Comparing efficacy of neoadjuvant therapy of triple-negative breast cancer: A Bayesian network meta-regression analysis.

1. Introduction
Breast cancer is the most common malignant tumor among women in the world, and it is also the second leading cause of cancer death in women. During 2012 to 2021, the incidence rate of breast cancer increased by 1% every year.[1] It poses a great threat to the physical and mental health of patients worldwide.
The molecular subtypes of breast cancer also vary according to race and ethnicity. White women have the highest proportion of hormone receptor positive/human epidermal growth factor receptor 2 negative breast cancer, while Black women have a significantly higher proportion of triple-negative breast cancer (TNBC) than other groups.[1]
TNBC accounts for 12% to 17% of all breast cancers.[2] Compared to other subtypes of breast cancer, TNBC patients suffer worse clinical outcomes.[2,3] TNBC patients who achieve pathological complete response (pCR) will have a better prognosis than non-pCR.[4] In early stage TNBC patients, neoadjuvant chemotherapy has become a standard approach and is more likely to achieve pCR than non-TNBC patients.[4]
Neoadjuvant chemotherapy is a treatment approach administered before primary surgery. It has revolutionized the management of breast cancer, particularly in aggressive subtypes, such as TNBC. Its primary goal is to shrink tumors, making them more amenable to conservative surgical procedures and reducing the presence of early micrometastases.[5] Accordingly, regulatory guidance supports the use of the pCR as an end point for clinical testing of neoadjuvant treatment in patients with early TNBC.[6,7]
Anthracyclines, cyclophosphamides, and taxanes are the essential drugs in neoadjuvant chemotherapy regimens for TNBC.[8] As the higher pCR rate and event-free survival (EFS)/overall survival (OS) benefit, platinum plays an important antitumor roles in neoadjuvant chemotherapy for TNBC.[9,10] Platinum is a DNA cross-linking agent, which can cross-connect with the DNA after entering the tumor cells, cause DNA strand breaks in tumor cells.[11]
Currently, poly-ADP-ribose polymerase inhibitors (PARPi) are used in neoadjuvant chemotherapy for TNBC patients, especially with BRCA gene mutations.[12] PARPi are a class of drugs that target the DNA repair mechanism of tumor cells. Their core mechanism of action is based on the principle of “synthetic lethality,” which selectively kills cancer cells by interfering with the DNA repair process.[13]
The vascular endothelial growth factor, such as bevacizumab, is an important regulator of tumor angiogenesis and metastasis.[14,15] Bevacizumab plays various roles in the tumor blood vessels by specifically binding to vascular endothelial growth factor and blocking its interaction with receptors.[16]
In recent years, immune checkpoint inhibitor (ICI) therapy has been successful in metastatic TNBC.[17] ICI therapy is directed against the interaction between the programmed death 1 (PD-1) and programmed death ligand 1 (PD-L1).[18] PD-1 is a co-inhibitory molecule expressed by activated T-cells when antigen-presenting cells or tumor cells are combined with PD-L1, which further lead to inhibiting the T-cell activation and suppressing the body’s antitumor immune response.[19]
Chemotherapy schemes composed of various chemotherapeutic drugs were used in neoadjuvant chemotherapy of TNBC. Some obtained high pCR rates. However, the improvement in pCR did not translate into benefits for EFS or OS.
In recent years, there have been an increasing number of Bayesian network meta-analyses (NMAs) on neoadjuvant chemotherapy regimens for TNBC, with nonconclusive evidence. In 2022, Li et al[20] published an NMA evaluating 8 neoadjuvant treatment options for TNBC. The treatment regimen included the combination of platinum, bevacizumab, PARPi, and ICI. In this previous study, the observation indicator was pCR.[20] Similarly, in 2024, Liu et al[21] showed that bevacizumab associated with platinum-containing regimens is likely to be the optimal treatment option for neoadjuvant TNBC, but this results are limited. In terms of data analytics, they only considered the covariates of hazard ratio (HR) and inferred the result by surface under the cumulative ranking curve (SUCRA), making the final results less reliable.[21]
Hence, we conducted this Bayesian NMA of randomized controlled trials (RCTs) about neoadjuvant therapy for TNBC, which includes direct and indirect comparisons among regimens, to identify the most effective maintenance treatment for TNBC patients that, in turn, could improve the oncological clinical practice.

