults develop sepsis and almost 270,000 of them die as a result of it, according to the Centers for Disease Control and Prevention.

The mission control staff will be alerted when the application detects a risk and care teams will be deployed as needed.

“Quick recognition and management of sepsis is critically important to increase the survivability and requires a system-level response,” wrote Heilman and Merkel. “Identifying patients and tracking their care progression along established sepsis care bundles seemed a logical expansion of the work we have done on daily operations.”

Another company that takes a health system-wide approach to AI is Palo Alto-based Ayasdi, which built an AI platform that their partners and customers can run applications on.

The company has developed a few applications of its own, including one called Clinical Variation Management (CVM), which analyzes hospital EMR and financial data and looks for clinical variation.

It automatically surfaces groups of similar procedures and generates clinical pathways that help to achieve the best patient outcome at the lowest cost.

“It discovers the sequences that are similar amongst patients to group them together,” said Gurjeet Singh, CEO and co-founder of Ayasdi. “Then having them grouped together, it predicts the best sequence per patient, and having those predictions in hand, it justifies these to a clinician.”

Four months after Flagler Hospital in St. Augustine, Florida implemented the CVM application, it was able to reduce the average cost of treating patients with pneumonia by 30 percent, reduced admissions by seven times, and decreased the average length of stay by about 2.5 days.

“For this one protocol, Flagler officials have discovered that by using our software, it will benefit them by roughly a million dollars per year,” said Singh. “They expect to have 18 protocols in production over the next 12 months, and expect to save about $20 million over the next three years.”

What constitutes artificial intelligence?
“AI” has become a very general term in recent years, and Singh would go as far as to say that it doesn’t mean anything specific anymore. He jokes with his friends that anyone with a Microsoft Excel spreadsheet can claim themselves to be AI these days.

According to him, an AI system should have five important characteristics in order to be considered legitimate. Those include the ability to discover, predict, justify, act and learn.

First off, an AI system needs to be able to learn from large, complex data sets without any human interaction. That is a phenomenon known as unsupervised learning and it’s so important because the system must discover all of the patterns that exist in the data without a human having to ask a question.

The AI system should also be able to use the large data sets to predict what is likely to happen in the future with a high degree of accuracy. Clinicians are still the main clinical decision makers, but AI can provide them with information on future needs, costs, disease burdens and patient risks.

The most critical characteristic of an AI system is its ability to justify or explain its results. That includes every recommendation, prediction and segment of the anomaly.

“If we are going to put our trust in these heavily automated augmented systems, the human operators of these systems have to be able to build intuition about what the system is doing and why it’s taking certain actions, and so on,” said Singh.

Many AI systems are worked on in the research setting, but not all translate into a product. If they are stuck in the academic setting, they are simply explorations of hypotheses instead of actual AI systems.

Lastly, the system has to be able to learn as the data evolves.

“For example, if the system spots that a new patient risk segment has emerged or a new type of payment fraud has begun to exist, then it should be able to update the human operator that something has changed, and [suggest] an action [they] might take in fixing that,” said Singh.

AI and radiology
There is no shortage of AI systems and applications for the radiology field. In the past year, more than a handful of new products came to market.

One of the most recent is Philips Healthcare’s Illumeo PACS with adaptive intelligence, which was a major focus at this year’s Radiological Society of North America (RSNA) annual meeting. The University of Utah Health recently became one of the first to leverage this technology.

The health system is tasked to review about 500,000 cases per year due to its large referral base. Before Illumeo was installed, the radiologists had to pull up a case in the PACS and search for the matching new and old images.

Illumeo leverages adaptive intelligence and analytics to automatically find those matching images. It then positions them side-by-side so that the radiologist can determine if the lesions got bigger over time.

“One of the biggest things is the amount of data,” said Dr. Richard Wiggins, a neuroradiologist at the health system. “We all know that we have to look at more and more data all the time for our studies and [some may be] getting reimbursed even less now than they were years ago for those studies.”

When he started his career, he was looking at between 2,000 and 2,500 images per day, but now he looks at about 250,000 per day. With the help of Illumeo, Wiggins and his colleagues are saving a lot of time.

MaxQ Artificial Intelligence also showed off its new AI technology at this year’s RSNA annual meeting. Its Accipio Ix application automatically flags non-contrast head CT scans with suspected intracranial hemorrhages so that those cases go to the top of the radiologist’s list.

MaxQ AI has formed partnerships with GE to incorporate Accipio Ix into its subscription program, and with IBM to add the application to its PACS. The company also has partnerships with Samsung NeuroLogica and EnvoyAl.

“Our goal as a company is to provide intelligent diagnostic decision support tools to the acute care setting,” said Gene Saragnese, chairman and CEO of MaxQ AI. “Many decisions that are made in the ER are time sensitive so we want to make accurate timely decisions. We think these types of tools can help in that type of situation.”

He added that patients are aggressively treated in academic institutions because the physicians have a high degree of expertise, but that is not always the case in community hospitals. MaxQ AI’s aim is to provide these AI-driven applications to physicians so they can have the confidence of an expert.

Accipio Ix received an accelerated pathway through the FDA’s Breakthrough Devices Program. Then, in early November, the company scored 510(k) clearance from the FDA.

Viz.ai is another promising company in the AI for radiology field. In April, its Viz CTP software, which automatically analyzes CT perfusion images, was cleared by the FDA. The technology’s advanced image analysis algorithm automatically generates parametric CT perfusion color maps based on the dynamic effect of the contrast agent through the brain.

The company also has another product called the LVO Stroke Platform that detects and alerts specialists to potential large vessel occlusion (LVO) strokes. It connects to a CT scanner and leverages an AI algorithm to detect the suspected strokes.

Aidoc, an Israel-based start-up is also making big moves in this field. Early last year, it launched the first AI-powered full-body solution for CT analysis.

The Aidoc Full Body Solution is an extension of the company’s head and spine AI solution. This new solution can help radiologists identify medical findings in the head, c-spine, chest, abdomen, etc.

Aidoc has big plans for the future. It is working on AI solutions for MR and is also looking into extending its AI solutions to other major imaging modalities.

When asked what the most exciting aspect of the untapped potential of AI is, Ariella Shoham, vice president of marketing for Aidoc, said, “If I had to say one thing it would be assessing the actual value and showing real influence on patient care. There are so many AI companies that are actually just algorithms that show levels of sensitivity and specificity but are nowhere near an actual implementation in a live setting.”

What AI is and what it ain’t
More and more, an industry consensus has emerged that AI will not replace radiologists anytime in the near future. So what does it mean for the practice itself?

The University of Utah Health System’s Higgins scoffed as he spoke about people who refer to themselves as “futurists”, who predict that there are going to be no more radiologists in the next few years. He believes that AI systems are going to be assisting radiologists and changing what they do, but that they stand no chance of replacing radiologists anytime soon.

“In reality, what I do as a radiologist is way more complicated than what people are doing with these algorithms right now,” he added.

He compared his job to figuring out what’s wrong in those pictures in the Highlights Magazines. Everyone is raving that AI algorithms can recognize a stop sign 70 percent of the time, but the job of a radiologist is much more complex than that.

“Being a radiologist is not really recognizing something like a stop sign, but more the fact that the stop sign is underwater and there are fish stopping at the stop sign,” said Higgins. “I think that’s closer to what we actually do as radiologists when we are interpreting images and trying to figure out not only if there is an abnormality there but what does it mean in terms of the clinical care and the patient’s clinical history.”