From interpreting CT scans to diagnosing eye disease, artificial intelligence is taking on medical tasks once reserved for only highly trained medical specialists — and in many cases outperforming its human counterparts.
Now AI is starting to show up in intensive care units, where hospitals treat their sickest patients. Doctors who have used the new systems say AI may be better at responding to the vast trove of medical data collected from ICU patients — and may help save patients who are teetering between life and death.
“Critical care is essentially this interface between humans and technology,” says Peter Laussen, chief of critical care medicine at Toronto’s Hospital for Sick Children. “The amount of data streaming from the patient in the ICU is huge,” encompassing readings of blood pressure, heartbeat, oxygen levels and other vital signs.
“We’re still at that very early phase of being able to use and implement it at the bedside,” Laussen, who co-chairs his hospital’s AI steering committee, said of AI. But in recent years, a handful of pilot programs have shown positive results.
Pilot programs
From 2012 to 2014, researchers tested a “smart” electronic medical record system — sort of a precursor to true AI — across 15 ICUs in the U.S. and found that it radically transformed them. Patients’ risk of dying was cut by half; in some cases, the system accurately identified potentially deadly conditions that doctors missed.
In 2016, researchers at the University of San Francisco piloted a new system that uses AI to detect a deadly blood infection called sepsis. The death rate fell more than 12 percent, meaning patients whose treatment involved the system were 58 percent less likely to die in the ICU.
In addition to saving lives, the system seemed to speed patients’ recoveries. ICU patients monitored by the AI system were discharged from the hospital an average of three days earlier than those who were not.
Within the past two years, Duke University and Johns Hopkins have launched similar systems that use AI to detect sepsis.
For the most part, medical AI systems are designed to sit quietly in the background of the hospitals’ computer systems, diligently tracking vital sign monitors and then sending doctors a text message or other notification at the first sign of trouble.
Doctors who have used these systems say AI gives them a better sense of their patients’ conditions. It also helps diagnose and predict critical illnesses, essentially saving lives by buying time.
“We’re at a point of being able to predict the likelihood of a cardiac arrest in 70 percent of occasions, five minutes before the event occurs,” Laussen said, adding that Toronto’s Hospital for Sick Children should have an AI system for heart attack prediction up and running within the next two years.
With AI, he said, “you can get ahead of these events.”
Information overload
Christopher Barton, an emergency medicine doctor at the University of California-San Francisco Medical Center who has championed the hospital’s AI initiative, said that in addition to improving medical care for patients, AI systems make life in the ICU less chaotic for doctors and nurses.
Existing vital sign monitoring systems “use very simple rules that set off an alarm or an alert to the patient’s provider or clinicians,” Barton said. “And those simple rules do have a high rate of false positive — or a high rate of ‘alarming’ when only a few criteria are met.”
In a traditional ICU, nurses respond to an alarm every 90 seconds, two thirds of which turn out to be false alarms, meaning they don’t signal real danger. Some doctors and nurses have been known to tune out — and even turn off — alarms, a phenomenon called “alarm fatigue.” In 2011, the FDA warned that alarm-related problems contributed to more than 500 patient deaths from 2005 to 2008.
Unlike a conventional monitoring system, AI is often able to predict problems hours in advance; doctors and nurses get a calm, text-message warning rather than having to respond to an urgent alarm signaling that a patient is already in trouble.
Barton said he received positive feedback from floor nurses during the AI study at UCSF. “The charge nurse of the floor will report that the number of false alarms has gone way down,” he said. “It’s making their work faster and more efficient.”
Learning from experience
Most of the intelligent ICU systems that have been implemented so far rely on a form of artificial intelligence known as machine learning. Unlike a typical piece of computer software, which requires an explicit set of instructions, machine learning algorithms can make decisions with limited human involvement.
Within ICUs, these systems “learn” from the constant stream of data from vital monitors and electronic medical records. In so doing, AI assistants can ignore false red flags and focus on meaningful patterns that indicate real risk. The more data that is fed into these machines, the more accurate their predictions.
As a bonus, AI models can hunt for meaningful patterns among massive databases of electronic medical records — absorbing far more data than a human would ever be able to review.
“I would love to have a 100-year work experience to detect some of these rare cases, but I simply don’t,” says Alexander Meyer, a computer scientist and cardiovascular surgeon at the German Heart Center Berlin in Germany. “If a machine has this experience and can decode this [data], it’s extremely helpful.”
Want more stories about technology?
SIGN UP FOR THE MACH NEWSLETTER AND FOLLOW NBC NEWS MACH ON TWITTER, FACEBOOK, AND INSTAGRAM.