A multi-institutional study to predict the benefits of DEB-TACE and molecular targeted agent sequential therapy in unresectable hepatocellular carcinoma using a radiological-clinical nomogram

Abstract

Objective

Exploring the efficacy of a Radiological-Clinical (Rad-Clinical) model in predicting prognosis of unresectable hepatocellular carcinoma (HCC) patients after drug eluting beads transcatheter arterial chemoembolization (DEB-TACE) to optimize the targeted sequential treatment.

Methods

In this retrospective analysis, we included 202 patients with unresectable HCC who received DEB-TACE treatment in 17 institutions from June 2018 to December 2022. Progression-free survival (PFS)-related radiomics features were computationally extracted from HCC patients to build a radiological signature (Rad-signature) model with least absolute shrinkage and selection operator regression. A Rad-Clinical model for postoperative PFS was further constructed according to the Rad-signature and clinical variables by Cox regression analysis. It was presented as a nomogram and evaluated by receiver operating characteristic curves, calibration curves, and decision curve analysis. And further evaluate the application value of Rad-Clinical model in clinical stages and targeted sequential therapy of HCC.

Results

Tumor size, Barcelona Clinic Liver Cancer (BCLC) stage, and radiomics score (Rad-score) were found to be independent risk factors for PFS after DEB-TACE treatment for unresectable HCC, with the Rad-Clinical model being the greatest predictor of PFS in these patients (hazard ratio: 2.08; 95% confidence interval: 1.56–2.78; P < 0.001) along with high 6 months, 12 months, 18 months, and 24 months area under the curves of 0.857, 0.810, 0.843, and 0.838, respectively. In addition, compared to the radiomics and clinical nomograms, the Radiological-Clinical nomogram also significantly improved the classification accuracy for PFS outcomes, based on the net reclassification improvement (45.2%, 95% CI 0.260–0.632, p < 0.05) and integrated discrimination improvement (14.9%, 95% CI 0.064–0.281, p < 0.05). Based on this model, low-risk patients had higher PFS than high-risk patients in BCLC-B and C stages (P = 0.021). Targeted sequential therapy for patients with high and low-risk HCC in BCLC-B stage exhibited significant benefits (P = 0.018, P = 0.012), but patients with high-risk HCC in BCLC-C stage did not benefit much (P = 0.052).

Conclusion

The Rad-Clinical model may be favorable for predicting PFS in patients with unresectable HCC treated with DEB-TACE and for identifying patients who may benefit from targeted sequential therapy.

Appendix E

DEB-TACE

After digital subtraction angiography (DSA) was performed to determine the tumor blood vessels, a 2.7-F micro-catheter (Cook, USA) was applied to inject 100 to 300um (or 300 to 500um) CalliSpheres drug elating hemispheres (Jiangsu Hengrui Medicine, Jiangshu, China) loaded with 80 mg pirarubicin for tumor embolization, and the main blood supply artery was embolized with a mixture of gelatin sponge particles and contrast agent until the blood supply was completely blocked. We used the modified Response Evaluation Criteria in Solid Tumors (mRECIST) standard to evaluate the changes in the lesions. If there are no active lesions at 1 month after DEB-TACE, enhanced CT imaging will be reviewed and judged every 3 months. If there are active lesions, DEB-TACE treatment will be repeated. The numbers of TACE treatment in this study was 1–5.

CT protocols of the seventeen institutions

All patients underwent unenhanced and triphasic dynamic contrast-enhanced CT scans before DEB-TACE. Enrolled patients at the 17 institutions had similar scan setup: Using the bolus-tracking method, the arterial phase was scanned at 15–20 s after 100-Houndsfeld unit attenuation of the abdominal aorta. The portal venous and delayed phases were obtained at 70 and 180 s, respectively, after contrast injection. Technical details of CT protocols at the seventeen institutions are outlined in Table S1.

Intra-observer and inter-observer agreement

The intra-observer agreement and inter-observer agreement of feature extraction were evaluated by correlation coefficients (ICCs). To compute the intra-observer ICC, 50 CT images were selected randomly and segmented twice in 1 month (At least 10 days apart) by reader A. To compute the inter-observer ICC, the selected images were segmented by two radiologists independently (reader A and reader B). Segmentation was performed to further obtain independent feature extraction to compute the intra-observer and inter-observer ICCs. When the ICC was greater than 0.75, it was considered good agreement, and the remaining segmentation was performed by reader A.

Features extraction and rad-signature model establishment

Feature screening unifies the scales of different features through the Z-score normalization method, so that the antiquated feature scales are distributed between 0 and 1. Firstly, pair-wise correlation analysis was performed to remove redundant radiological features, by using the “find Correlation” function in R package “caret” with the absolute correlation cutoff set at 0.9. Then, we employed the least absolute shrinkage and selection operator (LASSO) Cox regression, a qualified approach for regression of high-dimensional predictors by a penalty to shrink some regression coefficients to exactly zero. This approach selected the most predictive radiomics features from the training cohort. The penalty parameter (lambda) was determined by using a fivefold cross-validation based on minimum error criteria. Selected features were weighed by their respective coefficients obtained from LASSO, and were used to compute a radiomics signature (Rad-signature) with a linear combination. Rad-signature model was constructed by Cox regression.

The R packages used for statistical analysis

We used the “glmnet” package to perform the LASSO-Cox regression. The “rms” package was used for multivariable Cox regression analysis, nomogram construction, and calibration. The “DynNom” package was used to build dynamic nomograms on the web. The R function cox.zph was employed to test the proportional hazards assumption for a Cox regression model fit. The Cindex was calculated and compared using function concordance.index and cindex.comp in the “survcomp” package. Prediction error curves were generated using “pec” package. Decision curve analysis was performed using the “stdca.R” package. We used the “timeROC” package to do the time-dependent ROC curve analysis.