2. Methods
This Bayesian NMA was guided by the PRISMA guideline (Preferred Reporting Items for Systematic Reviews and Meta-analysis).[22]

2.1. Search strategy
We searched Google Scholars, Cochrane Library (CENTRAL), PubMed, Scopus, Embase, Web of ClinicalTrials.gov, and major international conference databases from inception to January 2025, with the following Mesh terms: (“Breast Cancer” OR “Breast Carcinoma” OR “Breast Neoplasm” OR “Breast Tumor” OR “Breast Malignant Tumor”) AND (“Neoadjuvant” OR “Neoadjuvant therapy” OR “Neoadjuvant Treatment”) AND (“Triple Negative” OR “ER Negative PR Negative HER2 Negative” OR “Endocrine Negative HER2 Negative”) AND (“Randomized” OR “Randomization” OR “Allocation Random”).

2.2. Selection criteria
The inclusion criteria were as follows: the study subjects were diagnosed TNBC patients without metastatic; the study design was RCT about neoadjuvant therapy; sufficient information was provided on pCR and/or PFS and/or OS.
The exclusion criteria were as follows: the data required for analysis was not reported; articles were observational studies, letters, or review; articles reported by non-English literature.

2.3. Data extraction and quality assessment
Two investigators independently searched and assessed the eligibility of each study by reading the title and abstract or even full-text when necessary. Data were independently extracted by them also. Any discrepancy was arbitrated by the senior reviewer. At last, the risk of bias for each included RCT was assessed by Cochrane Risk of Bias tool. The following information were collected: names of the first authors, publication year, country, number of patients, condition, therapeutic drugs, treatment dosage, and the outcomes included numbers of patients reached pCR, HRs, and confidence intervals (CIs) associated with EFS and OS. Subsequently, the data regarding EFS and OS at 18, 24, 30, 36, 42, and 48 months were harvested from Kaplan–Meier curve by GetData 2.26 (https://getdata.sourceforge.net/download).

2.4. Research endpoint
The primary research endpoint was the pCR rate. The secondary endpoint was the odds ratios (ORs) at different time points of EFS and OS. The tertiary endpoints were the HRs of EFS and OS compared by Bayesian NMA, as well as corresponding Bayesian network meta-regression analysis with the median follow-up time as the covariate.

2.5. Data analysis
For pCR and EFS/OS rate at each time node, ORs were generated by NMA using STATA 17.0 MP (https://www.stata.com/statamp/) to make pairwise comparisons among regimens. SUCRA was also formulated; a higher SUCRA indicates a higher probability of being the better treatment. However, whether the effect size between any pair with corresponding SUCRAs reached the significance was determined by net-league table, also called matrix in algebra. Inconsistency and consistency tests were performed to examine the existence of inconsistency. Publication bias was assessed by funnel plot as well.
For HRs regarding OS and EFS in each study Napierian logarithm HR (lnHR) and standard error of lnHR (selnHR) for each study were calculated by STATA 17.0 MP. Subsequently these data (lnHR and selnHR for OS and PFS) were input into RStudio 4.2.2 (https://cran.rstudio.com/bin/windows/base) by “gemtc” package to conduct Bayesian NMA to generate pairwise HRs, SUCRA, and matrix. Markov chain Monte Carlo was used to obtain posterior distributions, with 2000 burn-ins and 300,000 iterations of 4 each chain and a thinning interval of 10 for each outcome. Brooks–Gelman–Rubin diagnostics and Trace and density plots were used to evaluate and visualize the convergence of the model over iterations. For heterogeneity analysis, if I2 < 50% and P > .01, fixed effect model would be implemented; if 50% < I2 < 75%, random effect model would be carried out; if I2 > 75%, Galbraith plot would be drawn to preclude the studies outside the outlines.
Finally, in the sensitivity analysis, we used median follow up time as a covariate to perform meta-regression analyses to eliminate potential confounding factors.
If the 95% CI of the comparison value is greater than or <1, it indicates a significant difference in the comparison result.

