Machine Learning Model Integrating Conventional Ultrasound Radiomics and Clinicopathological Features for Early Recurrence Risk Prediction After Curative-Intent Hepatectomy in Hepatocellular Carcinoma.

Purpose: To develop and compare machine learning models integrating conventional ultrasound radiomic features with clinicopathological variables for predicting one-year postoperative recurrence in patients with hepatocellular carcinoma after curative-intent hepatectomy. Methods: This retrospective, single-center study included 154 patients with pathologically confirmed hepatocellular carcinoma who underwent curative-intent hepatectomy between January 2020 and December 2024. Patients were divided into a training cohort (n = 124) and a validation cohort (n = 30). Clinicopathological data and preoperative conventional ultrasound images were collected. Radiomic features were extracted and selected using least absolute shrinkage and selection operator (LASSO) regression. Predictive models were developed and compared using conventional logistic regression and four machine learning algorithms, including a generalized linear model (GLM), random forest (RF), support vector machine (SVM), and gradient boosting machine (GBM), based on different combinations of clinicopathological variables and radiomic features. Model performance was evaluated in the validation cohort. Results: Multivariable logistic regression identified Edmondson grade, satellite nodules, fibrinogen, and white blood cell count as factors independently associated with postoperative recurrence. Among the evaluated models, the RF model combining clinicopathological factors with the Radscore achieved the highest numerical validation area under the receiver operating characteristic curve (AUC) of 0.717 (95% confidence interval: 0.523-0.910), with an accuracy of 70.0%, a sensitivity of 42.9%, and a specificity of 93.8%. Conclusion: The RF model integrating clinicopathological factors with the Radscore showed moderate discrimination for one-year postoperative recurrence and may hold promise for postoperative risk assessment. Given the limited validation sample size and uncertainty in model performance, these findings should be considered exploratory and require further validation in larger, independent, multicenter cohorts before clinical implementation.