免疫检查点抑制剂联合化疗用于早期或局部晚期三阴性乳腺癌：一项随机试验的系统评价和贝叶斯网络荟萃分析。
全文（开放获取） · 中文译文
摘要 背景：免疫检查点抑制剂（ICIs）联合化疗已成为早期和局部晚期三阴性乳腺癌（TNBC）治疗的基石。然而，关于不同PD-1/PD-L1药物及治疗策略——尤其是仅新辅助与围手术期“全程”方案——相对疗效和安全性的直接比较证据仍然有限。为指导临床决策，我们进行了一项贝叶斯网络荟萃分析（NMA），比较六种具体基于ICI的方案和四种总体治疗策略。 方法：我们系统检索了PubMed、Embase、Cochrane Library和Web of Science，从数据库建库至2026年2月25日。符合条件的随机对照试验（RCTs）纳入既往未经治疗的早期或局部晚期TNBC患者，并比较ICI联合新辅助化疗与单纯化疗。主要终点为病理完全缓解（pCR）；次要终点包括总生存期（OS）和≥3级不良事件（AEs ≥3）。我们使用Cochrane RoB 2工具评估偏倚风险。采用贝叶斯随机效应模型进行NMA；效应以比值比（ORs）或风险比（HRs）及95%可信区间（CrIs）报告。治疗排名源自累积排名曲线下面积（SUCRA）值。证据确定性使用GRADE框架评估，并辅以CINeMA工具。 结果：共8项随机对照试验（RCT）涉及2，073例患者符合纳入标准，涵盖四种治疗策略类别和六种含免疫检查点抑制剂（ICI）的方案。在病理学完全缓解（pCR）方面，PD-1（新辅助）+化疗（OR 2.56，95% CrI 1.55-4.22）、PD-L1（围手术期）+化疗（OR 2.00，95% CrI 1.29-3.09）以及PD-1（围手术期）+化疗（OR 1.71，95% CrI 1.39-2.10）与单纯化疗相比均显著提高了pCR率，但证据确定性范围为中等至极低。PD-L1（新辅助）+化疗显示出pCR升高的数值趋势（OR 1.41，95% CrI 0.98-2.01），但未达到统计学显著性。在总生存期（OS）方面，PD-L1（新辅助）+化疗（HR 0.24，95% CrI 0.08-0.72）和PD-1（围手术期）+化疗（HR 0.63，95% CrI 0.48-0.82）均显示出相较于化疗的显著生存获益。在敏感性分析中，度伐利尤单抗（新辅助）+化疗和帕博利珠单抗（围手术期）+化疗产生的HR值相同（分别为0.24和0.63）。安全性特征存在差异：卡瑞利珠单抗（围手术期）+化疗（OR 0.60，95% CrI 0.34-1.04）和帕博利珠单抗（新辅助）+化疗（OR 0.53，95% CrI 0.28-1.01）提示≥3级不良事件（AEs）风险降低，但这些发现未达到统计学显著性，且支持证据的确定性为低至极低。 结论：ICI 联合化疗持续提高早期和局部晚期 TNBC 的 pCR 率。值得注意的是，尽管我们的分析提示新辅助 PD-L1 阻断和围手术期 PD-1 抑制可能带来生存获益，但这些信号仅来自三项研究，仍属于产生假设而非改变实践。有必要开展未来的头对头 RCT 以证实这些观察结果，并更好地界定长期安全性特征。 系统综述注册：https://www.crd.york.ac.uk/prospero/，注册号 CRD420261385223。 引言 乳腺癌仍然是全球女性中最常诊断的恶性肿瘤，并且在癌症相关死亡原因中位列第五。根据GLOBOCAN 2022年的估算，每年约有230万新发病例——占所有新发癌症的11.6%——全球约有67万人死于该病()。除了对患者及其家庭造成深远的生理和心理负担外，该疾病还通过高昂的治疗成本和劳动力生产力损失带来沉重的社会经济负担。三阴性乳腺癌（TNBC），定义为缺乏雌激素受体、孕激素受体和人表皮生长因子受体2表达，占所有乳腺癌的15–20%()。以显著的侵袭性、早期复发和有限的靶向治疗选择为特征，TNBC在乳腺癌亚型中预后最差()。数十年来，基于蒽环类和紫杉类的新辅助化疗一直是局部晚期或可手术TNBC患者的标准治疗。然而，病理学完全缓解（pCR）率仍不理想，通常在30%至40%之间()。未能达到pCR是长期预后不良的强预测因素，包括无事件生存期和无病生存期()。因此，克服传统化疗的疗效上限以提高pCR率——并且至关重要的是，将这些获益转化为生存优势——是早期和局部晚期TNBC的核心临床挑战。 靶向程序性死亡-1（PD-1）/程序性死亡配体1（PD-L1）轴的免疫检查点抑制剂（ICIs）的出现，重塑了实体瘤的治疗格局。TNBC特有的较高肿瘤突变负荷和密集的肿瘤浸润淋巴细胞群体，为免疫治疗提供了令人信服的生物学依据（）。在转移性背景下，随机试验已表明，PD-1/PD-L1抑制联合化疗可显著延长无进展生存期和总生存期，为向更早期疾病阶段的转化铺平了道路（, ）。里程碑式的III期试验——最值得注意的是KEYNOTE-522和IMpassion031——确立了在新辅助化疗中加入ICIs可使病理完全缓解（pCR）率绝对提高10至20个百分点，并改善无事件生存期。这些结果定义了高危早期TNBC的新标准（, ）。 尽管取得了这些进展，仍存在若干未解决的问题。首先，多种具有不同分子靶点的ICIs——抗PD-1与抗PD-L1药物——现已可用于临床，但它们在疗效或安全性方面是否存在有意义的差异仍不清楚()。其次，ICI给药的最佳持续时间和时机仍引发争论。以KEYNOTE-522为代表的“围手术期、全程”策略旨在通过延长免疫暴露来根除微小残留病灶，但引起了关于治疗持续时间延长、累积毒性和费用的担忧。相反，在GeparNuevo等试验中研究的“仅新辅助”方法提供了较短的治疗疗程，但其长期生存结局是否不劣于围手术期模式仍未得到证实()。关键在于，尚无头对头的随机对照试验（RCTs）直接比较这些药物或安排策略。 鉴于这种复杂性，传统的两两荟萃分析无法同时对所有干预措施进行排序。为填补这一证据空白，我们进行了一项贝叶斯网络荟萃分析（NMA），综合了所有可得的随机对照试验（RCT）数据。我们的目标是比较特定免疫检查点抑制剂（ICIs）（如帕博利珠单抗、阿替利珠单抗、卡瑞利珠单抗）及总体治疗策略（围手术期治疗对比单纯新辅助治疗）在关键终点——包括病理完全缓解（pCR）、总生存期和≥3级不良事件——方面的疗效与安全性。通过生成概率性治疗排序，本分析旨在提供高等级证据，为临床决策及未来指南制定提供参考。 材料与方法 本网络荟萃分析（NMA）依据系统综述与荟萃分析优先报告条目中关于网络荟萃分析的扩展声明（PRISMA-NMA；）（）实施。为确保透明度并减少结局报告偏倚，研究方案已在PROSPERO平台前瞻性注册（注册号：CRD420261385223）。 数据来源与检索策略 我们进行了一项系统性文献检索，检索范围涵盖PubMed、EMBASE、Cochrane图书馆、Web of Science、ClinicalTrials.gov以及世界卫生组织国际临床试验注册平台（ICTRP）。灰色文献通过手工检索过去五年主要肿瘤学会议（SABCS、ASCO、ESMO）的摘要，以及制药公司网站和相关学位论文进行补充。检索策略结合了医学主题词（MeSH）与与三阴性乳腺癌相关的自由文本词（如“Triple-Negative Breast Neoplasms”“ER-Negative PR-Negative HER2-Negative Breast Cancer”）、新辅助治疗（如“Neoadjuvant Therapies”）以及免疫检查点抑制剂（如“Checkpoint Inhibitors, Immune”“pembrolizumab”“atezolizumab”）（）。记录检索时间自数据库建库起至2026年2月25日，无语言限制。 纳入标准 纳入标准如下：（1）纳入经组织学确诊、既往未接受治疗、可手术的早期或局部晚期三阴性乳腺癌（TNBC）成人患者的随机对照试验（RCT）；（2）干预组接受PD-1或PD-L1抑制剂联合标准新辅助化疗——包括“仅新辅助”和“围手术期（全程）”两种策略；（3）对照组仅接受标准新辅助化疗；（4）可获得以下至少一项结局指标——病理学完全缓解（pCR）、总生存期（OS）或≥3级不良事件（AEs ≥3）。pCR定义为根据当地病理学评估，乳腺和腋窝均无浸润性癌残留。OS定义为从随机化至任何原因导致死亡的时间。AEs ≥3依照美国国家癌症研究所不良事件通用术语标准（CTCAE）进行分级。 排除标准为：（1）报道重叠患者人群或重复发表的RCT；（2）结局数据不完整或报告不充分的研究；（3）非随机研究、综述或病例报告。 