Purpose: The aim of this study is to create and validate an actionable interpretable survival model of hepatocellular carcinoma patients receiving Transarterial chemoembolization, by combining clinical data and CT radiomic data under a SHAP framework.
Materials and Methods: In the analysis, the publicly available WAW-TACE cohort (N = 233; 170 deaths, cadaveric data access group) was used. Of the 3,339 radiomics features, the 100 with the highest variance were selected. Following the preprocessing and one-hot encoding (features that did not converge were assigned non-significant p-values) and based on univariate Cox regression analysis, the 30 top- ranked predictors were selected. To perform the survival model, a ridge-penalized Cox proportional hazards with an α = 20, and 5-fold cross validation, was used to fit 70% of the data and the remaining 30% was used to evaluate the model. The performance of the model was evaluated using the Harrell's C-index and a time dependent AUC (Uno et al., 2013). 12-month calibration was performed along with an optimism correction (200 bootstrap iterations). Interpretability of the survival model was done using SHAP (Kernel Explainer).
Results: The test C-index was 0.654 (95% CI: 0.592–0.698). The internal bootstrap corrected C-index, reflecting a correction for optimism at a sample out of the distribution, was 0.728. The time- dependent AUC was noted at 0.874 (6 months), 0.803 (12 months), 0.663 (24 months), and 0.634 (36 months). The top predictors included serum albumin, number of lesions, and being female. The trained model had a higher C-index compared to the Hepatocellular carcinoma, Albumin, and Prothrombin Time-based Assessment of Liver Function and Risk of Death (HAP) score (C-index: 0.544) and the Assessment of Liver Function and Risk of Death (ALBI-TAE) (C-index: 0.624). Risk separation was significant and improved using the median risk threshold. The SHAP analysis presented plausible feature interactions and provided exogenous explanations.
Conclusion: The development of an explainable Cox-ridge model has established significant short-term survival prediction after TACE for HCC. The SHAP-based explanations provided for the model suggest that it can be incorporated into clinical decision support pathways with early clinical triage to facilitate access to systemic therapy.