CLN - Focus on Data Science
Sponsored By
Clinical laboratories can leverage secondary data — information collected for clinical care but useful for other purposes — to improve healthcare for a variety of patient populations.
The electronic health record (EHR) is full of these real-world data, and in recent years has become more accessible through data warehouses and related tools. “These data can be useful for understanding population health and the impact of laboratory testing at the population level,” said Lee F. Schroeder, MD, PhD, associate director in the division of clinical pathology at the University of Michigan.
“For example, the data can be used to evaluate how labs affect the cascade of care and demonstrate different diagnostic strategies’ impact on outcomes that are important to health systems, like identifying patients that would truly benefit from referrals, and in the process, reducing scheduling delays for specialty care,” Schroeder said. “Laboratory data can also be leveraged to understand gaps in health services delivery in different patient groups.”
Schroeder gave one of two presentations that demonstrated the usefulness of secondary data from EHRs at the Association for Diagnostics & Laboratory Medicine (ADLM, formerly AACC) July 2024 Data Science Summit. The event, held after ADLM’s annual meeting, showcased labs’ use of data to improve patient care, a strategy that is becoming ever more common and important in medicine.
Schroeder’s presentation showed how a celiac disease testing algorithm helped reduce unnecessary biopsies by improving how physicians ordered tests early in the diagnostic process. The second talk demonstrated how linking geospatial tools to EHR data can find inequities in kidney health screening among people with diabetes.
Improving the value of referrals for biopsy
As is the case in most specialties, gastroenterology patients often are subject to scheduling delays and many wait months for appointments. “Any change in the diagnostic process that could reduce low-value referrals could greatly benefit patient care,” Schroeder said.
In an interview, Schroeder explained how he used about 10 years’ worth of EHR data to study a University of Michigan celiac disease testing algorithm designed to identify the subset of tests that make the most sense for each patient, thus optimizing sensitivity and specificity.
The algorithm identifies the patient’s immunoglobulin A (IgA) levels and then reflexes to either IgA, immunoglobulin G (IgG), or both; versions of antitissue transglutaminase and antideaminated gliadin antibodies; and possibly antiendomysial antibodies. “What you want to do is identify which tests are going to pay off, and which may lead to unnecessary procedures, and just run those that will benefit each patient,” Schroeder said. As celiac screening tests lead directly to referrals for biopsy, it is particularly important to reduce false negatives as well as false positives, he added.
As part of his unpublished study, Schroeder filtered data for first-time celiac screening events using ICD codes from the EHR, labeling each screening event as either appropriate (consistent with the algorithm) or inappropriate and collected biopsy results. He found that when the study started in 2014, the proportion of appropriate orders consistent with the algorithm started out low, less than 10%. After an algorithm name change to make its purpose more obvious, the percentage of appropriate orders increased to about 60% — and was hovering at about 80% at the study’s end in 2023.
Schroeder also found that biopsy rate for patients whose clinicians ordered appropriately was 14.9%, versus 18.7% for clinicians who did not order tests consistent with the algorithm.
When Schroeder looked at changes in biopsy rate by physician specialty over the course of the study, he found notable differences among specialty services. The service with the largest volume of ordering was internal medicine, which saw its biopsy rate drop by nearly 30%. Other services increased their biopsy ordering rates. The largest increase was in pediatric emergency medicine, which saw its biopsy rate nearly triple. The emergency medicine and hospitalist services each saw gains of about 60%.
Importantly, biopsy positivity rates increased for services that decreased their biopsy ordering rates, thus suggesting higher value referrals. “It also shows that the levers we can control from the laboratory are finding purchase throughout the diagnostic cascade,” Schroeder said.
EHRs provide a rich set of observational, secondary data that is less controlled than data in randomized clinical trials, so inferring causality is much more difficult, Schroeder noted. To reduce potential bias in this process, he also used propensity scores to reduce the effect of different patient types receiving various test panels.
“If those at lower risk of disease receive the algorithm while those at greater risk receive other testing strategies, this can cause significant selection bias and confound the study results,” he said. “Propensity score matching reduces this bias by selecting a subset of patients so that the two patient populations have a similar likelihood of getting either inappropriate or appropriate testing, thus attempting to mimic a randomized trial.” Schroeder also used logistic regression to control for factors such as year and patient demographics. The results were largely unchanged with these controls.
Using geospatial data to increase health equity
The electronic health record (EHR) is full of these real-world data, and in recent years has become more accessible through data warehouses and related tools. “These data can be useful for understanding population health and the impact of laboratory testing at the population level,” said Lee F. Schroeder, MD, PhD, associate director in the division of clinical pathology at the University of Michigan.
“For example, the data can be used to evaluate how labs affect the cascade of care and demonstrate different diagnostic strategies’ impact on outcomes that are important to health systems, like identifying patients that would truly benefit from referrals, and in the process, reducing scheduling delays for specialty care,” Schroeder said. “Laboratory data can also be leveraged to understand gaps in health services delivery in different patient groups.”
Schroeder gave one of two presentations that demonstrated the usefulness of secondary data from EHRs at the Association for Diagnostics & Laboratory Medicine (ADLM, formerly AACC) July 2024 Data Science Summit. The event, held after ADLM’s annual meeting, showcased labs’ use of data to improve patient care, a strategy that is becoming ever more common and important in medicine.