去除重复文献后，两名评审员独立根据标题和摘要筛选所有记录。随后评估全文文章的符合性。筛选或数据提取过程中的分歧通过共识或咨询第三位资深研究者解决。对于每项纳入的试验，我们确保提取最新且最全面的数据集。 数据提取与质量评估 采用一份经过标准化设计和预试验的调查表，两名研究者独立地从每篇符合纳入标准的文献中提取数据。提取的变量包括试验名称、NCT标识符、第一作者、发表年份、试验分期、样本量、中位年龄、干预方案与对照方案，以及肿瘤分期。对于疗效和安全性结局，我们提取了病理完全缓解（pCR）和≥3级不良事件（AEs）的比值比（ORs）及其95%可信区间（CrIs），以及总生存期（OS）的风险比（HRs）及其95%可信区间（CrIs）。对于诸如I-SPY2等多臂平台试验，仅纳入三阴性乳腺癌（TNBC）特定队列的数据；HER2阳性和激素受体阳性亚组被排除在外。 采用Cochrane RoB 2工具（）评估方法学质量。该工具评估五个领域的偏倚风险：随机化过程、偏离预期干预措施、结局数据缺失、结局测量以及报告结果的选择性。每个领域以及总体偏倚风险均被归类为“低风险”、“存在一定问题”或“高风险”。 统计分析 采用R软件（R统计计算基金会，维也纳，奥地利）中的rjags和gemtc包进行贝叶斯网状meta分析（NMA）。随机效应模型作为主要分析框架，以考虑试验间潜在的临床和方法学异质性；同时拟合固定效应模型进行比较。先验分布设置如下：基线对数比值比采用正态分布N(0, 10²)，研究间异质性方差（τ²）采用均匀分布U(0, 2)。为检验这些假设的稳健性，敏感性分析采用替代的半柯西先验分布Cauchy+(0, 2.5)用于τ²，评估效应估计值和治疗排序（SUCRA）的稳定性。多臂随机对照试验（RCT）通过将共享对照组按比例分配到各主动比较组处理，以避免参与者重复计数。 模型参数通过马尔可夫链蒙特卡罗（MCMC）模拟进行估计，启动三条独立链。每条链经过20,000次迭代的预烧期，随后进行100,000次采样迭代用于后验估计。收敛性通过轨迹图和密度图进行可视化评估，并辅以Gelman–Rubin诊断（潜在尺度缩减因子，PSRF）；当PSRF接近1.0时，认为收敛充分。 治疗排名基于累积排序曲线下面积（SUCRA）值从后验分布中得出；数值越接近1表示疗效优越的概率越高。对于安全性结局（≥3级不良事件），SUCRA的解释被反转，即数值越高反映毒性风险越低。排名概率通过rank.probability函数提取，并使用pheatmap包以热图形式可视化，以展示各治疗在不同排名位置上的分布。 模型拟合通过残差偏差和偏差信息准则（DIC）进行评估。如果随机效应模型与固定效应模型相比，DIC差异（ΔDIC）超过3，则优先选择随机效应模型；ΔDIC < 3则提示模型性能差异可忽略不计。整体网络一致性通过比较一致性模型和不一致性模型进行评估；ΔDIC > 5被视为存在有意义的不一致性的指标。所有效应估计值均以比值比（OR）及其95%可信区间（CrIs）报告。 小样本效应和发表偏倚在STATA 18.0版本（StataCorp，College Station，TX，USA）中使用校正比较漏斗图进行评估。鉴于稀疏网络中漏斗图的可解释性有限，这些分析仅限于具有足够数量纳入研究的结局，并结合研究规模、效应量级和临床异质性进行谨慎解读。 使用GRADE和CINeMA进行证据分级。 证据确定性采用GRADE/CINeMA框架进行评估。随机对照试验（RCTs）被视为高确定性证据的起点。总体确定性根据六个领域的评估进行降级：偏倚风险、间接性、不精确性、异质性、不一致性以及发表偏倚或小研究效应。 偏倚风险使用Cochrane RoB 2工具在研究层面进行评估。随后使用CINeMA的贡献矩阵，根据每个试验对特定网络估计所贡献的信息比例，对这些风险进行加权和汇总。间接性通过检查传递性和可交换性假设进行评估。我们预设了潜在的效果修饰因素——包括基线疾病严重程度、干预强度和随访持续时间——并比较了直接和间接证据在人群特征、干预措施、对照措施和结局测量方面的一致性。 不精确性使用预设的最小重要差异（MID）阈值进行评估。对于二分类结局，比值比（OR）为1.25被定义为具有临床重要效果的阈值；如果95%可信区间（CrI）同时跨越了零值（OR = 1.0）和该MID阈值，则证据评级下调。异质性主要依据随机效应模型估计的研究间方差（τ²）来判断，并结合95%预测区间相对于MID的位置进行综合考虑。 对于包含闭合环路的网络，使用CINeMA内置的直接与间接证据比较来评估不一致性（不协调性），采用节点分裂法或设计-处理交互作用模型。通过将试验注册状态和灰色文献检索与校正比较漏斗图的目视检查相结合，评估发表偏倚和小样本效应。 每个领域被评定为“无问题”、“有一定问题”或“严重问题”。遵循GRADE指南，对于有一定问题的领域，证据确定性降级一级；对于存在严重问题的领域，降级两级。最终的确定性评级分为高、中、低或极低。 结果 研究选择与特征 系统检索共获得374条唯一记录。经过标题和摘要筛选后，对47篇全文文章进行了资格评估。最终，8项随机对照试验（RCT）符合纳入标准并被纳入分析（）（–）。这些试验共纳入2,073例患者，涉及六个特定的干预组：新辅助帕博利珠单抗联合化疗（Pembro-[NA]+CT）、新辅助阿替利珠单抗联合化疗（Atezo-[NA]+CT）、新辅助度伐利尤单抗联合化疗（Durva-[NA]+CT）、围手术期阿替利珠单抗联合化疗（Atezo-[P]+CT）、围手术期卡瑞利珠单抗联合化疗（Camre-[P]+CT）以及围手术期帕博利珠单抗联合化疗（Pembro-[P]+CT）。为便于分析，这些方案进一步归为四大总体策略：新辅助PD-1抑制剂联合化疗（PD-1[NA]+CT）、新辅助PD-L1抑制剂联合化疗（PD-L1[NA]+CT）、围手术期PD-1抑制剂联合化疗（PD-1[P]+CT）以及围手术期PD-L1抑制剂联合化疗（PD-L1[P]+CT）。 文献检索与研究筛选流程图[本研究依据PRISMA（系统综述和荟萃分析的首选报告项目）指南进行]。 所有纳入的试验均为比较含ICI方案与单独标准新辅助化疗的随机对照试验。核心化疗骨架包含紫杉类、蒽环类和铂类制剂。研究人群包括II–III期早期或局部晚期TNBC患者。样本量从每项试验数十名到数百名参与者不等。方法学设计分布均匀，有四项开放标签和四项双盲研究；随机化方案主要为1:1（n = 5），其次为1:4（n = 1）和2:1（n = 2）。 不同终点的数据可用性各不相同。全部八项试验均提供了病理完全缓解（pCR）数据。总生存期（OS）数据来自三项随访成熟的试验，涵盖1,123名患者。≥3级不良事件（AEs）数据从涉及905名患者的五项试验中提取。纳入研究的具体特征汇总于中。 纳入网络荟萃分析的研究基线特征。 db，双盲；NA，仅新辅助；Peri，围手术期（新辅助+辅助）；Pembro，帕博利珠单抗；Durva，度伐利尤单抗；Atezo，阿替利珠单抗；Camre，卡瑞利珠单抗；ICI，免疫检查点抑制剂；pCR，病理完全缓解。 偏倚风险评估 采用Cochrane RoB 2工具对纳入的8项随机对照试验（RCT）进行偏倚风险评估（）。总体而言，3项试验被判定为低偏倚风险，5项引起一些关注；无试验被判定为高偏倚风险，支持总体稳健的证据基础。在领域层面，全部8项试验在随机化过程和报告结果的选择方面均显示低偏倚风险。同样，所有试验在结局数据缺失方面均被评为低风险。相比之下，仅4项试验在偏离预期干预方面被认为低风险，同等数量的试验在结局测量方面符合低偏倚风险标准。 使用Cochrane偏倚风险工具2.0评估研究的结果总结。 按免疫治疗靶点类别的网络荟萃分析 总生存期和病理完全缓解 对于总生存期（OS），汇集了3项试验、涵盖1,123例患者的数据，涉及三个干预臂（）。与单纯化疗（CT）相比，新辅助PD-L1抑制联合CT（PD-L1[NA]+CT；HR 0.24，95% CrI 0.08–0.72）和围手术期PD-1抑制联合CT（PD-1[P]+CT；HR 0.63，95% CrI 0.48–0.82）均显著改善了早期TNBC的生存结局（）。 按抑制剂类别比较免疫检查点抑制剂的网络几何结构。（A）总生存期（OS）的网络图。（B）病理完全缓解（pCR）的网络图。（C）AEs ≥3的网络图。 按抑制剂类别划分的免疫检查点抑制剂在OS、pCR和≥3级不良事件方面的联赛表与GRADE证据概要。符号表示根据GRADE分类的证据确定性：*，高；&，中等；+，低；-，极低。 对于病理完全缓解（pCR），八项试验涉及2,073例患者，为四种干预策略提供了数据（）。相对于CT，新辅助PD-1抑制联合CT（PD-1[NA]+CT；OR 2.56，95% CrI 1.55–4.22）与pCR率显著升高相关，尽管基于极低确定性证据。围手术期PD-L1抑制联合CT（PD-L1[P]+CT；OR 2.00，95% CrI 1.29–3.09）和围手术期PD-1抑制联合CT（PD-1[P]+CT；OR 1.71，95% CrI 1.39–2.10）相似