3. Results

3.1. Characteristics of the included studies
A detailed description of the included studies can be found in Table 1. Initially, we retrieved 1038 articles from all the 5 databases; after meticulous screening, only 34 studies included 25 RCTs reporting 6838 patients were eligible for our study (Fig. 1).[9,10,12,23–53] Seventeen different maintenance regimes were evaluated within the included studies, namely, containing anthracyclines and taxanes (AT); containing anthracyclines, taxanes, and bevacizumab (ATBev); containing anthracyclines, taxanes, and everolimus (ATEve); containing anthracyclines, taxanes, and gemcitabine (ATGem); containing anthracyclines, taxanes, and PD-1 (ATPD1); containing anthracyclines, taxanes, and PD-L1 (ATPDL1); containing anthracyclines, taxanes, and platinum (ATPt); containing anthracyclines, taxanes, platinum, and bevacizumab (ATPtBev); containing anthracyclines, taxanes, platinum, and PARPi (ATPtPARPi); containing anthracyclines, taxanes, platinum, and PD-1 (ATPtPD1); containing anthracyclines, taxanes, platinum, and PD-L1 (ATPtPDL1); containing anthracyclines, taxanes, and capacitabine (ATX); containing taxanes only (T); containing taxanes and PARPi (TPARPi); containing taxanes and platinum (TPt); containing taxanes, platinum, and everolimus (TPtEve); containing taxanes, platinum, and PD-L1 (TPtPDL1). Patients included come from Asia, Europe, South/North America, and Africa. As shown in Figure S1, Supplemental Digital Content, https://links.lww.com/MD/R101, the risk of bias assessment conducted showed that there was no high risk of bias in the included studies.

3.2. Primary endpoint
Figure 2 shows the network graphs of pairwise comparison of regimens on pCR rate. Compared with AT, only ATPtPD1 (OR = 5.68, 95% CI: 1.49–21.62) significantly increased the pCR rate (Table S1, Supplemental Digital Content, https://links.lww.com/MD/R101)

3.3. Second endpoint

3.3.1. EFS at each time point
Figure 3A shows the network graphs of pairwise comparison of regimens on each time point of the EFS.
At 12th month, compared with AT, there was no treatment significantly increased the 12 months EFS rate (Table S2A, Supplemental Digital Content, https://links.lww.com/MD/R101).
At 18th month, compared with AT, only ATPtPD1 (OR = 2.28, 95% CI: 1.24–4.19), significantly increased the 18 months EFS rate (Table S2B, Supplemental Digital Content, https://links.lww.com/MD/R101).
At 24th month, compared with AT, only ATPtPD1 (OR = 2.43, 95% CI: 1.01–5.89), significantly increased the 24 months EFS rate (Table S2C, Supplemental Digital Content, https://links.lww.com/MD/R101).
At 30th month, compared with AT, only ATPtPD1 (OR = 3.021, 95% CI: 1.40–7.38), significantly increased the 30 months EFS rate (Table S2D, Supplemental Digital Content, https://links.lww.com/MD/R101).
At 36th month, compared with AT, ATPtPD1 (OR = 4.23, 95% CI: 1.15–15.57), and ATPt (OR = 2.54, 95% CI: 1.16–5.57). ATPtPD1, as the top SUCRA-ranked intervention, did not show significant advantage when compared with ATPt (Table S2E, Supplemental Digital Content, https://links.lww.com/MD/R101).
At 42nd month, compared with AT, ATPtPD1 (OR = 4.62, 95% CI: 1.38–15.50), and ATPt (OR = 2.68, 95% CI: 1.27–5.69). ATPtPD1, as the top SUCRA-ranked intervention, did not show significant advantage when compared with ATPt (Table S2F, Supplemental Digital Content, https://links.lww.com/MD/R101).
At 48th month, compared with AT, ATPtPD1 (OR = 4.04, 95% CI: 1.47–11.13), and ATPt (OR = 2.18, 95% CI: 1.12–4.23). ATPtPD1, as the top SUCRA-ranked intervention, did not show significant advantage when compared with ATPt (Table S2G, Supplemental Digital Content, https://links.lww.com/MD/R101).
As described, the regimen with the best significant effect on EFS from 12 to 48 months compared to AT was ATPtPD1 (Table 2).

