Prognostic Models and Predictors of Recurrence in Resectable Colorectal Liver Metastases: A Systematic Review and Meta-analysis.
TL;DR
It was showed that current prognostic models incompletely capture the biological heterogeneity underlying CRLM recurrence, and dynamic biomarkers and treatment-response indicators may support future multimodal prognostic frameworks and improve risk stratification following curative-intent treatment.
Abstract
We systematically reviewed prognostic models for recurrence after curative-intent locoregional treatment of colorectal liver metastases (CRLM) and quantitatively synthesized prognostic factors associated with recurrence-free survival (RFS). From 2,208 records across 26 years of literature, 293 studies were included, encompassing 85,150 patients in development cohorts and 5,549 patients in validation cohorts. Most models targeted risk stratification rather than clinically actionable prediction, and Cox proportional hazards regression remained the dominant modelling approach despite increasing use of artificial intelligence and machine learning terminology. External validation was uncommon (33/293 studies), performance reporting was heterogeneous, and overall methodological quality was limited, restricting cross-study comparability and clinical translation. Imaging-based modelling represented a small subset (19 studies; 23 RFS models), generally characterized by small cohorts and predominantly conventional regression-based pipelines. Meta-analysis identified several clinicopathological, molecular, and treatment-related predictors consistently associated with recurrence, including primary lymph node positivity (HR 1.58), multiple liver metastases (HR 1.49), positive resection margins (HR 1.73), postoperative carcinoembryonic antigen >5ng/mL (HR 2.79), poor response to neoadjuvant therapy according to RECIST criteria (HR 2.76), and postoperative circulating tumor DNA positivity (HR 4.78), the strongest prognostic factor identified. Adjuvant and peri-operative systemic therapies were associated with lower recurrence risk. The study showed that current prognostic models incompletely capture the biological heterogeneity underlying CRLM recurrence. Dynamic biomarkers and treatment-response indicators may support future multimodal prognostic frameworks and improve risk stratification following curative-intent treatment.