Fibrosing interstitial lung diseases (ILDs), including idiopathic pulmonary fibrosis (IPF), exhibit heterogeneous clinical courses after diagnosis. Existing prognostic tools commonly require imaging, pulmonary function testing, laboratory data, or clinician-entered notes, which may be unavailable or time-lagged in routine practice. The authors sought to determine whether longitudinal administrative claims could support a test-free, time-updated forecasting approach for multiple clinically actionable postdiagnosis trajectory events in patients with fibrosing ILD and IPF.
The study used de-identified longitudinal administrative claims from the Merative MarketScan Commercial Claims and Encounters and Medicare Supplemental and Coordination of Benefits databases. From these data the investigators constructed code-based digital twins, termed ZeBRA, which encode each patient's evolving diagnosis, pharmacy, and procedure history into a model-ready representation. Horizon-specific predictive models were trained on these encoded histories to produce individualized risk forecasts without reliance on imaging, pulmonary function tests, laboratory results, clinical notes, or direct patient data collection.
The modeling framework produced forecasts for seven claims-observable trajectory events chosen to reflect clinically relevant respiratory deterioration or management-altering complications. These were:
Predictions were evaluated in a time-updated follow-up setting at three horizons: 1 month, 6 months, and 1 year, enabling both near-term and longer-term risk assessment.
The analytic cohort included 345,918 patients identified with fibrosing ILD from the MarketScan databases. A subcohort of 17,284 patients met criteria for IPF. All analyses reported use of de-identified administrative claims; the dataset and cohort selection details beyond aggregate counts were not reproduced in the summary beyond what the source reported.
Model discrimination and predictive value were evaluated using area under the receiver operating characteristic curve (AUC) and positive predictive value (PPV) across events and horizons. Performance was reported with uncertainty estimates for key results. In the overall fibrosing ILD cohort, discrimination varied by outcome and horizon:
Specific 1-month results in the fibrosing ILD cohort included:
These results indicate stronger near-term discrimination for respiratory outcomes such as oxygen dependence and pulmonary hypertension compared with some non-respiratory complications.
The IPF subcohort demonstrated analogous, horizon-dependent predictive performance with somewhat different ranges:
At the 1-month horizon within the IPF group, reported PPVs were:
These higher PPVs for near-term respiratory events in the IPF subcohort suggest particularly strong positive predictive performance for short-horizon forecasts in that subgroup.
The authors conclude that a test-free digital-twin framework derived from routine longitudinal claims can provide individualized, time-updated forecasts of actionable fibrosing ILD and IPF trajectory events without imaging, pulmonary function tests, laboratory data, clinical notes, or patient-facing data collection. They propose that such forecasts could support low-burden reassessment, anticipatory care planning, and earlier recognition of elevated near-term risk for respiratory deterioration or management-altering complications.
The source reports no competing interests declared by the authors.
A functioning API implementation of the ZeBRA pulmonary hypertension prediction pipeline, including model input processing and risk-score generation, is available at the authors' GitHub repository: https://github.com/zeroknowledgediscovery/ZeBRA_nhfild. Source claims data used for model training and validation consist of de-identified MarketScan administrative claims and are subject to data-use agreements; these data cannot be publicly released. The authors indicate that derived summary statistics and scripts sufficient to reproduce main analyses are available upon reasonable request.
The University of Kentucky institutional review board approved use of the de-identified MarketScan database for UK researchers under the reported protocol, and the authors state that all relevant ethical guidelines and reporting checklists were followed. No additional limitations or external validation results beyond those summarized in the source were reported in the provided text.