3.3.2. OS at each time point
Figure 3B shows the network graphs of pairwise comparison of regimens on each time point of the OS.
At 12th month, compared with AT, there was no treatment significantly increased the 12 months OS rate (Table S3A, Supplemental Digital Content, https://links.lww.com/MD/R101).
At 18th month, compared with AT, ATPtPD1 (OR = 3.56, 95% CI: 1.33–9.52), and ATPt (OR = 2.72, 95% CI: 1.18–6.28). ATPtPD1, as the top SUCRA-ranked intervention, did not show significant advantage when compared with ATPt (Table S3B, Supplemental Digital Content, https://links.lww.com/MD/R101).
At 24th month, compared with AT, there was no treatment significantly increased the 24 months OS rate (Table S3C, Supplemental Digital Content, https://links.lww.com/MD/R101).
At 30th month, compared with AT, ATPtPD1 (OR = 2.49, 95% CI: 1.27–4.78), and ATPt (OR = 1.98, 95% CI: 1.15–3.40). ATPtPD1, as the top SUCRA-ranked intervention, did not show significant advantage when compared with ATPt (Table S3D, Supplemental Digital Content, https://links.lww.com/MD/R101).
At 36th month, compared with AT, ATPtPD1 (OR = 2.49, 95% CI: 1.34–4.65), and ATPt (OR = 1.94, 95% CI: 1.18–3.19). ATPtPD1, as the top SUCRA-ranked intervention, did not show significant advantage when compared with ATPt (Table S3E, Supplemental Digital Content, https://links.lww.com/MD/R101).
At 42nd month, compared with AT, ATPtPD1 (OR = 3.17, 95% CI: 1.74–5.78), and ATPt (OR = 2.32, 95% CI: 1.43–3.77). ATPtPD1, as the top SUCRA-ranked intervention, did not show significant advantage when compared with ATPt (Table S3F, Supplemental Digital Content, https://links.lww.com/MD/R101).
At 48th month, compared with AT, only ATPtPD1 (OR = 2.97, 95% CI: 1.12–7.91), showed the significant advantage on the 48 months OS rate (Table S3G, Supplemental Digital Content, https://links.lww.com/MD/R101).
As described, the regimen with the best significant effect on OS from 12 to 48 months compared to AT was ATPtPD1 (Table 2).