Schroeder’s presentation showed how a celiac disease testing algorithm helped reduce unnecessary biopsies by improving how physicians ordered tests early in the diagnostic process. The second talk demonstrated how linking geospatial tools to EHR data can find inequities in kidney health screening among people with diabetes.
Improving the value of referrals for biopsy
As is the case in most specialties, gastroenterology patients often are subject to scheduling delays and many wait months for appointments. “Any change in the diagnostic process that could reduce low-value referrals could greatly benefit patient care,” Schroeder said.
In an interview, Schroeder explained how he used about 10 years’ worth of EHR data to study a University of Michigan celiac disease testing algorithm designed to identify the subset of tests that make the most sense for each patient, thus optimizing sensitivity and specificity.
The algorithm identifies the patient’s immunoglobulin A (IgA) levels and then reflexes to either IgA, immunoglobulin G (IgG), or both; versions of antitissue transglutaminase and antideaminated gliadin antibodies; and possibly antiendomysial antibodies. “What you want to do is identify which tests are going to pay off, and which may lead to unnecessary procedures, and just run those that will benefit each patient,” Schroeder said. As celiac screening tests lead directly to referrals for biopsy, it is particularly important to reduce false negatives as well as false positives, he added.
As part of his unpublished study, Schroeder filtered data for first-time celiac screening events using ICD codes from the EHR, labeling each screening event as either appropriate (consistent with the algorithm) or inappropriate and collected biopsy results. He found that when the study started in 2014, the proportion of appropriate orders consistent with the algorithm started out low, less than 10%. After an algorithm name change to make its purpose more obvious, the percentage of appropriate orders increased to about 60% — and was hovering at about 80% at the study’s end in 2023.
Schroeder also found that biopsy rate for patients whose clinicians ordered appropriately was 14.9%, versus 18.7% for clinicians who did not order tests consistent with the algorithm.
When Schroeder looked at changes in biopsy rate by physician specialty over the course of the study, he found notable differences among specialty services. The service with the largest volume of ordering was internal medicine, which saw its biopsy rate drop by nearly 30%. Other services increased their biopsy ordering rates. The largest increase was in pediatric emergency medicine, which saw its biopsy rate nearly triple. The emergency medicine and hospitalist services each saw gains of about 60%.
Importantly, biopsy positivity rates increased for services that decreased their biopsy ordering rates, thus suggesting higher value referrals. “It also shows that the levers we can control from the laboratory are finding purchase throughout the diagnostic cascade,” Schroeder said.
EHRs provide a rich set of observational, secondary data that is less controlled than data in randomized clinical trials, so inferring causality is much more difficult, Schroeder noted. To reduce potential bias in this process, he also used propensity scores to reduce the effect of different patient types receiving various test panels.
“If those at lower risk of disease receive the algorithm while those at greater risk receive other testing strategies, this can cause significant selection bias and confound the study results,” he said. “Propensity score matching reduces this bias by selecting a subset of patients so that the two patient populations have a similar likelihood of getting either inappropriate or appropriate testing, thus attempting to mimic a randomized trial.” Schroeder also used logistic regression to control for factors such as year and patient demographics. The results were largely unchanged with these controls.
Using geospatial data to increase health equity
Labs, as well as all parts of the healthcare system, increasingly focus on providing equitable access to quality healthcare, with a focus on patient socioeconomic factors. One challenge to determining equitable access is getting data that can help characterize patients’ vulnerability. Such data are not always collected comprehensively at the patient level, Schroeder noted.
To determine nonmedical factors that affect patient health — also called social determinants of health — Schroeder and Vahid Azimi, MD, of Washington University in St. Louis School of Medicine teamed up on a study that used geospatial tools to determine equity in access to testing for kidney health, which is important in diabetes care. The pair conducted geospatial analyses on U.S. Census tracts where patients lived and used that population-level data as a proxy for patient-level features. Azimi presented this research at the ADLM Data Science Summit.
The unpublished study relied on both the census tract data and information from Schroeder and Azimi’s respective health systems’ EHRs to determine whether social vulnerability affected kidney health testing in people with diabetes.
By looking at census data — including age, sex, employment status, education, housing, and other social factors — the pair found that social vulnerability is significantly associated with whether diabetic patients get appropriate kidney tests during encounters with the health system.
Azimi found that many patients went to the emergency room, likely for something other than a kidney issue, and received a serum creatinine test because it’s part of the basic metabolic panel, but they didn’t get the urine albumin test. Azimi suggested that these hospital encounters are opportunities to fix the gaps in kidney care for patients with diabetes. Azimi also showed that 80% of those with only a serum creatinine result also had at least one outpatient encounter that provided another opportunity for urine testing.
EHR and geospatial data were just two types of data discussed at the summit. A running theme through many talks was use of artificial intelligence and large language models throughout the pre-analytic, analytic, and post-analytic phases of the diagnostic cascade. These new tools are revealing patterns that may not be obvious to humans, creating new ways to access knowledge — a situation that Schroeder likened to switching to the internet as a means of getting information after decades of looking it up in printed encyclopedias.
Schroeder predicted that data science will be essential for laboratory medicine to enhance patient care and demonstrate the impact of diagnostic strategies on health systems. “That goes beyond analytic accuracy and turnaround times to impacting health outcomes at the population level,” he said.
Deborah Levenson is Deborah Levenson is a freelance writer in College Park, Maryland. +Email: [email protected].
Read the full November/December 2024 CLN issue here.