Hospitals have an understandable goal for their intensive care units: to reduce “dead in bed” events.
With streams of data coming from equipment that monitors patients’ vital signs, the ICU seems the perfect setting to deploy artificially intelligent tools that could judge when a patient is likely to take a turn for the worse. “A lot of hospitals are interested in developing early warning systems that can predict life-threatening events like sepsis, cardiac arrest, and respiratory arrest,” says Priyanka Shah of the ECRI Institute, a nonprofit that evaluates medical procedures, devices, and drugs for the health care industry.
Both academic researchers and medical device companies are now trying to figure out which combinations of measurements can provide the best indication of patient deterioration, Shah says. Once that technical challenge is met, researchers will still have to prove “clinical relevance,” she says—not just proof that the technology works, but also that it can be integrated into a hospital’s workflow and that it will save money.
Dealing with FDA regulators, set-in-their-ways clinicians, and money-conscious hospital administrators may be the more daunting part of the mission to smarten up the ICU. Because on the technical front, the research is promising.
Predicting life and death in the PICU
The scenes inside a pediatric intensive care unit (PICU) can be heart-wrenching. In the neonatal ward, fragile newborns lie inside plastic incubators surrounded by machines and monitors. Down the hall, kids connected to tubes and IVs smile bravely under bright cartoon murals.
At Children’s Hospital Los Angeles, data scientists Melissa Aczon and David Ledbetter have come up with an AI system that could give doctors a better—and earlier—sense of which kids are likely to get worse. Aczon and Ledbetter work within a hospital research unit called the virtual PICU, where they partner with clinicians who are eager to see improvements in operations. “Their perspective is: All these ICU encounters are taking place and generating data,” says Aczon. “We have a moral responsibility to learn from these encounters and apply the lessons learned to future patients.”
Aczon and Ledbetter wanted to start by training an AI system to make an unambiguous prediction, so they designed an experimental system to predict mortality in the PICU. Drawing from the hospital’s electronic health records, they used data about the kids’ vital signs (which are sometimes measured as often as every few minutes), lab results, and information about drugs administered and procedures carried out.
Using the records of more than 12,000 patients who had passed through the PICU, their machine learning program found patterns in the data that distinguished the 5 percent of patients who died. The program could then predict mortality with 93 percent accuracy, a significantly better score than that achieved by simpler rating systems currently used in hospital PICUs. Aczon and Ledbetter published their results on the preprint server Arxiv.
Their key advance was using a type of machine learning called a recurrent neural network that’s designed to process an ongoing flow of data, rather than drawing conclusions from a “snapshot” of data from one moment in time. “RNNs are an elegant way to deal with the sequential nature of clinical data,” says Aczon. “New information is always coming in, and you want to integrate that.” Their RNN performed better as it gathered data over time.
This system was experimental, but Aczon and Ledbetter say such a tool would be of great use in a PICU. Of course, if mortality-predicting software were really deployed in a hospital, doctors wouldn’t be satisfied with simply getting a grim reaper risk score. That assessment would only be a first step, says Ledbetter. “Once you understand what’s going to happen to the patient, then you can start thinking about ways to intervene and make something different happen,” he says.
Preventing crisis in the ICU
“Our mission is to decrease mortality by automating critical care in the ICU unit,” says Wassim Haddad, cofounder of AreteX Systems. The company (which is in the process of changing its name to AutoMedica) focuses on two essential parts of ICU care: managing patients’ breathing through a mechanical ventilator, and managing the amount of fluids they’re given through an IV drip.
In the United States, 5.7 million people are admitted to the ICU each year, says Haddad, and 2.3 million will require a mechanical ventilator to help them breathe. About 800,000 of those people will experience a problem called patient-ventilator dyssynchrony. “If they’re not properly sedated,they tend to fight the ventilator,” Haddad explains. “If they want to take an inbreath, but the machine is saying, no, exhale, it can cause extreme anxiety in the patient.”
AreteX’s engineers created a machine learning tool that can identify distinct types of dyssynchrony based on data from a patient’s ventilator. Their system sends an alarm to the nurse or respiratory therapist who can hustle over to increase the sedation and prevent the patient from fighting the life-supporting machine. The company has recently started a clinical trial at Northeast Georgia Medical Center to test the system.
And the current solution is just a first step in fulfilling the company’s mission. “We’d like to go beyond clinical decision support, which is what we offer now,” says Haddad. “Ultimately, a fully automated system could change the pacing of the ventilator to correct the problem.” Getting to that stage, however, will require extensive clinical trials to prove the system’s safety to regulators and wary hospital administrators.
AreteX has a similar program to monitor the fluids that most ICU patients receive through an IV drip to increase the volume of blood and raise blood pressure. Haddad says today’s fluid management process is inefficient: “Typically the physician gives the order to a nurse, who manually changes the infusion rate. After a couple of hours, the doctor reassesses the patient,” he says. The ICU staff has to strike a careful balance, he says, because patients can go into shock and suffer serious complications from either too little or too much fluids.
Haddad’s system employs machine learning to gage an individual patient’s ongoing response to the amount of fluid being administered, adapting over time to keep the patient’s condition steady. AreteX is testing the fluid management system at Northeast Georgia Medical Center as well.
With the U.S. population aging fast, Haddad says, and with a recognized shortage in the “intensivist” clinicians being trained to work in emergency rooms and ICUs, automation may be the only answer. “There are 3 million people over the age of 85 today, and there will be 9 million by 2030,” he says. “This is going to put tremendous stress on the nation’s ICUs.”
Eliza Strickland is a senior editor at IEEE Spectrum, where she covers AI, biomedical engineering, and other topics. She holds a master’s degree in journalism from Columbia University.