Predictive modeling based on radiomics and hematological parameters for recurrence-free survival after curative resection in patients with hepatocellular carcinoma.

Abstract

e16280 Background: Early recurrence following curative resection represents a significant challenge in the management of hepatocellular carcinoma (HCC). Identifying patients at high risk of recurrence is crucial, enabling timely implementation of appropriate interventions postoperatively. This study aims to develop and validate a radiomics-based predictive model to forecast recurrence-free survival (RFS) in patients undergoing curative resection for HCC. Methods: We retrospectively analyzed data from 184 early-stage HCC patients who underwent curative resection. These patients were randomly assigned to a training cohort (n = 128) and a validation cohort (n = 56) in a 7:3 ratio. Liver tumors were delineated as regions of interest (ROI) using 3D-Slicer software. Radiomic features were extracted with Pyradiomics. Feature selection was performed using the interclass correlation coefficient (ICC) and least absolute shrinkage and selection operator (LASSO) regression. The most predictive features for RFS and their corresponding weights were identified and combined to calculate the Rad-score. Both the Rad-score and clinical variables, including biochemical parameters, were incorporated into univariate and multivariate analyses to construct a Cox proportional hazards model. A radiomics-based nomogram was subsequently developed to predict recurrence risk by integrating multiple influencing factors. The model’s performance was evaluated using the area under the curve (AUC), Harrell’s concordance index (C-index), and calibration curves. Results: The Rad-score for each patient was derived from 15 radiomic features. Multivariate analysis revealed that Rad-score, lactate dehydrogenase (LDH), and alpha-fetoprotein (AFP) were independent predictors of RFS. These factors effectively stratified patients into distinct recurrence risk groups, with RFS being significantly prolonged in the low-risk group in both the training cohort (p < 0.001) and validation cohort (p < 0.001). The Rad-score-based composite prediction model demonstrated robust predictive performance, with AUC of 0.756, 0.729, and 0.765 for 1, 2, and 3 years RFS in the training set, and 0.796, 0.798, and 0.920 for the same time points in the validation set. The C-index for the two cohorts was 0.718 and 0.787, respectively. Calibration curves further confirmed the favorable predictive accuracy of the nomogram model. Conclusions: This postoperative predictive model enhances the ability to identify patients at high risk of recurrence, serving as a valuable tool to inform clinical treatment decisions.