第一部分

 

建立肝细胞癌靶向联合免疫治疗全新疗效评价体系的研究

 

目的

通过整合实体瘤疗效评价标准1.1版（Response Evaluation Criteria in Solid Tumors 1.1, RECIST 1.1）与治疗后第6周甲胎蛋白（alpha-fetoprotein, AFP）应答水平，评估肝细胞癌（hepatocellular carcinoma, HCC）患者接受抗血管生成联合免疫治疗后的生存预后。

 

方法

本研究回顾性纳入150例（训练队列）和214例（验证队列）接受靶向联合免疫治疗的HCC患者。收集基线及第6周医学影像与AFP数据。AFP反应分层标准：部分缓解（partial response, PR）定义为AFP下降≥75%；疾病稳定（stable disease, SD）为AFP下降<75%且升高≤10%；疾病进展（progressive disease, PD）为AFP升高>10%。alpha-RECIST标准定义为：PR需满足RECIST 1.1-PR或AFP反应-PR；PD为AFP-PD或RECIST 1.1-PD且AFP反应PR；SD为既非PR也非PD。采用Kaplan-Meier曲线比较总生存期（overall survival, OS），通过一致性指数（concordance index, C-index）及时间依赖性受试者工作特征曲线下面积（AUC, area under the curve）评估不同标准的预测效能。

 

结果

RECIST 1.1标准在AFP < 20 ng/mL亚组中展现出显著的OS分层能力（P =0.020）。对于AFP ≥ 20 ng/mL患者，alpha-RECIST的预测效能（C-index=0.73）优于RECIST 1.1（0.66）、改良实体瘤疗效评价标准（modified Response Evaluation Criteria in Solid Tumors, mRECIST, 0.68）及单纯AFP应答（0.69）。国家癌症中心（National Cancer Center, NCC）策略在AFP < 20 ng/mL亚组采用RECIST 1.1标准，AFP ≥ 20 ng/mL亚组采用alpha-RECIST标准，其整体预测效能（C-index=0.73）显著优于其他标准（RECIST 1.1: 0.67；mRECIST: 0.69；AFP应答: 0.64）。验证队列中alpha-RECIST与NCC策略的C-index分别达0.77和0.74。

 

结论

在AFP ≥ 20 ng/mL及整体HCC患者中，alpha-RECIST标准与NCC策略相较于传统RECIST 1.1、mRECIST及AFP应答标准，展现出更优的生存分层能力与疗效预测效能。

 

第二部分

 

建立预测肝细胞癌靶向联合免疫治疗疗效的可解释性磁共振自动机器学习模型：一项多中心研究

 

目的

肝细胞癌（hepatocellular carcinoma, HCC）是全球癌症相关死亡的主要病因之一，抗血管生成联合免疫治疗已成为晚期 HCC 的一线方案，但疗效受肿瘤异质性限制。本研究旨在构建可解释性磁共振（magnetic resonance imaging, MRI）自动机器学习模型，预测 HCC 靶向联合免疫治疗的疗效，并解析其生物学机制。

 

方法

研究纳入 390 例接受抗血管生成（贝伐珠单抗/仑伐替尼）联合免疫治疗（抗 PD-1/PD-L1 单抗）的不可切除 HCC 患者，分为训练队列（n=188）、测试队列（n=47）、外部验证队列（n=81）及前瞻性验证队列（n=74）。基于 MRI 影像组学特征，采用 AutoGluon 框架构建HCC抗血管生成治疗与免疫治疗反应特征（Anti-angiogenic therapy and Immunotherapy Response Signature，HAIRS）模型，并结合传统机器学习（LASSO、SVM）进行对比。通过单因素/多因素 Logistic 回归及 Cox 回归分析模型预测效能及生存预后价值。利用 RNA 测序及单细胞 RNA 测序解析模型的生物学基础。

 

结果

HAIRS 模型在训练队列中 AUC 达 0.979（95% CI: 0.961-0.996），在测试队列、外部验证队列及前瞻性验证队列中 AUC 分别为 0.959、0.954 及 0.981，显著优于传统模型（P < 0.001）。HAIRS 亚型 1 组（低风险）的无进展生存期（progression free survival, PFS）及总生存期（overall survival, OS）均显著优于亚型 2 组（高风险）（P < 0.05）。多因素分析证实 HAIRS 模型是 PFS（风险比[hazard ratio, HR]=3.030, 95%置信区间[confidence interval, CI]: 1.915-4.792）及 OS（HR=2.610, 95% CI: 1.668-4.084）的独立预后因子。生物学机制表明，HAIRS 亚型 1 组肿瘤微环境（tumor microenvironment, TME）中 T 细胞受体信号通路及 VEGFA 通路显著激活，而亚型 2 组呈现免疫耗竭表型，CD8+ 效应记忆 T 细胞（CD8+ effector memory T Cells, CD8+ Tem）高表达且 EMT 通路富集。

