(263) Clinical Evaluation of Code-Based Algorithms to Identify Pulmonary Arterial Hypertension Patients in Coded Healthcare Databases

Publication Number: 594

Background: Pulmonary Arterial Hypertension (PAH) is a clinical sub-classification of Pulmonary Hypertension (PH). It is a severe, complex, and rare disease, and real-world data are scarce. Claims and Electronic Heath Record (EHR) databases can support observational research in PAH; however, differentiating PAH patients from the overall PH patient population based on administrative codes is challenging.

Objectives: The study objective was to develop guidance on the use of combinations of administrative codes for PAH patient identification (PAH algorithms), for aligned research in coded healthcare databases. PAH algorithm performance was evaluated by checking PAH patient cohorts identified based on administrative codes against their true clinical diagnoses.

Methods: Patient records from the Vera Moulton Wall Center clinical PH database (true diagnoses) were linked with coded EHR data from Stanford University Medical Center. Over a dozen published PAH algorithms and clinically meaningful algorithm modifications were applied to separate PAH from non-PAH patients in the coded EHR data. Using the clinical PH database, each patient’s cohort assignment was assessed against their true diagnoses.

Results: Records from 720 PH patients (558 PAH and 162 non-PAH patients) from the clinical PH database could be linked with EHR data. Algorithms consisting of only 1 P(A)H-specific diagnosis code separated PAH patients from non-PAH patients with sensitivity >97% and positive predictive value (PPV) >79%, while specificity was <12%. Such algorithms tend to classify all PH patients as PAH patients. Highest specificity (100%) and PPV (100%) were obtained for an algorithm requiring ≥1 PAH diagnosis code OR ≥1 PH diagnosis code + ≥1 diagnosis code for a PAH-associated disease, AND ≥1 PAH drug code, AND exclusionary codes for other forms of PH. This algorithm did not classify non-PAH patients as PAH patients but missed 520 true PAH patients (sensitivity: 7%). The algorithm that was most balanced in terms of sensitivity, NPV and PPV (all > 70%) included ≥1 PH diagnosis AND ≥1 PAH drug AND exclusionary codes.

Conclusions: There is no universal algorithm appropriate across all research questions. To identify a “clean” PAH cohort – e.g., for diagnostic studies –, a complex algorithm with combinations of diagnosis, procedure, drug, and exclusionary codes should be used. However, such an algorithm likely misses true PAH patients. To identify all PAH patients – e.g., for treatment pattern studies –, one diagnosis code might be sufficient, but the resulting cohort likely contains non-PAH patients, too. Algorithm performance might vary between databases and countries.