This study's protocol was duly registered with the International Prospective Register of Systematic Reviews (registration no. CRD42023458837) (14) and was executed in strict compliance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines (15).

To identify relevant studies, a systematic search was conducted in the PubMed, EMBASE, Web of Science and Cochrane Library (16) databases. All prospective randomized controlled trials (RCTs) and retrospective comparative cohort studies evaluating frontline rituximab therapy for PCNSL were included, with publication dates spanning from database inception to May 5, 2025. Eligible studies were required to report clinical response or survival endpoints.

The search employed a combination of keywords, subject terms and free-text terms to ensure a comprehensive retrieval of pertinent literature. The simplified search string utilized the following terms: ‘Rituximab’ OR ‘rituximab’ OR ‘CD20 monoclonal antibody (rituximab)’ AND ‘primary central nervous system lymphoma’ OR ‘PCNSL’ OR ‘PCNSLs’. The specific search formula is provided in detail in Table SI.

The inclusion criteria were as follows: i) Studies should be high-quality retrospective analyses or RCTs; ii) studies should compare a rituximab-containing treatment protocol vs. the same treatment protocol without rituximab; iii) studies should provide complete data sets for analysis; and iv) a minimum sample size of 10 cases was required.

The following exclusion criteria were applied: i) Editorials, single-arm trials, reviews, case reports and animal studies were excluded; ii) studies on recurrent or secondary CNSL (SCNSL) were not considered; and iii) studies with incomplete data were excluded.

A total of two authors (LCYS and WSL) extracted data by using one of three methods: Direct extraction from the original text, contacting original authors for permission to obtain raw data and utilizing Engauge Digitizer Version 4.1 (17) to extract data from Kaplan-Meier (K-M) survival curves if necessary. For studies that directly reported hazard ratios (HRs) with 95% CIs, HRs were extracted from the published values. For studies that did not report HRs explicitly, the validated survival curve reconstruction method established by Tierney et al (18) in 2007 was applied to estimate HRs and their corresponding standard errors. This method accounts for the number of patients at risk and the number of events at each time interval via the following workflow: i) Engauge Digitizer Version 4.1 was used to digitize the published K-M curve and extract numerical coordinates of survival probability at multiple discrete follow-up time-points; ii) combined with the numbers at risk and event counts reported at each time interval in the original study, the algorithm reconstructed the number of events and censored observations within each follow-up interval; iii) the observed minus expected number of events and its variance were calculated for each treatment arm based on log-rank theory, from which the hazard ratio and its 95% CI were finally derived; iv) the uncertainty associated with indirectly estimated HRs was quantified using the standard error derived from the K-M curve reconstruction. Raw individual patient-level survival data were obtained by contacting the corresponding authors of Dina et al (19) from 2019; these data were used directly for HR calculation without curve digitization.

The quality of RCTs was assessed using the Cochrane risk of bias (RoB) tool (20). Retrospective studies were evaluated with the Newcastle-Ottawa scale (NOS), which assigns a total score of 0–9, with studies scoring ≥5 considered high quality (21). Quality assessments were conducted independently by two researchers (LCYS and WSL), with a third researcher (DXW) involved in cases of disagreement. Only studies meeting predefined quality thresholds were incorporated into the meta-analysis.

Quantitative meta-analyses were conducted using Review Manager 5.3 (The Cochrane Collaboration) and Stata16 software (StataCorp LP) (22) for statistical analysis. The Eastern Cooperative Oncology Group performance status (ECOG-PS) was used to assess the baseline status of patients before participating in the clinical trials, with lower ECOG-PS scores indicating better patient status (23).

The primary outcome measures focused on efficacy indicators, namely OS, PFS and complete remission (CR). Given that certain studies only reported ~3-year OS and PFS following rituximab treatment, this study conducted a binary variable analysis of patients achieving 3-year OS and 3-year PFS to investigate whether the addition of rituximab benefits patients in reaching these milestones. The clinical efficacy of PCNSL treatment, with or without rituximab, was evaluated using the OR along with the 95% CI. The secondary outcome measure was the incidence of adverse drug reactions (ADRs) and serious adverse events (SAEs; Grade ≥III), which were primarily quantified using the OR.

OS was defined as the time from randomization (or treatment initiation in single-arm trials) to death from any cause (24). PFS was defined as the time from randomization (or treatment initiation in single-arm trials) to tumour progression or death from any cause, whichever occurred first (25). According to the International PCNSL Collaborative Group (IPCG) criteria for evaluating PCNSL efficacy, CR is defined as no detectable lesions on enhanced imaging, normal ophthalmologic examination results and negative cerebrospinal fluid cytology (26,27).

All pooled analyses were conducted using a random-effects model, regardless of the magnitude of the I² statistic. Heterogeneity across included studies was quantified using the I² statistic test. The heterogeneity threshold was predefined as follows: Low heterogeneity (I²≤50%, P≥0.10) and substantial heterogeneity (I²>50%). Subsequently, sensitivity analysis and subgroup analysis were conducted to further explore and explain the observed heterogeneity. Furthermore, funnel plots (Fig. S1) were employed to assess the potential for publication bias within the included studies. A two-sided P-value <0.05 was considered statistically significant. All pooled results were graphically displayed as forest plots.

