Blood tests ordered in intensive care units (ICU) have pros and cons: While they provide vital information, they also come with risks and additional costs to patients. At a recent symposium, a team of researchers from Princeton University unveiled a computational algorithm that assigns rewards and penalties to a specific test, with a goal to tamp down on lab tests and improve timing of therapies.
Sifting through records of 58,000 critical care admissions at Beth Israel Deaconess Medical Center in Boston, Barbara Engelhardt, PhD, an associate professor of computer science at Princeton, and graduate students Li-Fang Cheng and Niranjani Prasad zeroed in on more than 6,000 records of adult patients admitted to the ICU between 2001 and 2012 and who met certain criteria on ICU length of stay, vital signs, and common lab tests. The investigators looked at four tests commonly ordered in the ICU to diagnose kidney failure or sepsis: creatinine, blood urea nitrogen, lactate, and white blood cells (WBC).
“These medical data, at the scale we’re talking about, basically became available in the last year or two in a way that we can analyze them with machine learning methods,” said Engelhardt, the senior author of the study, in a statement. “That’s super exciting and a great opportunity.”
The computational algorithm she and her colleagues devised incorporates a “reward function” that supports a test order based on how informative the test is, while adding a penalty that factors in patient risk and the cost of the test. Clinicians can weigh both these elements to determine whether or not to order a test and if that test might lead to a certain intervention such as antibiotics. The key in this is to approach medical testing “like the sequential decision-making problem it is, where you account for all decisions and all the states you’ve seen in the past time period and decide what you should do at a current time to maximize long-term rewards for the patient,” said Prasad.
The investigators compared the machine learning algorithm’s reward function values against the testing regimens actually used with the 6,000-plus patients, in addition to those values resulting from randomized lab testing policies. In both scenarios, their system outperformed actual policies used in the hospital as well as random policies, yielding more information than the actual testing regimens clinicians used. While the results weren’t as promising for lactate tests, use of the algorithm could have reduced lab test orders by about 44% for WBC tests—and led to earlier interventions by clinicians to assist patients.
Next steps are to collaborate with data scientists on Penn Medicine’s Predictive Healthcare Team to bring this approach to clinical practice in the next few years. “This is one of the first times we’ll be able to take this machine learning approach and actually put it in the ICU, or in an inpatient hospital setting, and advise caregivers in a way that patients aren’t going to be at risk,” said Engelhardt. “That’s really something novel.”