Individualised and Time-Dependent Prediction of Recurrence After Colorectal Cancer Liver Metastases Resection

Introduction: Treatment strategies for colorectal cancer liver metastases (CRLM) are increasingly complex, driving interest in advanced prognostic tools. Externally validated benchmarks based on established clinical risk scores are needed to contextualise emerging models. However, existing scores remain difficult to apply and compare, because they provide categorical risk strata rather than individualised probabilities, and because model parameters are rarely published. Here, we validate and recalibrate seven established risk scores to derive individualised recurrence probabilities for future score benchmarking. Method: Patients undergoing index CRLM resection from two tertiary centres were included: Karolinska University Hospital (retrospective KaroLiver cohort [2013-2021, follow-up to 2025, n=709]) and Oslo University Hospital (OSLO-COMET trial cohort [2012-2016, follow-up to 2022, n=198]). Performance of the GAME, m-CS, Fong, Nordlinger, Nagashima, Konopke and Helsinki scores was assessed using recurrence-free survival as the primary outcome. Results: Among 709 patients in KaroLiver (median age 67 years [IQR 59-74]; 434 men [61%]), 422 (60%) had synchronous metastases and 444 (63%) received neoadjuvant chemotherapy. All scores significantly stratified recurrence risk (log-rank P<0.001; time-dependent AUC 0.63-0.74 at 3 years), albeit with performance differences across time points. Recalibrated models showed good agreement between predicted and observed recurrence and were externally validated in OSLO-COMET with broadly comparable net benefit. Models were implemented in the online tool PLATO (Personalised Liver metastases risk AssessmenT in Oncology). Conclusion: Established CRLM risk scores can be translated into individualised probabilistic recurrence estimates at clinically relevant time points, supporting patient communication, risk-adapted surveillance and transparent benchmarking of next-generation prognostic models.