Multimodality-Integrated Deep Survival Analysis of Hepatocellular Carcinoma After Transarterial Chemoembolization: A Two-Center Collaborative Study

Background & Aims: There is high heterogeneity in clinical outcome for patients with hepatocellular carcinoma (HCC) following transarterial chemoembolization (TACE) and the accurate survival prediction remains challenging. We aimed to develop and validate a deep survival approach for prognostic prediction in HCC with TACE. Methods: We built the survival models by integrating clinical variables, radiological identitifications, and an entire-volumetric radiomics signature in 200 training datasets using a series of survival analysis algorithms such as Cox extreme gradient boosting (Cox-XGBoost), Cox gradient boosting machine (Cox-GBM), deep Cox-Time, and stepwise Cox proportional hazard (Cox-PH) regression. Models were validated in two independent cohorts of 86 and 84 patients from two institutions, respectively. Concordance index (C-index) was used as the model performance metric. Results: A radiomics survival signature (Rad-S), constructed from 4,840 imaging features by Random survival forest and least absolute shrinkage and selection operator (LASSO) Cox analysis, demonstrated favorable prognostic performance with C-indexes of 0.610 to 0.663 in training and validation sets, respectively. Rad-S, modified Response Evaluation Criteria in Solid Tumors (mRECIST), and neutrophil to lymphocyte ratio (NLR) were selected as top variables. Among all survival analysis algorithms and existing prognostic systems, the Cox-GBM model achieved the best prediction performance, with C-indexes of 0.687 to 0.763 in training and validation cohort, respectively; it also gave two risk strata with significant different clinical outcome. Conclusions: The newly designed Rad-S can be a favorable prognostic predictor of HCC with TACE. The Cox-GBM model integrating multimodal data can enhance the prognostic prediction, thus help personalized clinical management. Funding: None to declare. Declaration of Interest: None to declare. Ethical Approval: Ethics committee approval was granted by the local institutional ethics review board (protocol 2018-SR-025), and the requirement of written informed consent was waived. All procedures involving human participants were performed in accordance with the 1975 Helsinki declaration and its later amendments.