Meeting News

Novel algorithms identify patients with axSpA

CHICAGO — Researchers developed three algorithms that successfully identified patients with axial spondyloarthritis.

During a presentation at the ACR/ARHP 2018 Annual Meeting, Rebecca S. Overbury, MD, MS, assistant professor of internal medicine and pediatrics in the division of rheumatology at the University of Utah, said the algorithms could help expand research in axSpA, which is currently limited.

“There is an unmet need for big data research in axSpA for real-life outcomes, including comorbidities, mortality, health care utilization and associated costs,” she said. “Traditional methods of cohort identification for axSpA in large datasets are currently insufficient.”

According to Overbury, there are no billing codes for most axSpA subtypes. In addition, she said that previous research has shown ICD-9 codes for ankylosing spondylitis (AS) lack sensitivity in “identifying data for the broader spectrum of axSpA patients.”

To address this unmet need for big data research, Overbury and colleagues set out to develop accurate methods for identifying axSpA in large datasets. To do this, the researchers used common strategies for cohort identification, including a rule-based strategy, natural language processing (NLP), and combinations of both. For the rule-based strategy, the researchers relied on structured data, such as combinations of ICD codes, laboratory data, medication dispensations and health care utilization. For NLP, computers were used to identify words and phrases in clinical documents that are suggestive of specific conditions.

The researchers used a reference population of 600 patients from facilities affiliated with the Veterans Health Administration with risk factors for axSpA to develop and test new methods for axSpA identification. After reviewing medical records and determining axSpA status, data from a subset of 451 patients were used to develop algorithms for axSpA identification. The algorithms were assessed in the remaining 149 patients.

Overbury and colleagues identified 47 variables that could potentially differentiate between patients with or without axSpA. With these variables, three algorithms were developed: the Full Algorithm, which included all 47 variables; the High Feasibility algorithm, which included 16 of the most predictive coded variables; and the Spond NLP algorithm, which used NLP as the single variable.

Of the 600 patients in the reference population, 27% (n = 162) had axSpA. In the testing population, the Full Algorithm predicted axSpA with 87.5% sensitivity and 91.7% specificity; the High Feasibility Algorithm predicted axSpA with 85% sensitivity and 93.6% specificity; and the Spond NLP Algorithm predicted axSpA with 95% sensitivity and 78% specificity.

The false-positive rate was 6% for the Full Algorithm, 4.7% for the High Feasibility Algorithm and 16.1% for the Spond NLP Algorithm. Meanwhile, the false-negative rate was 3.4% for the Full Algorithm, 4% for the High Feasibility Algorithm and 1.3% for the Spond NLP Algorithm.

“Overall, our Full Algorithm demonstrated the highest specificity and may be appropriate when low false-positive rates are desired, such as for treatment considerations,” Overbury said. “Our High Feasibility Algorithm was the simplest model, and so may be applied and tested relatively easily in different datasets. Our NLP model demonstrated the highest sensitivity, suggesting it may be best used when more inclusive identification is desired, such as for screening.”

Overbury noted several limitations. NLP, she said, is expensive and requires bioinformatics expertise as well as computing resources. In addition, since the study used ICD-9 codes, more research is underway to test the algorithms using ICD-10 codes. Finally, the results cannot be generalized to non-VA datasets without further validation.

Still, Overbury said the “successfully-designed novel methods for the development of inclusive axSpA cohorts” resulted in sensitive and specific algorithms.

“The algorithms offer a range of performance and feasibility attributes that may be appropriate for a broad array of axSpA research,” she said. – by Stephanie Viguers

Reference:

Overbury RS, et al. Abstract 893. Presented at: ACR/ARHP Annual Meeting, Oct. 20-24, 2018; Chicago.

Disclosure: Overbury reports no relevant financial disclosures.

CHICAGO — Researchers developed three algorithms that successfully identified patients with axial spondyloarthritis.

During a presentation at the ACR/ARHP 2018 Annual Meeting, Rebecca S. Overbury, MD, MS, assistant professor of internal medicine and pediatrics in the division of rheumatology at the University of Utah, said the algorithms could help expand research in axSpA, which is currently limited.

“There is an unmet need for big data research in axSpA for real-life outcomes, including comorbidities, mortality, health care utilization and associated costs,” she said. “Traditional methods of cohort identification for axSpA in large datasets are currently insufficient.”

According to Overbury, there are no billing codes for most axSpA subtypes. In addition, she said that previous research has shown ICD-9 codes for ankylosing spondylitis (AS) lack sensitivity in “identifying data for the broader spectrum of axSpA patients.”

To address this unmet need for big data research, Overbury and colleagues set out to develop accurate methods for identifying axSpA in large datasets. To do this, the researchers used common strategies for cohort identification, including a rule-based strategy, natural language processing (NLP), and combinations of both. For the rule-based strategy, the researchers relied on structured data, such as combinations of ICD codes, laboratory data, medication dispensations and health care utilization. For NLP, computers were used to identify words and phrases in clinical documents that are suggestive of specific conditions.

The researchers used a reference population of 600 patients from facilities affiliated with the Veterans Health Administration with risk factors for axSpA to develop and test new methods for axSpA identification. After reviewing medical records and determining axSpA status, data from a subset of 451 patients were used to develop algorithms for axSpA identification. The algorithms were assessed in the remaining 149 patients.

Overbury and colleagues identified 47 variables that could potentially differentiate between patients with or without axSpA. With these variables, three algorithms were developed: the Full Algorithm, which included all 47 variables; the High Feasibility algorithm, which included 16 of the most predictive coded variables; and the Spond NLP algorithm, which used NLP as the single variable.

Of the 600 patients in the reference population, 27% (n = 162) had axSpA. In the testing population, the Full Algorithm predicted axSpA with 87.5% sensitivity and 91.7% specificity; the High Feasibility Algorithm predicted axSpA with 85% sensitivity and 93.6% specificity; and the Spond NLP Algorithm predicted axSpA with 95% sensitivity and 78% specificity.

The false-positive rate was 6% for the Full Algorithm, 4.7% for the High Feasibility Algorithm and 16.1% for the Spond NLP Algorithm. Meanwhile, the false-negative rate was 3.4% for the Full Algorithm, 4% for the High Feasibility Algorithm and 1.3% for the Spond NLP Algorithm.

“Overall, our Full Algorithm demonstrated the highest specificity and may be appropriate when low false-positive rates are desired, such as for treatment considerations,” Overbury said. “Our High Feasibility Algorithm was the simplest model, and so may be applied and tested relatively easily in different datasets. Our NLP model demonstrated the highest sensitivity, suggesting it may be best used when more inclusive identification is desired, such as for screening.”

Overbury noted several limitations. NLP, she said, is expensive and requires bioinformatics expertise as well as computing resources. In addition, since the study used ICD-9 codes, more research is underway to test the algorithms using ICD-10 codes. Finally, the results cannot be generalized to non-VA datasets without further validation.

Still, Overbury said the “successfully-designed novel methods for the development of inclusive axSpA cohorts” resulted in sensitive and specific algorithms.

“The algorithms offer a range of performance and feasibility attributes that may be appropriate for a broad array of axSpA research,” she said. – by Stephanie Viguers

Reference:

Overbury RS, et al. Abstract 893. Presented at: ACR/ARHP Annual Meeting, Oct. 20-24, 2018; Chicago.

Disclosure: Overbury reports no relevant financial disclosures.

    See more from Discoveries from ACR: Spondyloarthropathies