LASSO-based nomogram to predict early treatment failure in advanced hepatocellular carcinoma treated with SBRT, lenvatinib, and tislelizumab

Background Stereotactic body radiotherapy (SBRT) combined with lenvatinib and tislelizumab has shown potential activity in advanced hepatocellular carcinoma (HCC), but early treatment failure (ETF) remains common. We developed an exploratory model to estimate ETF, defined as radiologically confirmed progressive disease (PD) or all-cause death within 6 months after treatment initiation. Methods This single-center retrospective study included advanced HCC patients treated with the triple regimen from June 2021 to June 2024. Baseline clinical, imaging, and hematological variables were analyzed. LASSO regression with 10-fold cross-validation was used for feature selection. Restricted cubic splines assessed linearity, and λ.1se was examined as a sensitivity analysis. A multivariable logistic regression model was presented as a nomogram. Discrimination, calibration, and decision curve analysis (DCA) were evaluated, with 1000-resample bootstrap internal validation repeating the modeling pipeline. Results Among 118 patients, 33 (27.9%) experienced ETF; 28 had radiologically confirmed PD and 5 died without prior documented PD. At λ.min, LASSO selected maximum tumor diameter, Vp4 portal vein tumor thrombosis (PVTT), baseline systemic immune-inflammation index (SII), ALBI score, and log10-transformed alpha-fetoprotein (AFP). Under λ.1se, four predictors were retained, with SII excluded. The final model had an apparent ROC-AUC of 0.903 and an optimism-corrected ROC-AUC of 0.866. Calibration showed acceptable agreement between predicted and observed risks. DCA suggested net benefit over an exploratory threshold range of approximately 0.10-0.50, which should not be interpreted as a validated treatment-switching threshold. Conclusion The nomogram showed promising internal discrimination for estimating 6-month ETF risk in advanced HCC patients receiving SBRT, lenvatinib, and tislelizumab. Given the modest sample size and single-center design, the model remains exploratory and hypothesis-generating. External multicenter validation is required before clinical implementation or use for treatment selection.