Dynamic survival prediction using longitudinal AGR in unresectable locally advanced esophageal squamous cell...

Immune checkpoint inhibitor (ICI)-based induction therapy followed by definitive radiotherapy has emerged as a promising treatment strategy for patients with unresectable locally advanced esophageal squamous cell carcinoma (LA-ESCC). However, many patients still experience disease progression and poor clinical outcomes. Although longitudinal circulating biomarkers may provide valuable prognostic information, their dynamic predictive value in this treatment setting remains largely unexplored.

We retrospectively enrolled patients with unresectable LA-ESCC from a single medical center who received PD-1inhibitor-based induction therapy followed by definitive radiotherapy. In addition to baseline clinical and histopathological characteristics, 110 circulating biomarkers were serially collected throughout treatment to characterize longitudinal changes. A Bayesian joint model integrating baseline predictors with longitudinal biomarker trajectories was developed for dynamic survival prediction. Model discrimination and calibration were evaluated using time-dependent area under the receiver operating characteristic curve (AUCs) and Brier scores, respectively.

A total of 376 patients were included, comprising a training cohort (n = 221) and an independent temporal test cohort (n = 155), yielding 328,104 longitudinal biomarker measurements. Following multistage variable selection procedure, latent class growth analysis identified distinct longitudinal trajectories of the albumin-to-globulin ratio (AGR) that are significantly associated with survival outcomes. Dynamic AGR value and trajectory emerged as the most informative longitudinal predictors among all candidate biomarkers. In the training cohort, the 12-, 24-, and 36-month AUCs were 0.85 (95% CI, 0.82-0.89), 0.92 (95% CI, 0.90-0.94), and 0.87 (95% CI, 0.84-0.91), respectively. In the test cohort, the corresponding 12- and 24-month AUCs were 0.80 (95% CI, 0.75-0.85) and 0.82 (95% CI, 0.78-0.86), respectively.

By integrating longitudinal AGR trajectories with baseline clinical information, the Bayesian joint model enables individualized, real-time prognostic updating and may facilitate dynamic risk stratification and risk-adapted management for patients with unresectable LA-ESCC.