全文（开放获取） · 英文原文
【Abstract】 BACKGROUND: Immune checkpoint inhibitors (ICIs) combined with chemotherapy have become a cornerstone in managing early and locally advanced triple-negative breast cancer (TNBC). Yet direct comparative evidence on the relative efficacy and safety of different PD-1/PD-L1 agents and treatment strategies-particularly neoadjuvant-only versus perioperative, "full-course" approaches-remains limited. To guide clinical decision-making, we performed a Bayesian network meta-analysis (NMA) comparing six specific ICI-based regimens and four overarching treatment strategies. METHODS: We systematically searched PubMed, Embase, Cochrane Library, and Web of Science from database inception through February 25, 2026. Eligible randomized controlled trials (RCTs) enrolled patients with previously untreated early or locally advanced TNBC and compared ICI plus neoadjuvant chemotherapy with chemotherapy alone. The primary endpoint was pathological complete response (pCR); secondary endpoints included overall survival (OS) and grade ≥3 adverse events (AEs ≥3). We assessed risk of bias using the Cochrane RoB 2 tool. A Bayesian random-effects model was applied for the NMA; effects are reported as odds ratios (ORs) or hazard ratios (HRs) with 95% credible intervals (CrIs). Treatment rankings were derived from surface under the cumulative ranking curve (SUCRA) values. Certainty of evidence was evaluated using the GRADE framework, supplemented by the CINeMA tool. RESULTS: Eight RCTs involving 2,073 patients met the inclusion criteria, encompassing four strategic treatment categories and six specific ICI-containing regimens. For pCR, PD-1 (neoadjuvant) + CT (OR 2.56, 95% CrI 1.55-4.22), PD-L1 (perioperative) + CT (OR 2.00, 95% CrI 1.29-3.09), and PD-1 (perioperative) + CT (OR 1.71, 95% CrI 1.39-2.10) each significantly improved pCR rates versus chemotherapy alone, though certainty of evidence ranged from moderate to very low. PD-L1 (neoadjuvant) + CT showed a numerical trend toward higher pCR (OR 1.41, 95% CrI 0.98-2.01) without reaching statistical significance. Regarding OS, both PD-L1 (neoadjuvant) + CT (HR 0.24, 95% CrI 0.08-0.72) and PD-1 (perioperative) + CT (HR 0.63, 95% CrI 0.48-0.82) demonstrated significant survival benefits over chemotherapy. In sensitivity analyses, durvalumab (neoadjuvant) + CT and pembrolizumab (perioperative) + CT yielded identical HRs (0.24 and 0.63, respectively). Safety profiles varied: camrelizumab (perioperative) + CT (OR 0.60, 95% CrI 0.34-1.04) and pembrolizumab (neoadjuvant) + CT (OR 0.53, 95% CrI 0.28-1.01) suggested reduced risks of AEs ≥3, but these findings did not achieve statistical significance and were supported by low-to-very-low certainty evidence. CONCLUSIONS: ICI-chemotherapy combinations consistently improve pCR rates in early and locally advanced TNBC. Notably, while our analysis hints at potential survival advantages with neoadjuvant PD-L1 blockade and perioperative PD-1 inhibition, these signals derive from only three studies and remain hypothesis-generating rather than practice-changing. Future head-to-head RCTs are warranted to confirm these observations and to better define long-term safety profiles. SYSTEMATIC REVIEW REGISTRATION: https://www.crd.york.ac.uk/prospero/, identifier CRD420261385223. 