The initial literature search retrieved a total of 653 unique citations. After the removal of duplicates and the exclusion of reviews and case reports, a total of 101 papers remained under consideration. Following a thorough review of the titles and abstracts, 61 studies were excluded, narrowing down the selection to 40 full-text references. Upon further assessment of these full-text references, studies were excluded based on several criteria: The absence of a control group in single-arm studies, the exclusion of basic and animal studies and the non-fulfilment of data requirements for experimental interventions. This meticulous evaluation process culminated in the inclusion of 15 studies in the systematic review. The PRISMA flow diagram, illustrating the study selection process, is presented in Fig. 1.

The study included 15 articles, comprising 3 RCTs and 12 retrospective trials. Details of the study characteristics were presented in Table I. The three RCTs, including HOVON105/ALLGNHL24 (28) (ACTRN12610000908033), IELSG32 (11) (NCT01011920) and Luo et al (29) from 2016, were complemented by cohort studies. Studies such as Luo et al (29) from 2016, Chen et al (30) from 2019, Dina et al (19) from 2019, Mou et al (31) from 2017, Sun et al (32) from 2017, Patekar et al (33) from 2019 and Pang et al (34) from 2021, have investigated the impact of rituximab on individuals of Asian descent. Conversely, ALLGNHL24 (28), IELSG32 (11), Birnbaum et al (35) from 2012, Gregory et al (36) from 2013, Houillier et al (37) from 2017, Kansara et al (38) from 2015, Holdhoff et al (39) from 2014 and Mocikova et al (40) from 2016 mainly focused on studying the effects of rituximab on Caucasians.

The mean age of the study population exceeded 50 years, with a predominance of elderly patients (Table I). Except for Luo et al (29) from 2016 and Gregory et al (36) from 2013, most studies exhibited a higher proportion of male participants (>50%). The proportion of patients with ECOG-PS ≤1 in the rituximab group was like that in the control group, indicating comparable baseline characteristics. The treatment regimens in this study encompassed methotrexate-based protocols for both control and experimental cohorts, potentially augmented with additional therapeutic modalities such as whole brain radiotherapy and autologous stem cell transplantation. Furthermore, based on specific experimental designs and groupings, patients were assessed for the addition of rituximab to these treatments. Most literature reported OS, PFS and CR rates, while ADRs were documented across all studies. Due to the different outcome indicators reported in various trials, some lacked data for certain metrics, such as PFS-HR and OS-HR results from 5 trials, 3-year OS from 10 trials, 3-year PFS from 9 trials and CR results from 12 trials.

Funnel plots were used to assess publication bias for effectiveness and safety indicators separately (Fig. S1), and their funnel plots were generally symmetrical, with no evidence of publication bias. Funnel plots for 3-year OS and 3-year PFS were generally symmetrical with no obvious asymmetry. Egger's linear regression test for 3-year OS yielded P=0.32, and Begg's rank correlation test gave P=0.47. For 3-year PFS, Egger's test yielded P=0.41 and Begg's test P=0.59. These results suggest no significant evidence of publication bias for the primary outcomes. Nevertheless, given the relatively small number of studies included in certain analyses, the possibility of unpublished negative trials cannot be entirely excluded. First, omission of studies with null or unfavorable effects may overestimate the pooled survival benefit of rituximab; inclusion of missing negative trials could potentially attenuate the effect size and eliminate the statistical significance of primary outcomes. Second, conventional publication bias tests have limited power with small study sets. Accordingly, these conclusions should be interpreted with caution, and further large, well-designed multicenter RCTs are required to confirm the survival benefit of rituximab in PCNSL. The detailed Egger's and Begg's test results are presented in Table SIII.

For this study, the 3-year PFS data were primarily extracted from 9 studies: ALLGNHL24 (28), Birnbaum et al (35) from 2012, Chen et al (30) from 2019, Holdhoff et al (39) from 2014, IELSG32 (11), Kansara et al (38) from 2015, Mocikova et al (40) from 2016, Mou et al (31) from 2017 and Patekar et al (33) from 2019. Additionally, the HR of PFS was derived from an analysis of 5 studies: ALLGNHL24 (28), Dina et al (19) from 2019, IELSG32 (11), Kansara et al (38) from 2015 and Pang et al (34) from 2021. As shown in Fig. 4A, the combined HR after applying the random-effects model was 0.54 (95%CI: 0.37–0.79, P=0.002), suggesting a heterogeneous yet robust and high-certainty benefit of incorporating rituximab into the treatment regimen for patients with PCNSL. Furthermore, the results, as depicted in Fig. 4B, indicated that the addition of rituximab significantly improved the 3-year PFS in patients with PCNSL (OR=2.30, 95% CI: 1.47–3.58, P=0.0002).

In the present study, an analysis of CR rates was performed. The results from 12 studies were pooled and are illustrated in Fig. 5. The inclusion of rituximab in the treatment protocol was associated with a significantly higher CR rate among patients with PCNSL (OR=2.04, 95% CI: 1.50–2.77, P<0.00001, I²=13%), indicating a substantial and statistically significant benefit.

