Prognosis prediction and risk stratification of transarterial chemoembolization or intraarterial chemotherapy for unresectable hepatocellular carcinoma based on machine learning

Objective To develop and validate a risk scoring scale model (RSSM) for stratifying prognostic risk after intra-arterial therapies (IATs) for hepatocellular carcinoma (HCC). Methods Between February 2014 and October 2022, 2338 patients with HCC who underwent initial IATs were consecutively enrolled. These patients were divided into training datasets (TD, n = 1700), internal validation datasets (ITD, n = 428), and external validation datasets (ETD, n = 200). Five-years death was used to predict outcome. Thirty-four clinical information were input and five supervised machine learning (ML) algorithms, including eXtreme Gradient Boosting (XGBoost), Categorical Gradient Boosting (CatBoost), Gradient Boosting Decision Tree (GBDT), Light Gradient Boosting Machine (LGBT), and Random Forest (RF), were compared using the areas under the receiver operating characteristic (AUC) with DeLong test. The variables with top important ML scores were used to build the RSSM by stepwise Cox regression. Results The CatBoost model achieved the best discrimination when 12 top variables were input, with the AUC of 0.851 (95% confidence intervals (CI), 0.833–0.868) for TD, 0.817 (95%CI, 0.759–0.857) for ITD, and 0.791 (95%CI, 0.748–0.834) for ETD. The RSSM was developed based on the immune checkpoint inhibitors (ICI) (hazard ratios (HR), 0.678; 95%CI 0.549, 0.837), tyrosine kinase inhibitors (TKI) (HR, 0.702; 95%CI 0.605, 0.814), local therapy (HR, 0.104; 95%CI 0.014, 0.747), response to the first IAT (HR, 4.221; 95%CI 2.229, 7.994), tumor size (HR, 1.054; 95%CI 1.038, 1.070), and BCLC grade (HR, 2.375; 95%CI 1.950, 2.894). Kaplan–Meier analysis confirmed the role of RSSM in risk stratification (p < 0.001). Conclusions The RSSM can stratify accurately prognostic risk for HCC patients received IAT. On the basis, an online calculator permits easy implementation of this model. Clinical relevance statement The risk scoring scale model could be easily implemented for physicians to stratify risk and predict prognosis quickly and accurately, thereby serving as a more favorable tool to strengthen individualized intra-arterial therapies and management in patients with unresectable hepatocellular carcinoma. Key Points • The Categorical Gradient Boosting (CatBoost) algorithm achieved the optimal and robust predictive ability (AUC, 0.851 (95%CI, 0.833–0.868) in training datasets, 0.817 (95%CI, 0.759–0.857) in internal validation datasets, and 0.791 (95%CI, 0.748–0.834) in external validation datasets) for prediction of 5-years death of hepatocellular carcinoma (HCC) after intra-arterial therapies (IATs) among five machine learning models. • We used the SHapley Additive exPlanations algorithms to explain the CatBoost model so as to resolve the black boxes of machine learning principles. • A simpler restricted variable, risk scoring scale model (RSSM), derived by stepwise Cox regression for risk stratification after intra-arterial therapies for hepatocellular carcinoma , provides the potential forewarning to adopt combination strategies for high-risk patients. Supplementary Information The online version contains supplementary material available at 10.1007/s00330-024-10581-2.


Supplementary Tables
Table S1.The data source form multi-center hospitals Table S2.The details of the 38 variables in this study.Table S3.The parameters of the five ML Algorithms Table S4.Multivariable Regression Analysis for OS in training datasets Table S5.Five ML based models for prediction of 1-year, 2-year and 3-year OS in training datasets Table S6.Five ML based models for prediction of 1-year, 2-year and 3-year OS internal test datasets Table S7.Five ML based models for prediction of 1-year, 2-year and 3-year OS in external test datasets

Supplementary Information E1.1TACE or HAIC procedures
TACE or HAIC procedures have been described in our previous report [1][2][3].All equipment of IAT procedures included i) digital subtraction angiography (Philips, type FD 20 1250 mA); ii) the artery sheath catheter was inserted into the femoral artery using the modified Seldinger technique; iii) A 5-Fr Yashiro catheter (Terumo) was advanced into the celiac trunk and superior mesenteric artery to assess the feeding hepatic artery; iv) A 2.7-Fr micro-catheter (Terumo) was inserted in the feeding artery.
The therapeutic principles of IAT procedures were as follows: 1).TACE: the feeding artery was selected or super-selected whenever possible.Emulsion, which consisted of 10-20 ml lipiodol, 30-50 mg platinum drugs, and 20-40-mg epirubicin was injected slowly until the offending vessel occluded.If necessary, embolization using gel foam mixed with contrast medium was injected to reduce the residual blood flow until there was no longer any tumor staining after repeat angiography.2).HAIC: all chemo-drugs were given by HAIC through the micro-catheter.A modified FOLFOX6 regimen, including oxaliplatin (130 mg/m 2 infusion for 3 h on day 1), leucovorin (200 mg/m 2 for 3-5 hours on day 1), and Fluorouracil (400 mg/m 2 in bolus, and then 2,400 mg/m 2 continuous infusion 23-46 h) was applied.Treatment was repeated every 3 weeks and commonly 4-6 cycles unless intrahepatic lesions progress or toxicity became unacceptable.

