Preoperative prediction of microvascular invasion in hepatocellular carcinoma using routine clinical variables: development and temporal validation of an interpretable random forest model in a single-center retrospective cohort study
Year: 2026
Type: article
Study design: Observational Study
Abstract: Background: Microvascular invasion (MVI) is a postoperative pathological feature associated with recurrence and poor prognosis in hepatocellular carcinoma (HCC), but it cannot be directly determined before surgery. Existing MVI models are often limited by insufficient validation, uncertain transportability, and emphasis on discrimination without adequate assessment of calibration or clinical utility. This study aimed to develop and temporally validate a machine learning model based on routinely available preoperative variables in patients with HCC to estimate the probability of MVI before surgery. Methods: This retrospective study included patients with pathologically confirmed HCC who underwent surgery at The First Affiliated Hospital, Jiangxi Medical College, Nanchang University, between December 2018 and December 2024. Of the 756 patients screened, 95 (12.6%) were excluded because of missing or incomplete key data, leaving 661 patients for analysis. Eligible patients had documented MVI status, complete preoperative data, and no preoperative antitumor treatment. MVI was defined as cancer cell nests within small endothelial-lined vessels that were detectable only microscopically. Patients were chronologically assigned to a development cohort (December 2018 to December 2021; n=344) and an internal temporal validation cohort (January 2022 to December 2024; n=317). Candidate predictors were initially screened using univariable logistic regression (LR), and variables with P<0.05 were carried forward. These variables were subsequently evaluated using least absolute shrinkage and selection operator (LASSO) LR and Boruta, and final predictors were defined as those retained by LASSO at λ_min and confirmed by Boruta. Eight machine-learning models were compared. For each model, the classification threshold was derived from cross-validated predictions in the development cohort using the maximum Youden index and then fixed for temporal validation. The final model was evaluated for discrimination, calibration, clinical net benefit, and interpretability. Results: A total of 315 patients (47.7%) had MVI. The median age was 56 years (interquartile range, 49-65 years); 581 patients (87.9%) were men, 347 (52.5%) had cirrhosis, and 596 (90.2%) had solitary tumors. Six predictors were selected: tumor number, tumor size, platelet count (PLT), lactate dehydrogenase (LDH), alpha-fetoprotein (AFP) (log1p-transformed), and cancer antigen 125 (CA125) (log1p-transformed). For the random forest (RF) model, the area under the receiver operating characteristic curve (AUC) was 0.865 [95% confidence interval (CI): 0.827-0.903] in the development cohort and 0.857 (95% CI: 0.816-0.897) in the temporal validation cohort. In the temporal validation cohort, accuracy, sensitivity, specificity, F1 score, and Brier score were 0.748, 0.757, 0.740, 0.720, and 0.155, respectively. Calibration showed acceptable agreement, decision curve analysis (DCA) indicated positive net benefit, and Shapley additive explanations (SHAP) identified AFP (log1p) as the most important predictor. Conclusions: The RF model showed promising discrimination but moderate threshold-dependent classification performance in same-center temporal validation. Multicenter external validation is required before the model can be considered for clinical testing or implementation.
Language: English
Subfield: Hepatology
Field: Medicine
Domain: Health Sciences
Open Access status: diamond
OpenAlex ID: W7224131814