【Introduction】 Breast cancer remains the most frequently diagnosed malignancy among women worldwide and ranks fifth as a cause of cancer-related mortality. According to GLOBOCAN 2022 estimates, approximately 2.3 million new cases occur each year—accounting for 11.6% of all incident cancers—with roughly 670,000 deaths globally (). Beyond its profound physiological and psychological toll on patients and families, the disease imposes substantial socioeconomic burdens through high treatment costs and lost workforce productivity. Triple-negative breast cancer (TNBC), defined by the absence of oestrogen receptor, progesterone receptor, and human epidermal growth factor receptor 2 expression, accounts for 15–20% of all breast cancers (). Characterized by marked aggressiveness, early relapse, and limited targeted therapeutic options, TNBC carries the poorest prognosis among breast cancer subtypes ().For decades, anthracycline- and taxane-based neoadjuvant chemotherapy represented the standard of care for patients with localized or operable TNBC. Nevertheless, pathological complete response (pCR) rates have remained suboptimal, typically ranging from 30% to 40% (). Failure to achieve pCR is a strong predictor of poor long-term outcomes, including event-free and disease-free survival (). Overcoming the efficacy ceiling of conventional chemotherapy to improve pCR rates—and, critically, to translate these gains into survival benefits—therefore represents a central clinical challenge in early-stage and locally advanced TNBC. The advent of immune checkpoint inhibitors (ICIs) targeting the programmed death-1 (PD-1)/programmed death-ligand 1 (PD-L1) axis has reshaped the therapeutic landscape across solid tumours. The relatively high tumour mutational burden and dense tumour-infiltrating lymphocyte populations characteristic of TNBC provide a compelling biological rationale for immunotherapy (). In the metastatic setting, randomised trials have shown that PD-1/PD-L1 inhibition combined with chemotherapy significantly prolongs progression-free and overall survival, paving the way for translation into earlier disease stages (, ). Landmark phase III trials—most notably KEYNOTE-522 and IMpassion031—established that adding ICIs to neoadjuvant chemotherapy increases pCR rates by an absolute 10 to 20 percentage points and improves event-free survival. These results have defined a new standard for high-risk early TNBC (, ). Despite these advances, several unresolved questions persist. First, multiple ICIs with distinct molecular targets—anti-PD-1 versus anti-PD-L1 agents—are now clinically available, yet whether meaningful differences exist in their efficacy or safety profiles remains unclear (). Second, the optimal duration and timing of ICI administration continue to provoke debate. The “perioperative, full-course” strategy exemplified by KEYNOTE-522 aims to eradicate minimal residual disease through extended immune exposure but raises concerns about prolonged treatment duration, cumulative toxicity, and cost. Conversely, the “neoadjuvant-only” approach investigated in trials such as GeparNuevo offers a shorter treatment course, but whether its long-term survival outcomes are non-inferior to the perioperative model remains unproven (). Critically, no head-to-head randomised controlled trials (RCTs) have directly compared these agents or scheduling strategies. Given this complexity, conventional pairwise meta-analyses cannot rank multiple interventions simultaneously. To fill this evidence gap, we performed a Bayesian network meta-analysis (NMA) synthesising all available RCT data. Our objective was to compare the efficacy and safety of specific ICIs (e.g., pembrolizumab, atezolizumab, camrelizumab) and overarching treatment strategies (perioperative versus neoadjuvant-only) across key endpoints—including pCR, overall survival, and grade ≥3 adverse events. By generating probabilistic treatment rankings, this analysis aims to provide high-level evidence to inform clinical decision-making and future guideline development. 