A Validated, explainable machine learning-based preoperative risk model for microvascular invasion in hepatocellular carcinoma.
LX, Z., JH, C., CR, Z., TC, L., TM, G., KQ, X., S, D., C, C., BY, W., & H, D. (2026). A Validated, explainable machine learning-based preoperative risk model for microvascular invasion in hepatocellular carcinoma.. Langenbeck's archives of surgery. https://doi.org/10.1007/s00423-026-04143-x
LX Z, JH C, CR Z, TC L, TM G, KQ X, et al. A Validated, explainable machine learning-based preoperative risk model for microvascular invasion in hepatocellular carcinoma.. Langenbeck's archives of surgery. 2026; doi: 10.1007/s00423-026-04143-x
LX Z, JH C, CR Z, et al. A Validated, explainable machine learning-based preoperative risk model for microvascular invasion in hepatocellular carcinoma.[J]. Langenbeck's archives of surgery. 2026. DOI: 10.1007/s00423-026-04143-x.
@article{lx2026,
author = {Zhou LX and Cai JH and Zhu CR and Luo TC and Gao TM and Xiao KQ and Ding S and Chen C and Wan BY and Dong H},
title = {A Validated, explainable machine learning-based preoperative risk model for microvascular invasion in hepatocellular carcinoma.},
journal = {Langenbeck's archives of surgery},
year = {2026},
doi = {10.1007/s00423-026-04143-x},
note = {PMID: 42663705},
}
TY - JOUR AU - Zhou LX AU - Cai JH AU - Zhu CR AU - Luo TC AU - Gao TM AU - Xiao KQ AU - Ding S AU - Chen C AU - Wan BY AU - Dong H TI - A Validated, explainable machine learning-based preoperative risk model for microvascular invasion in hepatocellular carcinoma. T2 - Langenbeck's archives of surgery PY - 2026 DO - 10.1007/s00423-026-04143-x AN - PMID:42663705 ER -
PURPOSE: This investigation aimed to develop and validate a diagnostic algorithm for preoperatively assessing the likelihood of microvascular invasion (MVI) in patients with hepatocellular carcinoma (HCC). METHODS: Clinical and pathological information of patients with HCC who underwent curative resection was collected from two medical centers. Data from Nanjing Drum Tower Hospital were randomly split into training (80%) and internal validation (20%) cohorts, while data from Northern Jiangsu People's Hospital were employed as an independent external validation cohort. Feature engineering was performed using recursive feature elimination (RFE) within the training cohort. Various machine learning models were applied, and their performance was evaluated through diverse metrics, such as receiver operating characteristic (ROC) curves. Additionally, the Shapley Additive Explanations (SHAP) method, together with tumor differentiation, Ki-67, and other relevant markers, were employed to enhance model interpretability and reliability. RESULTS: Among the 1106 patients enrolled, 315 were pathologically confirmed to have MVI. RFE identified tumor diameter, alpha-fetoprotein (AFP), gamma-glutamyl transferase (GGT), and pan-immune-inflammation value (PIV) as key determinants of MVI risk in patients with HCC. With these variables, the XGBoost model reached area under the curve (AUC) values of 0.893 in the training cohort, 0.845 in the internal validation cohort, and 0.793 in the external validation cohort. Correlation analyses revealed that the risk score showed significant associations with tumor differentiation and the expression of Ki-67 proliferation index, Glypican-3 (GPC3), Cytokeratin 19 (CK19), and Vascular endothelial growth factor receptor 2 (VEGFR2). CONCLUSION: An XGBoost model incorporating tumor diameter, AFP, GGT, and PIV exhibited robust performance in assessing preoperative MVI among HCC patients.