E1.2 The molecular-targeted agents and immune checkpoint inhibitors protocol
During HAIC or TACE treatment, molecular-targeted agents (MTAs) and immune checkpoint inhibitors (ICIs) were performed to control the intrahepatic and extrahepatic progression.Oral first-line targeted chemotherapy including sorafenib and lenvatinib was started 1-5 days after the first HAIC or TACE session and continually administered.Once the disease progresses or 3-4 AEs occur, the second-line treatment regimen (regorafenib or apatinib) can be administered.Oral lenvatinib (Lenvimafi; Eisai Co., Ltd.) was administered to the patients with Ad-HCCs.The initial dose was determined based on the patient's body weight and liver function.Patients weighing > 60 kg with the Child-Pugh A classification started at a dose of 12 mg once daily.
Patients weighing < 60 kg with the same liver function started at a dose of 8 mg once daily.A reduction in dosage or interruption of treatment was implemented when AEs were detected.Lenvatinib was administered unless patients were intolerant of radiological tumor progression or AEs.ICI immunotherapy were performed after 1-3 days IAT treat and every 3 weeks intravenously.Fixed-dose administration of PD-1 was used until disease progression or unexpected toxicity.The dose and interval of TKIs allowed changes depending on toxicity and disease conditions.

E1.3 IAT conversion therapy protocol
The target tumors were down-staged to BCLC-A stage and the tumor burden was reduced met the Milan criterion after multi-cycle of IAT treatment, the HCC patients were conducted by surgical resection, imaging-guided thermal ablation or SBRT.Among them, resectable tumor was defined as the complete removal of all macroscopic tumor tissue and an expected remnant liver volume no less than 250 mL/ M.

E1.4 The inclusion and exclusion criteria
The inclusion criteria were as follows: (grade 1, 2, and 3 = ≤ -2.60, > -2.60 to -1.39, and > -1.39, respectively).For more detailed evaluations of patients with the middle grade of ALBI (grade 2), we used modified ALBI (mALBI) grading consisting of 4 levels, which included subgrading for the middle grade of 2 (2a and 2b) based on an ALBI score of -2.27 as the cut-off, which was previously reported as the value for indocyanine green retention after 15 min (ICG-R15) of 30% [4] ; (4) treatment parameters (IAT modalities, combination with TKI, combination with ICI, sequential local therapy and the response of first IAT).The responses to IAT was assessed by dynamic contrast enhanced CT or magnetic resonance imaging (MRI) based on modified Response Evaluation Criteria in Solid Tumor (mRECIST), including complete response (CR), partial response (PR), stable disease (SD), and progression disease (PD), which was performed every 4-6 weeks after initial IAT and evaluated independently by two radiologists (reader 1, L.Z.L., and reader 2, J. Z., with 10 years of experience) who were blinded to IAT procedures at the time of data collection.

Supplementary Figures
Figure S1.The importance ranking of 12 variables in the CatBoost model using the SHAP algorithm.

Figure S2 .
Figure S2.shows the influence of the response to first IAT on 5-years death in the CatBoost model through the SHAP algorithm.

Figure S3 .
Figure S3.shows the influence of the ICI on 5-years death in the CatBoost model through the SHAP algorithm.

Figure S4 .
Figure S4.shows the influence of the tumor size on 5-years death in the CatBoost model through the SHAP algorithm.

Figure S5 .
Figure S5.shows the influence of the BCLC stages on 5-years death in the CatBoost model through the SHAP algorithm.

Figure S6 .
Figure S6.shows the influence of the local therapy on 5-years death in the CatBoost model through the SHAP algorithm.

Figure S7 .
Figure S7.shows the influence of the PLT on 5-years death in the CatBoost model through the SHAP algorithm.

Figure S8 .
Figure S8.shows the influence of the PT on 5-years death in the CatBoost model through the SHAP algorithm.

Figure S9 .
Figure S9.shows the influence of the INR on 5-years death in the CatBoost model through the SHAP algorithm.

Figure S10 .
Figure S10.shows the influence of the CRP on 5-years death in the CatBoost model through the SHAP algorithm.

Figure S11 .
Figure S11.shows the influence of the TKI on 5-years death in the CatBoost model through the SHAP algorithm.

Figure S12 .
Figure S12.shows the influence of the AST on 5-years death in the CatBoost model through the SHAP algorithm.

Figure S13 .
Figure S13.shows the influence of the Cre on 5-years death in the CatBoost model through the SHAP algorithm.

Figure S14 .
Figure S14.A web tool for prediction OS of HCC patients after IAT.

Figure S2 .
Figure S2.shows the influence of the response to first IAT on 5-years death in the CatBoost model through the SHAP algorithm.

Figure S5 .
Figure S5.shows the influence of the BCLC stages on 5-years death in the CatBoost model through the SHAP algorithm.

Figure S6 .
Figure S6.shows the influence of the local therapy on 5-years death in the CatBoost model through the SHAP algorithm.

Figure S7 .
Figure S7.shows the influence of the PLT on 5-years death in the CatBoost model through the SHAP algorithm.

Figure S8 .
Figure S8.shows the influence of the PT on 5-years death in the CatBoost model through the SHAP algorithm.

Figure S9 .
Figure S9.shows the influence of the INR on 5-years death in the CatBoost model through the SHAP algorithm.

Figure S10 .
Figure S10.shows the influence of the CRP on 5-years death in the CatBoost model through the SHAP algorithm.

Figure S11 .
Figure S11.shows the influence of the TKI on 5-years death in the CatBoost model through the SHAP algorithm.

Figure S12 .
Figure S12.shows the influence of the AST on 5-years death in the CatBoost model through the

Figure S13 .
Figure S13.shows the influence of the Cre on 5-years death in the CatBoost model through the SHAP algorithm.

Figure S14 .
Figure S14.A web tool for prediction OS of HCC patients after IAT.

Table S1 .
The data source form multi-center hospitals

Table S2 .
The details of the 38 variables in this study.

Table S3 .
The parameters of the ML Algorithms