Early recurrence prediction and risk stratification of hepatocellular carcinoma after transarterial chemoembolization achieving radiological complete response based on contrast-enhanced CT machine learning

Objectives To develop and validate machine learning (ML) models using clinical and contrast-enhanced CT (CECT) parameters to assess recurrence risk in hepatocellular carcinoma (HCC) after transarterial chemoembolization (TACE) achieving radiological complete response (CR). Methods 122 HCC patients who underwent TACE and achieved radiological CR from two centers were divided into the development (n = 100) and external validation dataset (n = 22). Recurrence free survival (RFS) was tracked, and patients were categorized into early recurrence (ER) and non-ER groups based on a 1-year cutoff. Forty clinical and CECT parameters were collected and screened. Six ML models were constructed and compared using the area under the curve (AUC) and decision curve analysis (DCA). Key parameters were used to construct a Cox regression nomogram and stratify recurrence risk using log-rank test. Results The extreme gradient boosting (XGBoost) model demonstrated the best predictive performance based on 13 parameters, with AUCs of 0.913 and 0.812 for the internal and external validation datasets. SHapley Additive exPlanations (SHAP) analysis identified the top 10 parameters. The Cox regression nomogram was constructed with ECV, complete capsule, FIB-4 index, tumor size, platelet-to-neutrophil ratio, and delayed phase tumor CT value. Log-rank test demonstrated significant risk stratification between the two datasets (both p  < 0.01). Conclusion The XGBoost-based ER prediction model identifies 1-year recurrence following TACE with radiological CR. The Cox regression nomogram enables risk stratification, dividing patients into three subgroups.