A large retrospective study of test orders for almost 200,000 inpatients seen at three academic institutions -- Stanford University, the University of Michigan, and the University of California, San Francisco -- found that the best machine-learning models could have predicted with high accuracy the likelihood of normal and unnecessary results for commonly performed lab tests known to have low utility.
The study results were reported by Dr. Jonathan Chen, PhD, an assistant professor at the Stanford Center for Biomedical Informatics Research, and colleagues in JAMA Network Open online September 11.
"The findings suggest that data-driven methods can make explicit the level of uncertainty and expected information gain from diagnostic tests, with the potential to encourage useful testing and discourage low-value testing that can incur direct cost and indirect harm," Chen and colleagues wrote.
High rates of unnecessary testing
The study was conducted in recognition of the waste in the U.S. healthcare system associated with the ordering of unnecessary tests. The authors cited the Institute of Medicine's estimate that $200 billion a year goes toward unneeded tests and procedures.
"Laboratory testing, in particular, constitutes the highest-volume medical procedure, with estimates of up to 25% to 50% of all inpatient testing being medically unnecessary," they wrote. "The consequences of unnecessary testing are not simply financial but also include low patient satisfaction, sleep fragmentation, risk of delirium, iatrogenic anemia, and increased mortality."
There have been many interventions to rein in use, including clinical education, Choosing Wisely guidelines from the American Society for Clinical Pathology, and electronic medical record-based ordering restrictions. However, unnecessary tests "remain prolific" for many reasons, including fear of missing problems, malpractice concerns, and the "overall difficulty of systematically identifying low-value testing at the point of care, prompting behavior to check just in case," the authors wrote.
The study was backed by the U.S. National Institutes of Health's Big Data to Knowledge (BD2K) award and a grant from the Gordon and Betty Moore Foundation. The researchers tested machine-learning approaches in nearly 200,000 inpatients treated between 2008 and 2018 at the three institutions (116,637 from Stanford, 60,929 from the University of Michigan, and 12,940 from the University of California, San Francisco). Many high-volume tests had been repeated within 24 hours.
"For example, hundreds of tests for serum albumin, thyrotropin, and glycated hemoglobin levels were performed again within 24 hours, along with tens of thousands of repetitive tests for phosphorus and complete blood cell counts with differential," Chen and colleagues wrote. "This finding quantitatively supports issues suggested in previous guidelines that hospitals can immediately use to target unnecessary repeated tests, such as through best practice alerts showing recently available test results."
The researchers applied machine-learning models to the information available at the time orders were placed and determined whether it was possible to predict results. The best-performing machine-learning models predicted normal results with an area under the receiver operating characteristic curve (AUC) of 0.90 or greater for 12 standalone laboratory tests, such as sodium and lactate dehydrogenase, and 10 common lab test components, they reported.
The methods allowed the group to identify factors predictive of test results, including prior results and gender.
"The findings suggest that low-yield diagnostic testing is common and can be systematically identified through data-driven methods and patient context-aware predictions," Chen and colleagues wrote.
Range of lab decision-support systems
Several companies have developed utilization management platforms that assist doctors in selecting and ordering the most appropriate lab tests. Safedin Beqaj, president and CEO of Irvine, CA-based Medical Database, and colleagues shared experience with the company's Laboratory Decision System (LDS) in a report published online September 5 in the open-access Journal of Clinical and Laboratory Medicine.
Laboratory test utilization management platforms define clinical parameters, gather and analyze data to make recommendations, and then ensure they are in line with best practices. Systems also can identify tests that have been completed and can flag findings that require follow-up testing.
Medical Database's platform offers a test scoring system, which is based on medical evidence, clinical relevancy, and medical necessity, as well as a ranking system that allows clinicians to select the most relevant tests based on diseases and codes. Additionally, it follows medical necessity guidelines from the U.S. Centers for Medicare and Medicaid Services (CMS) to prevent doctors from ordering highly complex or costly tests that are not needed.
Beqaj and colleagues reported data from two studies of the LDS that flagged problems with claims for tests. In one study of 96,170 orders from a reference laboratory, they found many claims for tests that had ICD-10 codes described as "never covered" by Medicare, based on medical necessity.
In a second study of 294,870 laboratory test claims from a preferred provider organization (PPO) managing self-pay insurers, the group found many issues with missing or invalid ICD-10 codes. Overall, more than 50% of submitted orders did not meet medical necessity according to Medibase's system, and more than 20% did not meet national and local coverage determination policies from CMS.
When it comes to clinical laboratory test ordering procedures, many problems can occur. Clinicians sometimes fail to follow appropriate use guidelines, algorithms, or directives, the authors noted. However, there are other factors that influence test ordering, including the following:
- Ease of ordering
- Ignorance of test characteristics
- Peer or supervisor pressure to be extremely thorough
- Fear of litigation
- Desire for certainty of diagnosis
- Financial incentives
- Patient demand
The U.S. is in the process of implementing regulations requiring an appropriate use consultation and electronic clinical decision-support system to support orders of imaging services for Medicare patients. Lawmakers have also recommended the use of similar systems for ordering other types of diagnostic tests. Beqaj and colleagues see a clear need for a similar system for labs to select the right test for each condition and assign the correct ICD-10 code to show medical necessity.
Aside from Medical Database, companies developing computer systems for labs include Beacon Laboratory Benefit Solutions and Abbott Diagnostics. Additionally, the Mayo Clinic and National Decision Support partnered to develop CareSelect Lab, a decision-support tool that aims to help healthcare providers order appropriate lab tests, improve patient care, and reduce wasteful spending.
An LDS system "can assist providers in making appropriate utilization decisions while also supporting laboratories in reimbursement and streamline claim verification for [payors], all of which combined will serve to make the laboratory industry and overall healthcare more efficient and cost-effective," the authors wrote.
Copyright © 2019 LabPulse.com