3.4. Tertiary endpoints

3.4.1. HR for EFS and regression analysis
Due to limited EFS data, we treated data regarding disease-free survival and relapse-free survival reported in these studies as EFS data. A total of 13/25 trails reported outcomes associated with the HRs of EFS. The 10 included interventions were compared directly and indirectly (Fig. 4A). All articles reported the median follow-up time.
In the raw analysis: compared with AT, ATPtPD1 (HR = 2.24, 95% CI: 1.42–3.59), ATPtPARPi (HR = 1.64, 95% CI: 1.10–2.53), and ATPt (HR = 1.53, 95% CI: 1.17–2.08) showed significant advantage, as shown in Table S4A, Supplemental Digital Content, https://links.lww.com/MD/R101.
After the regression analysis: compared with AT, ATPtPD1 (HR = 2.29, 95% CI: 1.40–3.87), ATPtPARPi (HR = 1.66, 95% CI: 1.08–2.65), and ATPt (HR = 1.57, 95% CI: 1.17–2.22) showed significant advantage, as shown in Table S4B, Supplemental Digital Content, https://links.lww.com/MD/R101.
Time window analysis: when compared with AT, ATPtPD1 showed significant advantage in the time window of 0 to 104 months after the initiation of treatment (Fig. 5A). The time window for ATPt was 15 to 85 months(Fig. 5B) and for ATPtPARPi was 18 to 72 months (Fig. 5C).
As described, the regimen with the best significant effect on HR for EFS compared to AT was ATPtPD1 (Table 2).

3.4.2. HR for OS and regression analysis
A total of 9/25 trails reported outcomes associated with the HRs of OS. The 9 included interventions were compared directly and indirectly (Fig. 4B). All articles reported the median follow-up time.
In the raw analysis: compared with AT, only ATPtPD1 (HR = 2.67, 95% CI: 1.03–7.35) showed significant advantage, as shown in Table S5A, Supplemental Digital Content, https://links.lww.com/MD/R101.
After the regression analysis: compared with AT, only ATPtPD1 (HR = 2.71, 95% CI: 1.19–6.39) showed significant advantage, as shown in Table S5B, Supplemental Digital Content, https://links.lww.com/MD/R101.
Time window analysis: when compared with AT, ATPtPD1 showed significant advantage in the time window of 0 to 74 months after the initiation of treatment (Fig. 6).
As described, the regimen with the best significant effect on HR for OS compared to AT was ATPtPD1 (Table 2).

3.5. Inconsistency tests and heterogeneity analysis and small sample effect tests
There was no heterogeneity between studies included in the present Bayesian NMA and network regression analysis (Figures S2–S9, Supplemental Digital Content, https://links.lww.com/MD/R101). Small sample effect was explored by network funnel plot. P < .05 was considered to be statistically significant (Figures S10–S12, Supplemental Digital Content, https://links.lww.com/MD/R101).

4. Discussion
In recent years, various advanced diagnostic technologies have emerged, making more and more breast cancer found at early stagey.[54,55] Early breast cancer can improve the OS of the overall diseased population, especially in the hormone receptor positive and human epidermal growth factor receptor 2 positive subgroups. Currently, TNBC lacks targeted therapeutic targets, resulting in limited overall efficacy. TNBC is an aggressive form of breast cancer that is often associated with poor patient outcomes, largely due to the limited treatment options available.[56,57] Most TNBC patients with early-stage are now treated with neoadjuvant chemotherapy. This approach, primarily driven by the significant prognostic benefit associated with pCR, particularly in patients with more aggressive subtypes, has become the standard of treatment. New treatment methods, such as immunotherapy, have achieved definite results in the application of TNBC, improving EFS and OS.[58] Therefore, in the treatment of early TNBC, immunotherapy has been seeking a breakthrough. Some of these new treatments have significantly improved pCR, but have not effectively improved EFS or OS.[59] Such treatments do not fully comply with the principles and objectives of tumor treatment, so we do not recommend them. The aim of the present study was to discuss the best options for neoadjuvant chemotherapy of TNBC.
To the best of our knowledge, the present study reports the first Bayesian NMA and regression analysis comparing relative efficacy of all current available neoadjuvant therapies for TNBC. The findings were as follows.