 

结论

本研究首次构建基于 MRI 的可解释性自动机器学习模型 HAIRS，为 HCC 靶向联合免疫治疗的个体化决策提供了可靠的影像学工具，并揭示了TME免疫状态与治疗反应的潜在关联。

 

第三部分

 

阿帕替尼联合卡瑞丽珠单抗作为伴微血管侵犯的肝细胞癌术后辅助治疗的应用价值：一项多中心真实世界研究

 

目的

目前肝细胞癌（hepatocellular carcinoma, HCC）缺乏高级别证据支持的辅助治疗方案以降低术后复发风险。本研究旨在评估卡瑞利珠单抗联合阿帕替尼在合并微血管侵犯（microvascular invasion, MVI）的HCC患者术后辅助治疗中的安全性及疗效。

 

方法

回顾性纳入2019年10月至2022年6月期间于三家医疗中心接受根治性切除术且病理确诊MVI阳性的HCC患者。通过倾向性评分匹配（PSM）、时序检验、Cox回归分析及亚组分析评估辅助治疗与无复发生存期（recurrence-free survival, RFS）和总生存期（overall survival, OS）的关联，并报告≥3级治疗相关不良事件（treatment-related adverse events, TRAEs）。

 

结果

初始队列包含辅助治疗组111例和观察组276例，PSM后分别为99例和172例。辅助治疗组RFS显著延长（风险比[hazard ratio, HR]=0.52，95%置信区间[confidence interval, CI]：0.39-0.69，P < 0.001），但OS差异无统计学意义（HR=0.62，95%CI：0.39-0.99，P = 0.079）。PSM后结果一致。亚组分析显示卡瑞利珠单抗联合阿帕替尼在多数亚组中RFS获益显著。≥3级TRAEs发生率为20.7%，最常见为高血压（12.6%）和蛋白尿（9.0%）。

 

结论

卡瑞利珠单抗联合阿帕替尼作为MVI阳性HCC患者术后辅助治疗方案可显著改善RFS且安全性可控。

 

Part 1

 

Reclassification of therapeutic response of unresectable hepatocellular carcinoma to anti-angiogenic therapy and immunotherapy using alpha RECIST

 

Abstract

Objectives

To assess the therapeutic response of HCC to antiangiogenic therapy plus immunotherapy by integrating RECIST 1.1 and alpha-fetoprotein (AFP) response at the 6th week to predict overall survival (OS).

 

Methods

This retrospective study included 150 and 214 patients with HCC who received combination therapy in training and validation cohorts. The medical images and AFP levels obtained at baseline and 6th week were collected. AFP response stratification: partial response (PR): AFP% ≥ 75% decline; stable disease (SD): AFP% < 75% decline and  ≤ 10% elevation; progressive disease (PD): AFP% > 10% elevation. The alpha-RECIST was: PR: RECIST 1.1-PR or AFP-PR; PD: AFP-PD or RECIST 1.1-PD and does not satisfy AFP-PR; SD: neither PR nor PD. OS was compared using Kaplan–Meier curves. The predictive ability of various criteria was evaluated using the concordance index and time-dependent area under the receiver-operating characteristic curve.

 

Results

RECIST 1.1 achieved significant OS stratification (P = 0.020) for AFP < 20 ng/mL. For AFP ≥ 20 ng/mL, alpha-RECIST showed better performance than RECIST 1.1, mRECIST, and AFP response according to C-index (0.73 vs 0.66 vs 0.68 vs 0.69). The National Cancer Center (NCC) strategy utilized RECIST 1.1 for AFP < 20 ng/mL and alpha-RECIST for AFP ≥ 20 ng/mL and showed better performance than RECIST 1.1, mRECIST and AFP response according to C-index (0.73 vs 0.67 vs 0.69 vs 0.64). The performances of alpha-RECIST and NCC Strategy were confirmed in the validation cohort (C-index = 0.77 and 0.74).

 

Conclusions

The alpha-RECIST and NCC Strategy achieved better survival stratification in patients with HCC under combination therapy in the AFP ≥ 20 ng/mL group and the whole cohort compared to the RECIST 1.1, mRECIST, and AFP response.