【Materials and methods】 This network meta-analysis (NMA) was conducted in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension statement for network meta-analyses (PRISMA-NMA; ) (). To ensure transparency and mitigate outcome reporting bias, the study protocol was prospectively registered with PROSPERO (CRD420261385223). 【Data sources and search strategy】 We performed a systematic literature search across PubMed, EMBASE, the Cochrane Library, Web of Science, ClinicalTrials.gov, and the WHO International Clinical Trials Registry Platform (ICTRP). Grey literature was supplemented by hand-searching abstracts from major oncology conferences (SABCS, ASCO, ESMO) over the preceding five years, along with pharmaceutical company websites and relevant dissertations. The search strategy combined Medical Subject Headings (MeSH) and free-text terms related to triple-negative breast cancer (e.g., “Triple-Negative Breast Neoplasms,” “ER-Negative PR-Negative HER2-Negative Breast Cancer”), neoadjuvant therapy (e.g., “Neoadjuvant Therapies”), and immune checkpoint inhibitors (e.g., “Checkpoint Inhibitors, Immune,” “pembrolizumab,” “atezolizumab”) (). Records were searched from database inception through February 25, 2026, with no language restrictions. 【Eligibility criteria】 Inclusion criteria were as follows: (1) randomised controlled trials (RCTs) enrolling adults with histologically confirmed, treatment-naïve, operable early-stage or locally advanced TNBC; (2) intervention arms comprising a PD-1 or PD-L1 inhibitor combined with standard neoadjuvant chemotherapy—both “neoadjuvant-only” and “perioperative (full-course)” strategies were eligible; (3) control arms receiving standard neoadjuvant chemotherapy alone; and (4) availability of at least one of the following outcomes—pathological complete response (pCR), overall survival (OS), or grade ≥3 adverse events (AEs ≥3). pCR was defined as the absence of invasive carcinoma in both breast and axilla per local pathology assessment. OS was defined as the time from randomisation to death from any cause. AEs ≥3 were graded according to the National Cancer Institute Common Terminology Criteria for Adverse Events (CTCAE). Exclusion criteria were: (1) RCTs reporting on overlapping patient populations or duplicate publications; (2) studies with incomplete or inadequately reported outcome data; and (3) non-randomised studies, reviews, or case reports. After removing duplicates, two reviewers independently screened all records by title and abstract. Full-text articles were then assessed for eligibility. Discrepancies during screening or data extraction were resolved through consensus or by consulting a third senior investigator. For each included trial, we ensured that the most recent and comprehensive dataset was extracted. 