A computational phenotype combining ICD-10 codes, long-acting bronchodilator prescriptions, and lung function testing history identified true airflow obstruction in only 62% of patients.
Cross-Sectional (n=609)
No
Does a computational phenotype using structured EHR data accurately identify patients with COPD who meet spirometry criteria for airflow obstruction?
A computational phenotype using structured EHR data (ICD codes, medications, lung function testing history) had low accuracy (62%) in identifying true COPD patients with airflow obstruction.
Abstract Rationale Computational phenotypes are data algorithms used in clinical and epidemiologic research to identify patient cohorts with a disease based on structured or unstructured electronic health record (EHR) data elements. Computational phenotypes to identify patients with chronic obstructive pulmonary disease (COPD) commonly utilize ICD diagnosis codes. However, previous research suggests that just over half of patients with an ICD diagnosis of COPD meet guideline-based diagnostic criteria including airflow obstruction on post-bronchodilator spirometry. This study aimed to evaluate the performance of a computational phenotype that utilized a combination of structured data elements including ICD codes, medications, and a history of lung function testing for identifying patients with COPD who meet spirometry criteria for airflow obstruction. Methods We performed a cross-sectional study using EHR data from a single academic hospital in Chicago. We included patients aged 40 years with an ICD-10 diagnosis of COPD (J41.*, J42, J43.*, or J44.*) during an encounter in the previous two years, an active long-acting bronchodilator prescription, and a previous encounter for lung function testing. Demographic data was collected as part of the data query. We performed manual chart review of pulmonary function testing to determine if post-bronchodilator spirometry met criteria for airflow obstruction based on FEV1/FVC0.70. Patients with and without airflow obstruction were compared using chi-square and student t-tests as appropriate. Results We found 617 patients who met inclusion criteria, though 8 patients did not have post-bronchodilator spirometry available. Of the remaining 609 patients, the mean age was 65 years old, 354 (58%) were female, 432 (71%) were Black, 102 (17%) were White, and 75 (12%) were another race. Only 376 (62%) of the study sample had airflow obstruction with a post-bronchodilator FEV1/FVC0.70. Compared to those with airflow obstruction, those without airflow obstruction were significantly younger (mean age 66 years vs 63 years, respectively) and more likely to be female. Although non-White and Hispanic patients had lower rates of airflow obstruction, the differences were not statistically significant. There were no significant between-group differences in insurance, language, and smoking history. Conclusions A computational phenotype using a combination of structured data including ICD-10 codes, long-acting bronchodilator prescription, and history of lung function testing did not identify patients with COPD and airflow obstruction with high accuracy. Utilizing natural language processing models to extract unstructured spirometry data may enhance the accuracy of computational phenotypes for identifying patients with COPD who meet guideline-based diagnostic criteria. This abstract is funded by: The National Center for Advancing Translational Sciences and National Institutes of Health through Grant Award Numbers KL2TR002002 and UL1TR002003
Liu et al. (Fri,) conducted a cross-sectional in Chronic Obstructive Pulmonary Disease (COPD) (n=609). Computational phenotype (ICD-10 codes, medications, and lung function testing history) vs. Manual chart review (post-bronchodilator FEV1/FVC<0.70) was evaluated on Airflow obstruction with a post-bronchodilator FEV1/FVC<0.70. A computational phenotype combining ICD-10 codes, long-acting bronchodilator prescriptions, and lung function testing history identified true airflow obstruction in only 62% of patients.