 

Part 2

 

Development of an Explainable MRI-Based Automated Machine Learning Model for Predicting Therapeutic Response to Targeted-Immunotherapy Combination in Hepatocellular Carcinoma: A Multicenter Study

 

Abstract

 

Objectives

Hepatocellular carcinoma (HCC) is a leading cause of cancer-related mortality globally. Anti-angiogenic combined with immune checkpoint inhibitor (ICI) therapy has emerged as a first-line treatment for advanced HCC, but its efficacy is limited by tumor heterogeneity. This study aimed to develop an interpretable magnetic resonance imaging (MRI)-based automated machine learning (AutoML) model, termed HAIRS (HCC Anti-angiogenic and Immunotherapy Response Signature), to predict therapeutic response to targeted combination therapy and explore its biological mechanisms.

 

Methods

A total of 390 patients with unresectable HCC treated with anti-angiogenic agents (bevacizumab/lenvatinib) plus ICIs (anti-PD-1/PD-L1 mAbs) were enrolled, including training (n=188), testing (n=47), external validation (n=81), and prospective validation (n=74) cohorts. The HAIRS model was constructed using the AutoGluon framework based on MRI radiomic features and compared with traditional machine learning models (LASSO, SVM). Predictive performance and survival outcomes were evaluated via univariate/multivariate logistic regression and Cox regression. Bulk RNA sequencing and single-cell RNA sequencing were performed to dissect the biological basis of HAIRS.

 

Results

HAIRS demonstrated superior predictive efficacy with an AUC of 0.979 (95%CI: 0.961-0.996) in the training cohort, outperforming traditional models (P < 0.001). Validation cohorts achieved AUCs of 0.959, 0.954, and 0.981. Patients in HAIRS subtype 1 (low-risk) exhibited significantly prolonged progression-free survival (PFS) and overall survival (OS) compared with subtype 2 (high-risk) (P < 0.05). Multivariate analysis confirmed HAIRS as an independent prognostic factor for PFS (HR=3.030, 95%CI: 1.915-4.792) and OS (HR=2.610, 95%CI: 1.668-4.084). Mechanistically, HAIRS subtype 1 was associated with activated T-cell receptor and VEGFA signaling pathways, whereas subtype 2 showed immune exhaustion characterized by increased exhausted CD8+ effector memory T cells (CD8+ Tem) and epithelial-mesenchymal transition (EMT) enrichment.

 

Conclusions

This study establishes HAIRS as a novel interpretable MRI-based AutoML model, providing a non-invasive tool for personalized treatment decisions in HCC. The model also reveals distinct tumor microenvironment (TME) immune phenotypes underlying therapeutic responses to targeted combination therapy.

 

Part 3

 

Adjuvant camrelizumab plus apatinib in resected hepatocellular carcinoma with microvascular invasion: a multi-center real world study

 

Abstract

 

Objectives

Hepatocellular carcinoma (HCC) treatment currently lacks adjuvant therapy with a high level of supporting evidence to reduce recurrence after hepatectomy. This study aimed to assess the safety and efficacy of camrelizumab plus apatinib in the adjuvant therapy of patients with HCC with microvascular invasion (MVI).

 

Methods

Data were retrospectively collected on consecutive patients with HCC who underwent radical resection and were diagnosed with MVI-positive tumors between October 2019 and June 2022 at three centers. The association between adjuvant therapy and prognosis [recurrence-free survival (RFS), overall survival (OS)] was evaluated by propensity score matching (PSM), the log-rank test, Cox regression analysis, and subgroup analysis. Furthermore, grade 3 or 4 treatment-related adverse events (TRAEs) of adjuvant therapy were reported.

 

Results

Among the 111 patients in the adjuvant therapy group and 276 patients in the observation group at enrolment, there were 99 and 172 in the adjuvant therapy and observation groups after PSM, respectively. RFS was better in the adjuvant therapy group [hazard ratio (HR) 0.52; 95% confidence interval (CI): 0.39 to 0.69; P < 0.001], whereas OS was not (HR 0.62; 95% CI: 0.39 to 0.99; P = 0.079). These results were confirmed after PSM. Subgroup analyses were generally consistent in favour of adjuvant camrelizumab plus apatinib with better RFS. Grade 3 or 4 TRAEs accounted for 20.7% during adjuvant therapy; the most common TRAEs included hypertension and proteinuria.

 

Conclusions

Postoperative adjuvant camrelizumab plus apatinib significantly improved the RFS benefits with acceptable toxicities in patients with HCC with MVI.