Machine Learning-Based Radiopatho-Clinical Model Integrating Ultrasound Radiomics and Kleiner Score for Prognosis Prediction in NAFLD-Related Hepatocellular Carcinoma.
Authors
Affiliations (4)
Affiliations (4)
- Department of Hepatobiliary and Pancreatic Surgery, The Affiliated Hospital of Qingdao University, Qingdao, People's Republic of China.
- Qingdao Medical College, Qingdao University, Qingdao, Shandong, People's Republic of China.
- Department of Thoracic Surgery, The Affiliated Hospital of Qingdao University, Qingdao, People's Republic of China.
- Department of Abdominal Ultrasound, The Affiliated Hospital of Qingdao University, Qingdao, People's Republic of China.
Abstract
Nonalcoholic fatty liver disease (NAFLD) is an increasingly important etiologic factor in hepatocellular carcinoma (HCC), but the prognostic value of liver-background steatosis remains incompletely defined. We developed and internally validated an integrated machine learning model combining ultrasound radiomics, pathological steatosis grading, and clinicopathological variables for postoperative risk stratification in HCC. This retrospective study included 639 patients with HCC who underwent curative resection between 2010 and 2023. Radiomic features were extracted from preoperative ultrasound images, and a radiomics signature was generated using LASSO regression. Hepatic steatosis was graded using the Kleiner score, and clinicopathological variables were screened using the Boruta algorithm. A total of 101 machine learning models were developed and compared. Model performance was assessed using the concordance index, time-dependent area under the curve (AUC), Brier score, calibration, decision curve analysis, and SHapley Additive exPlanations. The random survival forest model showed the best overall performance for predicting overall survival (OS) and recurrence-free survival (RFS). In the validation cohort, the 1-, 3-, and 5-year AUCs were 0.863, 0.794, and 0.804 for OS, and 0.828, 0.811, and 0.823 for RFS, respectively. Brier scores remained below 0.20. Compared with BCLC and CNLC staging systems, the integrated model showed improved discrimination, calibration, and net clinical benefit. SHAP analysis indicated that microvascular invasion, tumor size, AFP, INR, Kleiner score, and Radscore contributed to individualized risk prediction. This integrated radiopatho-clinical machine learning model showed favorable internal performance for predicting OS and RFS after curative resection. Steatosis-related features may provide complementary prognostic information, supporting individualized postoperative surveillance, although external validation is required.