4.1. Core findings
For pCR, only ATPtPD1 can improve results compared with AT.
For EFS with a follow-up period of 12th to 48th month, ATPtPD1 demonstrated a significant advantage from 18th to 48th month compared with AT. The significant result of ATPt also showed, but only between 36th and 48th month after initial treatment.
For OS with a follow-up period of 12th to 48th month, when compared with AT, ATPtPD1 and ATPt both showed a significant efficacy only on the 18th month, while there is no improvement on the 12th and 24th months. However, at subsequent follow-up times, significant efficacy was found for ATPtPD1 extending from 30th to 48th months, whereas ATPt only extended from 30th to 42nd months.
For HR of EFS: ATPtPD1, ATPt, and ATPtPARPi demonstrated a significant advantage compared with AT. After regression analysis using follow-up time as a correcting factor, the significant improvement did not change significantly. However, from the time window analysis, ATPtPD1 showed significant advantage in a longer time window of 0 to 104 months after the initiation of treatment. While the time window for ATPt was 15 to 85 months, for ATPtPARPi it was 18 to 72 months.
For HR for OS: only ATPtPD1 significantly improves HR for OS compared with AT. Similarly, after regression analysis using follow-up time as a correcting factor, the significant improvement does not change significantly. ATPtPD1 was the only regimen that shows significant efficacy in pCR rate, EFS/OS during the follow-up period of 18 to 48 months after treatment, and HR for EFS/OS.

4.2. Clinical-pathological correlation
Anthracyclines insert themselves between DNA base pairs (intercalation), disrupting the DNA helix. This interference inhibits essential processes like DNA replication and RNA transcription, critical for rapidly dividing cancer cells. Anthracyclines undergo redox cycling, producing reactive oxygen species (ROS) such as superoxide radicals. These ROS induce oxidative stress, damaging DNA, proteins, and lipids. Cancer cells, often with weaker antioxidant defenses, are particularly vulnerable.[60] Iron-anthracycline complexes can exacerbate ROS production, contributing to both efficacy and side effects like cardiotoxicity.[60] Recent research explores strategies to mitigate toxicity while enhancing efficacy, such as liposomal formulations or targeting tumor-specific pathways.[61]
The antitumor mechanisms of taxanes (e.g., paclitaxel, docetaxel, cabazitaxel) primarily involve disrupting microtubule dynamics, which are critical for cell division and survival. Taxanes bind to β-tubulin subunits of microtubules, stabilizing them and preventing their disassembly (depolymerization). This disrupts the normal dynamic instability of microtubules, a process required for proper mitotic spindle formation during cell division. Cells become trapped in metaphase arrest (mitotic arrest), unable to complete chromosome segregation. Prolonged mitotic arrest triggers apoptosis (programmed cell death).[62] Taxanes also affect interphase microtubules, interfering with critical cellular functions in nondividing cells, such as intracellular transport (e.g., organelle movement, vesicle trafficking). Cells signaling pathways dependent on microtubule integrity. This broad disruption contributes to tumor cell death even in slowly proliferating cancers.[63]
The antitumor mechanisms of platinum-based chemotherapy drugs (e.g., cisplatin, carboplatin, oxaliplatin, lobaplatin) primarily involve DNA damage and disruption of cellular processes, leading to apoptosis (programmed cell death). The activated platinum binds to purine bases (mainly guanine) in DNA, including intrastrand crosslink and interstrand crosslink.[64] Finally, damaged DNA stalls replication forks, triggering cell cycle arrest (G2/M phase). Platinum-induced tumor cell death can release damage-associated molecular patterns, activating dendritic cells and enhancing antitumor immunity.[64,65]
The antitumor mechanism of PD-1/PD-L1 inhibitors (e.g., pembrolizumab, camrelizumab, atelizumab) involves disrupting immune evasion strategies employed by cancer cells, thereby reactivating T-cell-mediated tumor destruction.[66] PD-1 is a checkpoint receptor on activated T-cells that suppresses their activity when engaged. PD-L1 is a ligand often overexpressed by tumor cells and stromal cells in the tumor microenvironment, which binds to PD-1 to inhibit T-cell function.[67] Deregulation of T-cell inhibition as a core antitumor mechanism in immunotherapy. Tumors exploit the PD-1/PD-L1 pathway by upregulating PD-L1 expression. When PD-L1 binds to PD-1 on tumor-infiltrating T-cells, it delivers a co-inhibitory signal. After reducing proliferation, cytokine production, and cytotoxic activity, T-cell function is restored and kills tumor cells directly.[67]
PD-1 acts as a receptor and receives inhibitory signals directly from PD-L1/PD-L2. PD-L1 acts as a ligand and transmits inhibitory signals by binding to PD-1. Tumor cells directly inhibit T cell function through high expression of PD-L1, while inducing immunosuppressive activity of regulatory T cells. PD-1 inhibitors (e.g., pembrolizumab, camrelizumab) directly block the PD-1 receptor, preventing it from binding to PD-L1/PD-L2, and may also affect the 2 inhibitory pathways, PD-L1 and PD-L2. PD-L1 inhibitors (e.g., atelizumab) block PD-L1 only, preserving PD-L2 interaction with PD-1 (may reduce some autoimmune side effects). This is more specific for tumors with high PD-L1 expression.[68] PD-1 inhibitors may block the dual binding of PD-1 to PD-L1/PD-L2, leading to more extensive immune activation and a slightly higher incidence of side effects such as pneumonia and thyroiditis. PD-L1 inhibitors block only PD-L1, preserving PD-L2 binding to PD-1 (PD-L2 is expressed in some normal tissues), which may theoretically reduce some side effects, but the clinical differences are not yet significant.[69]

