Identification and validation of an explainable machine learning model for hepatocellular carcinoma at high risk: a
Shufang Hao1, Xiaoyu Liang2, Xue Li2
1Department of Interventional Therapy, Guangdong Provincial Hospital of Chinese Medicine and Guangdong Provincial Academy of Chinese Medical Sciences, Guangzhou, Guangdong, P. R. China.
Background:
Currently, there is still a lack of noninvasive, automated, and accurate machine learning (ML) model for prognostic risk stratification of high-risk hepatocellular carcinoma (hHCC) after transarterial chemoembolization (TACE).
Purpose:
To develop and test an ML-based model that integrates various clinical variables for prediction of early recurrence (ER, defined as recurrence within 24 months after treatment) after TACE for hHCC.
Methods:
From May 2008 to October 2022, consecutive patients with unresectable HCC who underwent initial TACE were reviewed at multicenter. We input 39 preoperative and 5 treatment-related variables into five ML-based models for ER prediction. Furthermore, we compared their performances using area under the receiver operating characteristic curve (AUC) with DeLong test between these ML models and the major staging systems including Barcelona Clinic Liver Cancer, China National Liver Cancer, American Joint Committee on Cancer, Japan Society of Hepatology, and Hong Kong Liver Cancer. Overall survival (OS) and recurrence-free survival (RFS) were analyzed using the Kaplan-Meier method with log-rank tests.
Results:
A total of 2068 eligible patients with hHCC were categorized into the training set ( n = 1220), the internal test set ( n = 306), and the external test set ( n = 542). ER rate was found in 78.4%, 78.8 %, and 62.5% of patients in three sets. Among all ML models, the eXtreme gradient boosting (XGBoost) model using top seven variables yielded the highest AUC of 0.888 in the internal test set and 0.854 in the external test set. Furthermore, it yielded better discriminatory ability for RFS with higher time-dependent AUC than all major staging systems (DeLong test, all P < 0.001). Kaplan-Meier analysis revealed significant OS and RFS risk stratification using the XGBoost model in three sets, respectively.
Conclusion:
The XGBoost model demonstrated superior predictive performance compared with other four ML models and all clinical staging systems in identifying ER, which may support individualized TACE management by guiding therapy selection.