【Data extraction and quality assessment】 Using a standardised, pilot-tested form, two investigators independently extracted data from each eligible publication. Extracted variables included trial name, NCT identifier, first author, publication year, trial phase, sample size, median age, intervention and control regimens, and tumour stage. For efficacy and safety outcomes, we extracted odds ratios (ORs) with 95% credible intervals (CrIs) for pCR and AEs ≥3, and hazard ratios (HRs) with 95% CrIs for OS. For multi-arm platform trials such as I-SPY2, only data from TNBC-specific cohorts were included; HER2-positive and hormone receptor-positive subgroups were excluded. Methodological quality was assessed using the Cochrane RoB 2 tool (). This instrument evaluates bias across five domains: the randomisation process, deviations from intended interventions, missing outcome data, measurement of the outcome, and selection of the reported result. Each domain, along with the overall risk of bias, was categorised as “low risk,” “some concerns,” or “high risk.” 【Statistical analysis】 A Bayesian network meta-analysis (NMA) was performed using the rjags and gemtc packages in R (R Foundation for Statistical Computing, Vienna, Austria). Random-effects models served as the primary analytical framework to account for potential clinical and methodological heterogeneity across trials; fixed-effect models were fitted for comparison. Vague priors were assigned as follows: a normal distribution N(0, 10²) for the baseline log-odds ratio, and a uniform distribution U(0, 2) for the between-study heterogeneity variance (τ²). To test the robustness of these assumptions, sensitivity analyses employed an alternative half-Cauchy prior, Cauchy+(0, 2.5), for τ², evaluating stability in effect estimates and treatment rankings (SUCRA). Multi-arm randomised controlled trials (RCTs) were handled by splitting the shared control arm proportionally across active comparators to avoid double-counting of participants. Model parameters were estimated via Markov Chain Monte Carlo (MCMC) simulation, initiating three independent chains. Each chain underwent a burn-in phase of 20,000 iterations, followed by 100,000 sampling iterations for posterior estimation. Convergence was assessed visually using trace and density plots, supplemented by the Gelman–Rubin diagnostic (potential scale reduction factor, PSRF); convergence was deemed adequate when PSRF approached 1.0. Treatment rankings were derived from the posterior distributions using surface under the cumulative ranking curve (SUCRA) values; values closer to 1 indicated a higher probability of superior efficacy. For safety outcomes (AEs ≥3), SUCRA interpretation was inverted such that higher values reflected a lower risk of toxicity. Ranking probabilities were extracted via the rank.probabilityfunction and visualised as heatmaps using the pheatmap package to illustrate the distribution of treatments across rank positions. Model fit was evaluated using residual deviance and the Deviance Information Criterion (DIC). A random-effects model was preferred if the DIC difference (ΔDIC) exceeded 3 compared with the fixed-effect model; ΔDIC < 3 suggested negligible differences in model performance. Global network consistency was appraised by comparing consistency and inconsistency models; a ΔDIC > 5 was considered indicative of meaningful inconsistency. All effect estimates are reported as odds ratios (ORs) with 95% credible intervals (CrIs). Small-study effects and publication bias were assessed in STATA version 18.0 (StataCorp, College Station, TX, USA) using comparison-adjusted funnel plots. Given the limited interpretability of funnel plots in sparse networks, these analyses were restricted to outcomes with an adequate number of contributing studies and interpreted cautiously alongside study size, effect magnitude, and clinical heterogeneity. 