Development and validation of claims-based algorithms to identify interstitial lung disease among Japanese patients with cancer in routine clinical practice using real-world data sources.
Authors
Affiliations (9)
Affiliations (9)
- Global Patient Safety, Pharmacoepidemiology, Eli Lilly and Company, Indianapolis, IN, United States.
- Department of Medical Informatics, Graduate School of Medicine, Osaka University, Osaka, Japan.
- Department of Medical Informatics, Osaka University Hospital, Osaka, Japan.
- Department of Respiratory Medicine and Clinical Immunology, Graduate School of Medicine, Osaka University, Osaka, Japan.
- Department of Respiratory Medicine, NHO Osaka National Hospital, Osaka, Japan.
- Department of Radiology, NHO Osaka National Hospital, Osaka, Japan.
- Global Statistical Sciences, Eli Lilly and Company, Indianapolis, IN, United States.
- Japan Drug Development and Medical Affairs, Eli Lilly Japan K.K., Kobe, Japan.
- NHO Osaka National Hospital, Osaka, Japan.
Abstract
Efforts to validate claims-based algorithms for identifying patients with interstitial lung disease (ILD), an important safety concern in Japan, are limited. We developed and validated a claims-based algorithm for identifying ILD in Japanese patients with cancer using machine-learning modeling. This Japanese observational study used administrative claims and electronic medical record data collected in January 2013-March 2019 (Phase 1) and January 2015-March 2021 (Phase 2). Patients were classified as ILD cases based on chest computed tomography reports using natural language processing, with confirmatory reviews (ILD<sub>CT+</sub>). Machine-learning modeling strategies (logistic regression, least absolute shrinkage and selection operator [LASSO] logistic regression, and eXtreme Gradient Boosting) selected ILD identification variables from prespecified candidates. Model performances were estimated. Approximately 30% of randomly selected ILD<sub>CT+</sub> cases were adjudicated using medical records; algorithm performance was adjusted using adjudication results. The best-performing algorithm was validated using an external claims database. Among 13,601 eligible patients, 415 were ILD<sub>CT+</sub> cases; 123 were selected for adjudication. The best-performing model was the LASSO reduced model (using only the top variables identified in the full model) (sensitivity: 33.5%; specificity: 99.3%; positive predictive value [PPV]: 76.7%); identified variables were confirmed ILD diagnosis codes, Krebs von den Lungen-6/serum surfactant protein-D codes, age, and sex. This model showed similar performance in an external database (sensitivity: 19.8%; specificity: 99.4%; PPV: 65.5%), when a cutoff of 0.5 was used as a threshold to classify patients per their modeled probability of having ILD. Despite limited sensitivity, this validated algorithm's acceptable PPV may enable confident identification of true positive cases of ILD when suspected positive from claims data in Japanese patients with cancer, potentially making it valuable as a case-confirmation tool for retrospective studies using healthcare claims databases, particularly for those comparing relative risks between treatments.