Background: Despite guideline recommendations, most patients with COPD do not undergo alpha-1 antitrypsin deficiency (AATD) testing and approximately 90% of people with AATD in the US remain undiagnosed. This study sought to develop a predictive model using real-world data to improve detection of AATD-positive patients in the general COPD population.
Methods: A predictive model using XGBoost was developed using the EVERSANA database, including longitudinal, patient-level medical claims, prescription claims, AATD-specific testing data, and electronic health records (EHR). The model was trained and then validated to predict AATD-positive status. Patients were coded as AATD positive based on the presence of any of the following criteria: 1) ≥2 AATD diagnosis codes in claims; 2) an AATD diagnosis code in the EHR; 3) a positive laboratory test for AATD; or 4) use of AATD-related medication. Over 500 variables were used to train the predictive model and >20 models were run to optimize the predictive power.
Results: 13,585 AATD-positive patients and 7,796 AATD-negative patients were included in the model. Inclusion of non-AATD laboratory test results was critical for defining cohorts and optimizing model prediction (e.g., respiratory comorbidities, and calcium, glucose, hemoglobin, and bilirubin levels). The final model yielded high predictive power, with an area under the receiver operating characteristic curve of 0.9.
Conclusion: Predictive modeling using real-world data is a sound approach for assessing AATD risk and useful for identifying COPD patients who should be confirmed by genetic testing. External validation is warranted to further assess the generalizability of these results.
Keywords: alpha-1 antitrypsin deficiency; detection; electronic health records; predictive model; testing.
JCOPDF © 2024.