Background/Purpose: Interstitial lung disease (ILD) affects 10-15% of RA patients leading to significant disability and premature mortality. While large administrative datasets have been widely leveraged in outcomes research, the accuracy of algorithms used to identify RA-ILD cases has not been adequately studied. Our objective was to evaluate the accuracy of administrative algorithms for identifying ILD in RA.

Methods: We selected participants in the Veterans Affairs Rheumatoid Arthritis (VARA) registry for detailed medical record review using a stratified sampling technique. Those with ≥2 outpatient or ≥1 inpatient ILD International Classification of Diseases (ICD) codes were reviewed as well as a 10% random sample of those without ILD codes. After standardizing the review process, three rheumatologists abstracted data from clinical notes, imaging/pathology reports, and pulmonary function testing (PFT). ‘Stringent’ and ‘relaxed’ ILD definitions based on medical record review were constructed a priori. Administrative data was obtained from the VA Corporate Data Warehouse and included outpatient and inpatient ICD codes, ICD codes from non-VA providers, provider specialty, diagnostic testing (imaging, biopsy, PFTs), and corresponding dates. We tested the algorithms in phases: 1) number and type of encounters, 2) specific ICD codes, 3) provider specialty and diagnostic testing, and 4) exclusion of ‘other’ ILD. Sensitivity, specificity, positive and negative predictive value (PPV), and agreement (Kappa statistic) with medical record classification were assessed. Analyses accounted for the stratified sampling design using inverse probability weighting. Several sensitivity analyses were completed.

Results: Characteristics of selected participants (n=536) were reflective of the VARA registry (n=2640) and VA population. ILD was confirmed in 182 (stringent) and 203 (relaxed), with pulmonologist diagnosis and chest computed tomography (CT) evidence in the majority. Initially, we identified ≥2 ILD ICD codes from inpatient or outpatient encounters as the best discriminating factors (PPV 65.5%, Kappa 0.70). Optimal ICD codes were ICD-9 515, 516.3, 516.8, and 516.9 and ICD-10 J84.1, J84.89, and J84.9 (PPV 69.5%, Kappa 0.72). Algorithms including a pulmonologist diagnosis, chest CT, PFTs, lung biopsy, or combinations improved performance (PPV 77.4-81.9%, Kappa 0.75). Exclusion of ‘other’ ILD modestly improved PPV further (78.5-82.9%). In sensitivity analyses, comparisons against even more sensitive ILD reference-standards resulted in PPVs of 82.4-86.3% while other minor algorithm modifications did not significantly affect algorithm performance.

Conclusion: The administrative algorithms identified in this study demonstrate substantial agreement with medical record review in classifying ILD in RA patients, though PPV is limited by the relatively low prevalence of RA-ILD. Optimal algorithms with PPV ≥80-85% contained multiple ILD encounters, specific ICD codes, pulmonologist diagnosis or diagnostic testing, and exclusion of ‘other’ ILD. These algorithms can be leveraged for RA-ILD outcomes research.

« Back to 2019 ACR/ARP Annual Meeting

ACR Meeting Abstracts - https://acrabstracts.org/abstract/accuracy-of-administrative-algorithms-for-identifying-interstitial-lung-disease-in-rheumatoid-arthritis/