【Grading of evidence using GRADE and CINeMA】 The certainty of evidence was evaluated using the GRADE/CINeMA framework.Randomised controlled trials (RCTs) were taken as the starting point for high-certainty evidence. Overall certainty was downgraded based on assessments across six domains: risk of bias, indirectness, imprecision, heterogeneity, incoherence, and publication bias or small-study effects. Risk of bias was evaluated at the study level using the Cochrane RoB 2 tool. CINeMA’s contribution matrix was then used to weight and aggregate these risks according to the proportion of information each trial contributed to the specific network estimate. Indirectness was assessed by examining the transitivity and exchangeability assumptions. We prespecified potential effect modifiers—including baseline disease severity, intervention intensity, and follow-up duration—and compared the consistency of direct and indirect evidence with respect to population characteristics, interventions, comparators, and outcome measurements. Imprecision was evaluated using prespecified thresholds for the minimal important difference (MID). For dichotomous outcomes, an odds ratio (OR) of 1.25 was defined as the threshold for a clinically important effect; evidence was rated down if the 95% credible interval (CrI) crossed both the null value (OR = 1.0) and this MID threshold. Heterogeneity was informed primarily by the between-study variance (τ²) estimated from the random-effects model, considered together with the location of the 95% prediction interval relative to the MID. For networks containing closed loops, incoherence (inconsistency) was assessed using CINeMA’s built-in direct-versus-indirect evidence comparisons, applying either node-splitting or design-by-treatment interaction models. Publication bias and small-study effects were evaluated by triangulating trial registration status and grey literature searches with visual inspection of comparison-adjusted funnel plots. Each domain was rated as “no concerns,” “some concerns,” or “serious concerns.” Following GRADE guidance, evidence certainty was downgraded by one level for domains with some concerns and by two levels for those with serious concerns. The final certainty ratings were categorised as high, moderate, low, or very low. 【Results】 【Study selection and characteristics】 The systematic search yielded 374 unique records. After title and abstract screening, 47 full-text articles were assessed for eligibility. Ultimately, eight randomised controlled trials (RCTs) met the inclusion criteria and were incorporated into the analysis () (–). These trials enrolled a total of 2,073 patients across six specific intervention arms: neoadjuvant pembrolizumab plus chemotherapy (Pembro-[NA]+CT), neoadjuvant atezolizumab plus chemotherapy (Atezo-[NA]+CT), neoadjuvant durvalumab plus chemotherapy (Durva-[NA]+CT), perioperative atezolizumab plus chemotherapy (Atezo-[P]+CT), perioperative camrelizumab plus chemotherapy (Camre-[P]+CT), and perioperative pembrolizumab plus chemotherapy (Pembro-[P]+CT). For analytical purposes, these regimens were further grouped into four overarching strategies: neoadjuvant PD-1 inhibition plus chemotherapy (PD-1[NA]+CT), neoadjuvant PD-L1 inhibition plus chemotherapy (PD-L1[NA]+CT), perioperative PD-1 inhibition plus chemotherapy (PD-1[P]+CT), and perioperative PD-L1 inhibition plus chemotherapy (PD-L1[P]+CT). Flow diagram of the literature search and study selection process [this study was conducted in accordance with the