Development of a machine learning model to predict overall survival for large hepatocellular carcinoma at BCLC stage A or B after curative hepatectomy

Tai-Xin Yang^1,2Jia-Yong Su^1,2Min-Jun Li^1,2Shen Shuang^1,2Yu Wang²Huan- Nan Wei²Ming-Jian Huang²Qing-Man Qin²You-Yin Ran²Yao-Ting Huang²Jin-Yan Huang²Jie Zhang^1,2*Bang-De Xiang^1,3,4*Wen-Feng Gong^1*

• ¹Guangxi Medical University Cancer Hospital, Nanning, China
• ²Guangxi Medical University, Nanning, China
• ³Key Laboratory of Early Prevention and Treatment for Regional High Frequency Tumors (Guangxi Medical University), Ministry of Education, Nanning, China
• ⁴Guangxi Key Laboratory of Early Prevention and Treatment for Regional High Frequency Tumors, Nanning, China

Introduction: Patients with large hepatocellular carcinoma (LHCC) have a poor prognosis even after curative hepatectomy. This study aimed to develop and validate an interpretable machine learning (ML) model to predict their overall survival (OS). Methods: This study included 2,565 patients with hepatocellular carcinoma (HCC) who underwent curative hepatectomy between January 2014 and December 2021. The LHCC patients were randomly assigned (7:3 ratio) to a training (n=1069) or validation (n=457) group. Independent risk factors for OS were identified using multivariable Cox regression. Eight ML models were developed and compared. The optimal model's interpretability was assessed using Shapley Additive Explanations (SHAP). Results: LHCC patients experienced a considerable reduction in OS (Hazard Ratio, HR: 1.810, 95% Confidence Interval, CI: 1.585-2.068) compared to SHCC patients. Among eight ML models, the gradient boosting machine (GBM) model demonstrated superior performance. In the validation group, the GBM model achieved area under the receiver operating characteristic curve (AUC) values of 0.742, 0.744, and 0.750 for 1-, 3-, and 5-year OS, respectively. These results were comparable with or superior to established postoperative predictive models. The GBM model showed the ability to stratify patients with LHCC into distinct prognostic groups. A web-based calculator was developed for risk score generation. Notably, the GBM model showed enhanced predictive accuracy in patients with a high neutrophil-lymphocyte ratio (C-index: 0.819). Conclusions: The GBM-based model demonstrated the potential to predict prognosis for patients with LHCC after curative hepatectomy. This interpretable model may assist in personalized risk assessment and tailoring postoperative management strategies.