4.3. Feasibility analysis
Chemotherapy containing anthracycline, taxanes, and platinum has been applied to breast cancer for a long time. The most common toxic side effects are bone marrow suppression, nausea and vomiting, and fatigue.[70] Colonizing human granulocyte stimulating factor is effective in ameliorating neutropenia induced during chemotherapy.[71] And nausea and vomiting are well controlled after application of a standard antiemetic protocol.[72] According to relevant trails, treatments combined with PD1 inhibitors did not increase in chemotherapy related toxic side effects. Only exhibiting adverse reactions related to immunotherapy like rash, interstitial pneumonia, and thyroid dysfunction. But all of adverse reactions related to immunotherapy were controllable and relievable.[73] To avoid excessive accumulation of toxic side effects, sequential administration can be used in order to be better accepted among patients.[74]

4.4. Limitations
First, the sample sizes of certain included studies were inadequate, resulting in the small sample effect and potential bias. Second, in the analyses of HR EFS/OS, certain studies had not reported 95% CIs or the interquartile range. These studies had to be excluded, which potentially shrank the sample size by another means, eventually increasing random error. Third, the limited resolution of survival curve images in certain studies was compromised. Finally, the quality of some of the studies was low, bringing potential interference.

4.5. Perspective
We hope that the design of future clinical trials will be more precise and the final OS data will be reported. The data from subgroup analysis should be more detailed. Adverse events should be evaluated after long-term follow-ups and could be compared at each time node. When making clinical decisions, adverse effects should be considered. FDA Adverse Event Reporting System database can provide some information on these adverse events.

5. Conclusion
In conclusion, considering the benefit of treatment on pCR rate and EFS/OS emerged earlier and over a long period of time, ATPtPD1 should be recommended as the optimal neoadjuvant therapy in TNBC. However, it is necessary to design more RCTs to confirm this result.

Author contributions
Conceptualization: Yicheng Jiang.
Data curation: Wenbo Deng, Meng Xu, Qiang Wang.
Formal analysis: Yicheng Jiang, Jiajia Zeng, Meng Xu.
Methodology: Jian Liu, Ruijun Tang.
Software: Yicheng Jiang, Jiajia Zeng, Ruijun Tang, Qiang Wang.
Writing – original draft: Wenjie Shi.
Writing – review & editing: Jian Liu, Ulf Dietrich Kahlert, Wenjie Shi, Qiang Wang.

Supplementary Material