Machine Learning Could Reduce Lab Testing for ICU Patients

January 18, 2019 – Princeton researchers have found that machine learning can be used to improve treatment and reduce lab testing for patients in intensive care units (ICUs).

The researchers used data from more than 6,000 ICU patients to design a machine learning-based system that could reduce frequency of lab tests and improve timing of critical treatments.

The researchers used the MIMIC III critical care database, which includes records of 58,000 critical care admissions at Beth Israel Deaconess Medical Center. For the 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 common vital signs and lab tests.

“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 Princeton Associate Professor of Computer Science Barbara Engelhardt, who is the senior author of the study.

The analysis focused on four blood tests measuring lactate, creatinine, blood urea nitrogen, and white blood cells. These were used to diagnose two critical issues for ICU patients—kidney failure and sepsis.

READ MORE: AWS Launches Machine Learning Service to Mine Patient Data

“Since one of our goals was to think about whether we could reduce the number of lab tests, we started looking at the [blood test] panels that are most ordered,” said Li-Fang Cheng, co-lead author of the study along with Niranjani Prasad.

The research team used a “reward function” that indicates a test order based on how informative the test will be at a given moment. There is more of a reward for administering a test when there is a higher probability that the state of the patient is substantially different from the previous measurement and when the test outcome is likely to indicate a clinical intervention, such as giving antibiotics or assisting breathing through mechanical ventilation.

On the other hand, the functionality adds a penalty for the cost of the test and risk to the patient from the test. Based on the situation, a clinician could opt to prioritize one of these components over others.

Known as reinforcement learning, this method’s goal is to recommend judgements that maximize the reward. This turns medical testing into a sequential decision-making problem, “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,” explained Prasad.

This method requires a lot of computing power to sort through the information quickly enough for a clinical setting, added Engelhardt.

READ MORE: Using Healthcare AI, Machine Learning for HIT Infrastructure

“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.”

To check the usefulness of the lab testing policy generated by machine learning, the researchers compared the reward function values that would have resulted from applying their approach to the testing regimens that were actually employed for the more than 6,000 patients in the data set and to randomized lab testing methods.

For each test and reward function, the lab test policy developed by the machine learning algorithm would have led to better reward values compared to the actual policies used in the hospitals and, in most cases, to randomized policies. A noteworthy exception was lactate testing; this result could be attributed to the low frequency of orders for lactate testing, which resulted in a high degree of variance in the informativeness of the test, the researchers noted.

The researchers’ analysis demonstrated that their optimized policy would have produced more data than did the testing regime that clinicians used. Utilizing the algorithm may have reduced the quantity of lab test orders by up to 44 percent in the case of white blood cell testing. In addition, the researchers demonstrated that this approach could have informed clinicians to intervene hours sooner when a patient’s condition started to deteriorate.

The researchers are working with data scientists on Penn Medicine’s Predictive Healthcare Team to launch the machine learning-based approach in the clinic within the next several years.

“Having access to machine learning, artificial intelligence and statistical modeling with large amounts of data” will help clinicians “make better decisions, and ultimately improve patient outcomes,” said Penn Senior Data Scientist Corey Chivers.