Identifying hepatocellular carcinoma patients at risk of early non-response after first-cycle transarterial chemoembolization: A reproducible machine learning study.

PURPOSE: Transarterial chemoembolization (TACE) is standard therapy for intermediate-stage hepatocellular carcinoma (HCC), but early non-response after the first cycle remains common and unpredictable. Existing clinical scores use limited variables and fail to capture heterogeneity influencing treatment response. This study developed reproducible machine learning (ML) models to predict early non-response after first-cycle TACE using an open-access dataset. METHODS: We analyzed 233 patients from the publicly available WAW-TACE dataset, which includes pre-TACE CT imaging data and 33 harmonized clinical variables. Early non-response at first follow-up was defined by mRECIST criteria as stable or progressive disease versus complete or partial response. Logistic regression, random forest, XGBoost, and support vector machine models were trained using five-fold cross-validation and evaluated on an independent test set using accuracy, F1-score, and AUC-ROC. RESULTS: Cross-validation F1-scores ranged from 0.78 to 0.83. On the independent test set, logistic regression achieved the most balanced performance (accuracy 0.73, AUC 0.76), followed by random forest (0.70, 0.73), SVM (0.66, 0.75), and XGBoost (0.66, 0.73). SVM demonstrated the highest sensitivity (recall 0.93), while XGBoost achieved perfect recall (1.00) with reduced precision, consistent with overclassification of the positive class. Feature importance analyses identified clinically relevant predictors, including bilirubin, the Six-and-Twelve score, lesion diameter, disease etiology, and BCLC stage. CONCLUSION: Reproducible ML models trained on open TACE data demonstrate modest performance for predicting early non-response. These findings provide a transparent benchmark for future model development and validation in interventional oncology and suggest that baseline clinical variables alone may have limited predictive utility.