Document resource
Background The standard of care first-line treatment for advanced biliary tract cancer (BTC) has been updated to immune checkpoint inhibitors combined with chemotherapy. However, immuno-chemotherapy for BTC still demonstrates significant population heterogeneity, requiring continuous prognostic monitoring and dynamic treatment modifications tailored to individual therapeutic responses. We aimed to develop an individualized AI-driven system to dynamically refine survival predictions for advanced BTC patients through updated multidimensional clinical data, enabling real-time adaptive prognostic stratification.Methods This multicenter study analyzed data from advanced BTC patients treated with immuno-chemotherapy across eight Chinese centers. Patients from four centers were randomized into development (70%) and internal validation (30%) cohorts, with two institutions as external validation cohorts. Prospective validation included five centers (original two and three new). We developed a Bayesian joint modeling framework to integrate patients‘ baseline clinical/imaging data with serial laboratory measurements (C-reactive protein [CRP], Carbohydrate antigen 19-9 [CA19-9], and total bilirubin [TBil]), establishing an Individualized Dynamic Risk Estimation Model (iDREM-BTC) for real-time survival prediction., and an advanced version (iDREM(pro)-BTC) added immunohistochemistry/genetic data. Model performance was evaluated via AUC and calibration. The study protocol is detailed in figure 1, and the study was registered on ClinicalTrials.gov (NCT06849193).Results 2314 patients were included (841 in the development cohort, 360 in the internal validation cohort and 1113 in the external validation cohorts). Machine learning and multivariate Cox regression analyses identified age, tumor burden, biliary obstruction severity, ascites degree, and levels of CRP, CA19-9, and TBil as independent predictors of mortality in advanced BTC patients receiving immuno-chemotherapy. The iDREM-BTC model demonstrated robust performance in the development cohort (overall dynamic AUC 0.730, 95% CI 0.689-0.794), showing progressive improvement in AUC values with each subsequent measurement (increasing from 0.688 to 0.798). This predictive capability was consistently validated across multiple cohorts: internal validation (AUC 0.718, 95% CI 0.670-0.778), external validation cohort1 (0.755, 0.707-0.808), external validation cohort2 (0.705, 0.639-0.773), and prospective external validation (0.745, 0.691-0.802). Furthermore, the iDREM(pro)-BTC model achieved better discrimination in the development cohort (AUC 0.807, 95% CI 0.781-0.839), with AUC values steadily improving from 0.780 to 0.839 across repeated measurements. External validation confirmed these findings (AUC 0.718, 0.669-0.787), showing comparable progression from 0.666 to 0.791.Conclusions The Individualized dynamic design of the model brings it closer to clinical practice and the needs of dynamic precision medicine, enabling real-time prediction of prognosis for advanced BTC patients during immuno-chemotherapy. To enhance practicality, two alternative models were designed for different clinical scenarios.Trial Registration The trial is registered at www.clinicaltrials.gov ( NCT06849193).Ethics Approval This study involving human subjects was approved by the Clinical Research Ethics Committee of Zhongda Hospital Affiliated to Southeast University (Ethics Approval ID: 2025ZDSYLL065-P01). Informed consent was waived for the retrospective development and validation cohorts, while all participants in the prospective external validation cohort provided written consent. The research adhered strictly to the Declaration of Helsinki principles and followed the STROBE guidelines.Abstract 626 Figure 1Study protocol