Construction of a Preoperative Prediction Model for TACE Resistance in Primary Hepatocellular Carcinoma Based on Machine Learning Algorithms.

Purpose: Transcatheter arterial chemoembolization (TACE) resistance compromises prognosis in unresectable hepatocellular carcinoma (HCC). This study aimed to develop an interpretable prediction model using machine learning (ML) and Shapley Additive Explanations (SHAP) for preoperative assessment of TACE resistance. Patients and Methods: A single-center retrospective analysis included 562 HCC patients who received ≥3 TACE sessions (2013-2024). Multi-modal features (blood routine, coagulation, biochemistry, imaging) were integrated. Seven ML models (LR, RF, DT, XGBoost, LightGBM, SVM, ANN) were constructed. Feature selection used univariate Logistic regression and Lasso regression. Model performance was evaluated via AUC, F1 score, and accuracy; SHAP analyzed feature importance. Results: Data were split into training (n=394, 70%) and validation (n=168, 30%) sets. Seven core predictors (NLR, tumor capsule integrity, AFP, etc.) were identified. XGBoost outperformed other models, with AUCs of 0.942 (95% CI: 0.919-0.966) and 0.898 (95% CI: 0.853-0.944) in training and validation sets, respectively, and an F1 score of 0.741. SHAP revealed NLR (mean Shapley value=0.13) and tumor capsule absence (0.08) as the strongest predictors. Conclusion: This interpretable ML model efficiently predicts TACE resistance using multi-modal data, with AUC>0.8. It offers a preoperative tool to identify high-risk patients, optimize treatment strategies, and holds significant clinical translational value.