Natural History of Fibrotic Interstitial Lung Disease using AI-driven Test-free Assessment of Routine EHR
Discuss this preprint
Start a discussion What are Sciety discussions?Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
Rationale
Fibrosing interstitial lung diseases (ILDs), including idiopathic pulmonary fibrosis (IPF), have heterogeneous postdiagnosis courses. Prognostic tools often rely on pulmonary function testing, imaging, or laboratory data that may not be uniformly available, and rarely provide individualized, time-updated forecasts of multiple trajectory events.
Objectives
Can longitudinal healthcare claims generate test-free, time-updated forecasts of clinically actionable postdiagnosis trajectory events in patients with fibrosing ILD and IPF?
Methods
Using de-identified longitudinal administrative claims from the Merative™ MarketScan ® Commercial Claims and Encounters and Medicare Supplemental and Coordination of Benefits databases, we constructed code-based digital twins (ZeBRA) encoding each patient’s evolving diagnosis, pharmacy, and procedure history. Horizon-specific models forecast seven claims-observable events: supplemental oxygen escalation, pulmonary hypertension, acute respiratory failure/ARDS composite, nausea, diarrhea, liver injury, and gastrointestinal bleeding. The analytic cohort included 345,918 patients with fibrosing ILD, including 17,284 with IPF. Predictions were evaluated in a time-updated follow-up setting over 1-month, 6-month, and 1-year horizons.
Results
Predictive discrimination was consistent across events and horizons. In fibrosing ILD, AUC ranged from 0.691 for liver injury at 1 year to 0.912 for oxygen dependence at 1 month, with PPV ranging from 0.189 to 0.714. At 1 month, oxygen dependence achieved AUC 0.912 ± 0.005 with PPV 0.473 ± 0.005, and pulmonary hypertension achieved AUC 0.881 ± 0.005 with PPV 0.539 ± 0.005. The IPF subcohort showed analogous horizon-dependent performance, with AUC ranging from 0.687 to 0.855 and PPV from 0.245 to 0.817. At 1 month in IPF, PPV was 0.753 ± 0.015 for oxygen dependence and 0.817 ± 0.011 for pulmonary hypertension.
Conclusions
A test-free digital-twin framework 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. These forecasts may support low-burden reassessment, anticipatory care planning, and earlier recognition of elevated near-term risk for respiratory deterioration or management-altering complications.
