NEW YORK – A recent study has demonstrated that machine learning methods can sort out which laboratory tests used in hospitals may have greater patient-care value, and which ones don't.
The study, which was led by Stanford University researchers and published this week in JAMA, analyzed data from 191,506 inpatients at three different hospitals to predict which diagnostic lab tests are medically unnecessary.
The researchers used prediction models to assess data from 116,637 inpatients treated at Stanford University Hospital from Jan. 1, 2008, to Dec. 3, 2017; 60,929 inpatients at the University of Michigan from Jan 1, 2015, to Dec. 31, 2018; and 13,940 inpatients at the University of California, San Francisco from Jan. 1 to Dec. 31, 2018.
They found that low-value diagnostic tests are commonly used in hospitals and can be identified systematically using machine learning models, which can quantify the level of uncertainty and expected info to be gained from these tests.
They noted that while "numerous interventions have been studied to reduce inappropriate laboratory testing … unnecessary tests remain prolific" due to fears of missing problems, the threat of lawsuits, patient preferences, and "the overall difficulty of systematically identifying low-value testing at the point of care."
Lab overutilization is a significant cost for many hospitals, and citing statistics from the Academy of Medicine, the researchers added that more than $200 million a year is wasted on unnecessary tests and procedures.
An estimated up to 25 to 50 percent of all inpatient lab testing is medically unnecessary, making reducing these tests a priority for many health systems, they said.
Just this week, laboratory testing firm Quest Diagnostics and healthcare solutions company Hc1 announced a partnership to analyze lab data to manage test usage and identify areas where lab tests are inappropriately ordered.
The study's researchers aimed to identify diagnostic lab tests with predictable results that were unlikely to deliver new information.
They used descriptive statistics to predict the result of each lab test using the information that was available before the test was ordered. This info included patient demographics, normality of the most recent test of interest, numbers of recent tests of interest, history of Charlson Comorbidity Index categories, specialty teams treating the patient, time since admission, time of the day and year of the test, and summary statistics of recent vitals and lab results.
Further, by utilizing machine learning models, the researchers were able to personalize results and provide more accurate predictions.
In particular, they noted that while some tests – such as serum albumin, thyrotropin, as well as phosphorus, and complete blood cell counts – have already been identified as being vulnerable to potential overutilization, other low-yield tests are more difficult to identify. A machine learning-based method, the researchers said, provides a systematic and quantitative approach to making more informed clinical decision about the appropriateness of tests.
The study "should encourage practitioners and quality improvement committees to make explicit and quantitative their own embedded assumptions on acceptable decision thresholds," they wrote.
The method, the researchers said, could be used to implement and optimize guideline and protocol-based testing. In the case of lactate testing for sepsis protocols, for example, "the results of this study can still inform the development of such regulatory requirement on the appropriate number and interval of screening tests that may otherwise be excessive of too rigid for individual cases," the study said.
It can also be used in hospitals' cost-effectiveness measures, since hospitals can cut back on those tests that are overused. However, the savings need to be compared against the potential harm or cost of missing an abnormal test, the researchers noted.