Abstract
IMPORTANCE: Current prognostic models for colorectal liver metastases (CRLM) primarily incorporate clinicopathologic features assessed at a single time point, resulting in a static risk assessment for individuals. Given that tumor progression is a dynamic process, especially for patients with CRLM, and patients' data are continuously collected during the follow-up visits, dynamic prediction is a natural model for risk assessments via reflecting the latest prognosis, whenever new marker measurements are available. OBJECTIVE: To develop CRLM prognostic models and a clinical web-based tool to facilitate dynamic predictions. DESIGN, SETTING, AND PARTICIPANTS: In this retrospective prognostic study, patients with CRLM who underwent resection between January 2014 and January 2021, were included in the training and validation cohorts. Clinicopathologic characteristics and preoperative and postoperative laboratory measurements taken within 12 months after surgery across 9 laboratory markers (carcinoembryonic antigen, carbohydrate antigen 19-9, γ-glutamyl transferase, red blood cell distribution width SD and coefficient of variance, aspartate aminotransferase to platelet ratio index, Fibrous-4 index, S-index, and neutrophil-to-lymphocyte ratio) were collected. Three prediction models for progression-free survival (PFS) and overall survival (OS) based on a functional random survival forest framework were constructed and compared: model A incorporated only clinicopathologic characteristics, model B included clinicopathologic characteristics and preoperative laboratory markers, and model C integrated clinicopathologic characteristics along with longitudinal laboratory markers. Data were analyzed from June 2024 to June 2025. EXPOSURE: Resection in patients with CRLM. MAIN OUTCOMES AND MEASURES: Performance metrics included area under the receiver operating characteristic curve (AUC) and Brier score (BS). RESULTS: A total of 976 patients (median [IQR] age, 59 [51-65] years; 612 [62.7%] male) were eligible for this study, with 758 patients in the training cohort (median [IQR] age, 59 [52-66] years; 487 [64.2%] male) and 218 patients in the validation cohort (median [IQR] age, 58 [49-64] years; 125 [57.3%] male).The training cohort included a total of 24 992 longitudinal measurements, and the external validation cohort included 7198 longitudinal measurements. In the external validation cohort, model C demonstrated an improved prognostic capability compared with models A and B, with AUC values of 0.796 (95% CI, 0.740-0.848) for 1-year progression-free survival (PFS), 0.837 (95% CI, 0.768-0.899) for 3-year PFS, and 0.850 (95% CI, 0.780-0.914) for 5-year PFS. The corresponding BSs were 0.246 (95% CI, 0.236-0.261) for 1 year, 0.205 (95% CI, 0.193-0.218) for 3 years, and 0.142 (95% CI, 0.132-0.153) for 5 years. Model C consistently outperformed models A and B for overall survival (OS) prognosis, with AUCs of 0.849 (95% CI, 0.768-0.914) for 1 year, 0.741 (95% CI, 0.667-0.815) for 3 years, and 0.753 (95% CI: 0.656-0.849) for 5 years, alongside BS values of 0.047 (95% CI, 0.045-0.048) for 1 year, 0.178 (95% CI, 0.168-0.195) for 3 years, and 0.144 (95% CI, 0.133-0.165) for 5 years. Additionally, dynamic individualized risk profiles for PFS and OS were developed for patients. A web-based tool was created to facilitate the practical application of these dynamic prediction models for new patients in clinical environments. CONCLUSIONS AND RELEVANCE: In this retrospective prognostic study, the dynamic models, along with the web-based tool for personalized prediction, demonstrated improved performance by incorporating multiple longitudinal makers.