An automated machine-learning model for prognostic risk stratification of intermediate-stage hepatocellular carcinoma...

Currently, there is still a lack of noninvasive, automated, and accurate machine-learning (ML) models for prognostic risk stratification of intermediate-stage hepatocellular carcinoma (HCC) after transarterial chemoembolization (TACE).

We aimed to develop an ML model for prognostic risk stratification of intermediate-stage HCC after TACE to assist physicians in decision-making.

Between April 2008 and October 2022, consecutive patients with intermediate-stage HCC undergoing initial conventional TACE were retrospectively enrolled from seven tertiary hospitals. A system utilizing natural language processing technology was used to extract clinical information from electronic medical records to develop the ML models. The primary outcomes were 2-year HCC-related death and cancer-related survival (CRS, defined as the interval from initial TACE to either HCC-related death or last follow-up). The ML models' performance and their comparison with various biomarkers were assessed.

A total of 4426 eligible patients were included (3906 males, 520 females; median age, 54 years ± 11 [standard deviation]; 2667 in the training cohort, 667 in the internal test cohort, and 1092 patients in the external test cohort). Six ML models were developed, with the XGBoost model demonstrating the best predictive performance. It achieved an AUC of 0.842 (95% CI, 0.827-0.857) in the training cohort, 0.815 (95% CI, 0.783-0.847) in the internal test cohort, and 0.798 (95% CI, 0.771-0.824) in the external test cohort. Among high-risk patients stratified by the XGBoost model, those who received TACE combined with microwave ablation had significantly higher cumulative CRS rates than those treated with TACE alone.

We developed a noninvasive, automated, and accurate ML model, the XGBoost model, with robust performance in prognostic risk stratification for intermediate-stage HCC following TACE.