Machine learning shows modest accuracy predicting early HCC non-response to first-cycle chemoembolization

Machine-learning models using baseline clinical variables showed modest performance (AUC 0.76) for predicting early non-response after first-cycle TACE in intermediate-stage HCC, suggesting these variables alone have limited predictive utility.

• Retrospective analysis of 233 patients from the public WAW-TACE dataset; logistic regression achieved the best test-set accuracy of 0.73 and AUC of 0.76.
• XGBoost achieved perfect sensitivity (1.00) at the cost of low precision, indicating a tendency to overcall non-response.
• Key limitation: single-center, retrospective data lacking external and prospective validation.

RadPigeon summaries are original and for information only. They are not clinical advice.