Advanced Models for Liver Fibrosis Assessment

Noninvasive tests (NITs) such as ELF, FIB-4, vibration-controlled transient elastography (VCTE), and magnetic resonance elastography (MRE) are increasingly used to assess liver fibrosis in both clinical practice and therapeutic trials. Despite widespread adoption, their diagnostic performance—particularly in longitudinal settings—remains difficult to evaluate rigorously. Conventional metrics such as sensitivity, specificity, and AUROC implicitly assume a perfect, static gold standard. In liver fibrosis research, however, the historical reference standard (liver biopsy) is invasive, infrequently measured, subject to sampling and observer error, and often unavailable at follow-up time points.

This project develops a longitudinal structural equation modeling (SEM) framework that redefines how diagnostic performance is evaluated when no perfect gold standard exists, and disease severity evolves over time. Rather than treating observed fibrosis measures as truth, the approach models fibrosis severity as a latent, unobserved disease process, inferred jointly from multiple imperfect indicators—including serum biomarkers, imaging-based stiffness measures, and biopsy proxies where available.

The proposed framework integrates measurement models (linking observed tests to latent fibrosis severity) with longitudinal structural models that characterize disease trajectories, treatment effects, and individual heterogeneity. By explicitly separating signal from measurement error, the models enable estimation of time-dependent diagnostic properties such as responsiveness to change, reliability, and comparative information content across biomarkers—quantities that are not identifiable using standard cross-sectional or repeated-measures approaches.

A distinguishing feature of this work is its emphasis on dynamic inference. Latent growth and latent change score models are used to quantify progression or regression of fibrosis over time, while accommodating irregular follow-up, missing data, and informative dropout common in real-world cohorts and clinical trials. Extensions incorporate joint modeling of clinical events, allowing latent fibrosis trajectories to be linked to downstream outcomes such as decompensation, transplantation, or mortality.

The methodology is designed to be platform-agnostic and multimodal, supporting integration of blood-based biomarkers, elastography, imaging, and clinical endpoints within a unified statistical framework. While motivated by liver fibrosis, the modeling strategy is intentionally generalizable to other disease areas where diagnostics are surrogate-based, longitudinal, and imperfectly observed. This project addresses a critical methodological gap in modern diagnostic research: how to rigorously evaluate and compare diagnostic tools when “ground truth” is latent, evolving, and only partially observed. The resulting framework has direct implications for trial design, biomarker qualification, and regulatory evaluation of noninvasive diagnostics, while also advancing the statistical foundations of diagnostic analytics.