Performance of Administrative Algorithms to Identify Interstitial Lung Disease in Rheumatoid Arthritis

Objective To determine the performance of administrative-based algorithms for classifying interstitial lung disease (ILD) complicating rheumatoid arthritis (RA). Methods Participants in a large, multicenter RA registry were screened for ILD using codes from the International Classification of Diseases, Ninth Revision (ICD-9) and the ICD-10. Medical record review confirmed ILD among participants screening positive and a random sample of those screening negative. ICD and procedure codes, provider specialty, and dates were extracted from Veterans Affairs administrative data to construct ILD algorithms. Performance of these algorithms against medical record review was assessed by sensitivity, specificity, positive predictive value (PPV), negative predictive value, and kappa using inverse probability weighting to account for sampling methods. Results Medical records of 536 RA patients were reviewed, confirming 182 (stringent definition) and 203 (relaxed definition) cases of ILD. Initially, we identified >= 2 ICD codes from inpatient or outpatient encounters as optimal discriminating factors (specificity 96.0%, PPV 65.5%; kappa = 0.70). Subsequently, we constructed a set of ICD-9 and ICD-10 codes that improved algorithm specificity (specificity 96.8%, PPV 69.5%; kappa = 0.72). Algorithms that included a pulmonologist diagnosis or chest computed tomography plus pulmonary function testing or lung biopsy had improved performance (specificity 98.0%, PPV 77.4%; kappa = 0.75). PPV increased with exclusion of other ILD causes (78.5%) in comparison with the relaxed ILD definition (82.4%) and in sensitivity analyses (83.4-86.3%). Gains in specificity and PPV with greater algorithm requirements were accompanied by declines in sensitivity. Conclusion Administrative algorithms with optimal combinations of ICD codes, provider specialty, diagnostic testing, and exclusion of other ILD causes accurately classify ILD in RA.