Machine Learning Can Reduce Laboratory Testing for ICU Patients
Princeton researchers have found that machine learning can be used to improve care and reduce laboratory testing for patients in the intensive care unit.
The researchers used data from more than 6,000 ICU patients to design machine-based learning systems that could reduce the frequency of laboratory tests and increase critical care time.
The researchers used the MIMIC III critical care database, which included a record 58,000 critical care admissions at Beth Israel Deaconess Medical Center. For this study, they selected a subset of 6,060 records of adults who stayed in the ICU for up to 20 days and had measurements for general vital signs and laboratory tests.
“This medical data, on the scale we’re talking about, basically becomes available in the last one or two years in a way that we can analyze with machine learning methods,” said Princeton Computer Associate Professor Barbara Engelhardt, who is the study’s senior author.
The analysis focused on four blood tests that measured lactate, creatinine, blood urea nitrogen, and white blood cells. This is used to diagnose two critical problems for ICU patients – kidney failure and sepsis.
“Because one of our goals is to think about whether we can reduce the number of lab tests, we began to see the most ordered [blood test] panel,” said Li-Fang Cheng, lead author of the joint study. with Niranjani Prasad.
The research team used a “gift function” that shows the sequence of tests based on how informative the test is at a given time. There are more prizes for testing when there is a higher probability that the patient’s condition is substantially different from the previous measurements and when the test results tend to show clinical interventions, such as giving antibiotics or helping breathing through mechanical ventilation.
On the other hand, the function adds a penalty to the cost of the test and the risk to the patient from the test. Based on the situation, a clinician can choose to prioritize one of these components than the other.
Known as reinforcement learning, the purpose of this method is to recommend assessments that maximize rewards. This turns medical testing into a matter of sequential decision making, “where you are responsible for all the decisions and all states that you have seen in the past and decide what you have to do at this time to maximize long-term benefits to be patient,” explained Prasad.
This method requires a lot of computing power to sort out information fast enough for clinical settings, Engelhardt added.
“This is one of the first times we will be able to take this machine learning approach and actually put it in the ICU, or in an inpatient hospital, and advise nurses in a way that patients will not be at risk. , “Engelhardt said. “That’s really something new.”
To examine the usefulness of laboratory testing policies produced by machine learning, the researchers compared the value of reward functions that would result from applying their approach to a testing regimen that was actually used for more than 6,000 patients in the data set and randomly. lab testing method.
For each test function and award, the lab test policy developed by machine learning algorithms will produce better prize values compared to the actual policies used in the hospital and, in most cases, random policies. A noteworthy exception is lactate testing; this result can be attributed to the low frequency of orders for lactate testing, which results in a high degree of variance in test informativeness, the researchers noted.
The researchers’ analysis showed that their optimized policies would produce more data than the testing regime used by doctors. Utilizing an algorithm might have reduced the number of laboratory test orders by 44 percent in the case of white blood cell testing. In addition, the researchers showed that this approach could tell doctors to intervene several hours faster when the patient’s condition began to deteriorate.
The researchers are working with data scientists at the Penn Medicine Predictive Health Team to launch a machine learning-based approach at the clinic in the next few years.
“Having access to machine learning, artificial intelligence and statistical modeling with a large amount of data” will help doctors “make better decisions, and ultimately improve patient outcomes,” said Penn Senior Data Scientist Corey Chivers.
The team algorithm uses a “gift function” that drives the sequence of tests based on how informative the test is at a given time. In other words, there is a greater reward in giving the patient a test if there is a higher probability that the patient’s condition is significantly different from the previous measurement.
To test the usefulness of the laboratory testing policies they made, the researchers compared the value of the reward function that would result from implementing their system with a testing regimen that was actually used for 6,060 patients in this study.
The researchers found that the policies produced by the machine learning algorithm would produce more information about the patient’s condition than the actual testing their doctors followed.
In addition, when looking at white blood cell tests, algorithms can reduce the number of lab test orders by 44 percent.
They also found their approach would help remind doctors to intervene sometimes several hours faster when the patient’s condition began to deteriorate.