PRISMA (preferred reporting items for systematic reviews and meta-analyses) guidelines]. All included trials were RCTs comparing ICI-containing regimens against standard neoadjuvant chemotherapy alone. Core chemotherapeutic backbones consisted of taxane-, anthracycline-, and platinum-based agents. The study populations comprised patients with stage II–III early or locally advanced TNBC. Sample sizes ranged from several dozen to several hundred participants per trial. Methodological designs were evenly split, with four open-label and four double-blind studies; randomisation schemes were predominantly 1:1 (n = 5), followed by 1:4 (n = 1) and 2:1 (n = 2). Data availability varied across endpoints. All eight trials provided pathological complete response (pCR) data. Overall survival (OS) data were available from three trials with mature follow-up, encompassing 1,123 patients. AEs ≥3 data were extracted from five trials involving 905 patients. Detailed characteristics of the included studies are summarised in . Baseline characteristics of studies included in the network meta-analysis. db, double-blind; NA, neoadjuvant only; Peri, perioperative (neoadj+adjuv); Pembro, pembrolizumab; Durva, durvalumab; Atezo, atezolizumab; Camre, camrelizumab; ICI, immune checkpoint inhibitor; pCR, pathological complete response. 【Risk of bias assessment】 Risk of bias for the eight included RCTs was evaluated using the Cochrane RoB 2 tool (). Overall, three trials were judged to be at low risk of bias, while five raised some concerns; no trial was deemed to be at high risk of bias, supporting a generally robust evidence base. At the domain level, all eight trials demonstrated low risk of bias regarding the randomisation process and the selection of reported results. Similarly, all trials were rated as low risk for missing outcome data. In contrast, only four trials were considered low risk with respect to deviations from intended interventions, and an equivalent number met the criteria for low risk of bias in outcome measurement. Summary of results from assessment of studies by using Cochrane risk of bias tool 2.0. 【Network meta-analysis by immunotherapy target class】 【Overall survival and pathological complete response】 For overall survival (OS), data from three trials encompassing 1,123 patients were pooled across three intervention arms (). Compared with chemotherapy alone (CT), both neoadjuvant PD-L1 inhibition plus CT (PD-L1[NA]+CT; HR 0.24, 95% CrI 0.08–0.72) and perioperative PD-1 inhibition plus CT (PD-1[P]+CT; HR 0.63, 95% CrI 0.48–0.82) significantly improved survival outcomes in early TNBC (). Network geometry for comparisons of immune checkpoint inhibitors by inhibitor class. (A) Network plot for overall survival (OS). (B) Network plot for pathological complete response (pCR). (C) Network plot for AEs ≥3. League tables and GRADE evidence profiles for immune checkpoint inhibitors by inhibitor class regarding OS, pCR, and AEs ≥3. Symbols indicate the certainty of evidence according to GRADE classification: *, high; &, moderate; +, low; -, very low. For pathological complete response (pCR), eight trials involving 2,073 patients contributed data across four intervention strategies (). Relative to CT, neoadjuvant PD-1 inhibition plus CT (PD-1[NA]+CT; OR 2.56, 95% CrI 1.55–4.22) was associated with a marked increase in pCR rates, albeit based on very low–certainty evidence. Perioperative PD-L1 inhibition plus CT (PD-L1[P]+CT; OR 2.00, 95% CrI 1.29–3.09) and perioperative PD-1 inhibition plus CT (PD-1[P]+CT; OR 1.71, 95% CrI 1.39–2.10) simila