The safety of rituximab in treating patients with PCNSL was comprehensively assessed for the following grade ≥3 adverse events: Neutropenia, leukopenia, thrombocytopenia, infection, hepatic toxicity, renal toxicity, neurotoxicity and anemia. The results are summarized in Tables II and SIV. Study-level availability of grade ≥3 SAE data is summarized in Table SIV. Of the 8 included studies, 3 studies [ALLGNHL24 (28), IELSG32 (11), Kansara et al (38) from 2015] reported complete data across all 8 SAE categories. Pooled analyses for all grade ≥3 SAE endpoints showed no statistically significant differences between the rituximab and control groups. Figs. S2 and S3 present the grade ≥3 SAEs from the combined studies.

For neutropenia ≥ grade 3, the OR was 1.06 (95% CI: 0.65–1.73, P=0.83; I²=0%), indicating no significant difference. For leukopenia ≥ grade 3, the OR was 1.74 (95% CI: 0.92–3.27, P=0.09; I²=0%), showing a trend toward increased risk in the rituximab group but not reaching statistical significance. Thrombocytopenia ≥ grade 3 showed an OR of 1.26 (95% CI: 0.80–1.99, P=0.32; I²=0%) and Anemia ≥ grade 3 and OR of 0.88 (95% CI: 0.34–2.25, P=0.79; I²=63%). Forest blots for specific hematological SAEs are provided in Fig. S2.

Infection ≥ grade 3 had an OR of 0.96 (95% CI: 0.60–1.51, P=0.84; I²=0%). Hepatic toxicity ≥ grade 3 had an OR=0.81 (95% CI: 0.46–1.41, P=0.45; I²=45%) and renal toxicity ≥ grade 3 an OR of 1.43 (95% CI: 0.63–3.27, P=0.39; I²=0%). None of the grade ≥3 SAEs reached statistical significance, indicating that rituximab did not substantially increase the risk of SAEs compared to chemotherapy alone. Forest blots for non-hematological SAEs are provided in Fig. S3.

For any-grade adverse events (Fig. 6), the comparative analysis revealed no significant differences in the incidence of gastrointestinal toxicity (OR=1.21, 95% CI: 0.47–3.11), thrombocytopenia (OR=1.22, 95% CI: 0.79–1.87) and neurotoxicity (OR=0.72, 95% CI: 0.45–1.15) between rituximab users and controls. An elevated risk of hepatic toxicity was observed but it did not reach statistical significance (OR=1.42, 95% CI: 0.98–2.05, P=0.06).

The study populations exploring the role of rituximab in PCNSL have focused on Asians and Caucasians, so in this paper, the 3-year OS and 3-year PFS in which heterogeneity exists were stratified by ethnicity, and the specific results were shown in Fig. 7 and Table SV. In this study, the Asian subgroup comprises populations from East Asia (Han Chinese) and South Asia (Indian subcontinent).

In the 3-year OS assessment, a combined analysis of four studies showed an OR of 2.62 (95% CI: 1.16–5.91, P=0.02) for Asians and a combined analysis of six studies showed an OR of 1.69 (95% CI: 1.23–2.32, P=0.001) for Caucasians; for the 3-year PFS, a combined analysis of three studies showed an OR of 2.98 (95% CI: 1.17–7.60, P=0.02) for Asians, while a six-study meta-analysis yielded an OR of 2.02 (95% CI: 1.22–3.34, P=0.006) for Caucasians, suggesting that rituximab may have more pronounced efficacy in Asians compared to Caucasians.

This study evaluated the efficacy of rituximab in patients with PCNSL based on study type. As illustrated in Fig. 8A and B, for OS, the HR for cohort studies was 0.38 (95% CI: 0.16–0.88, P=0.02), while the HR for RCTs was 0.76 (95% CI: 0.52–1.12, P=0.17). The I² statistic between groups was found to be 54.9%, indicating no significant difference between them (P=0.14). Regarding PFS, the HR for cohort studies was 0.39 (95% CI: 0.16–0.93, P=0.03), whereas the HR for RCTs was 0.63 (95% CI: 0.43–0.92, P=0.02), with an I² value of 0% between groups (P=0.32). These findings suggest that rituximab may demonstrate superior efficacy in retrospective cohort studies compared to RCTs; however, this difference is not statistically significant across groups.

Sensitivity analyses were performed by sequentially excluding individual studies from each pooled analysis. For 3-year OS, the pooled OR ranged from 2.01 to 2.51 upon sequential exclusion and no single study substantially altered the overall effect direction or statistical significance. For PFS-HR, exclusion of the Pang et al (34) from 2021 study (which reported the most extreme HR of 0.20) resulted in a pooled HR of 0.67 (95% CI: 0.50–0.89, P=0.005), still indicating significant benefit. Sensitivity analyses excluding studies with NOS scores <7 yielded consistent results: 3-year OS OR=2.36 (95% CI: 1.42–3.93, P=0.0008) and 3-year PFS OR=2.14 (95% CI: 1.54–2.97, P<0.0001). Excluding HRs estimated from K-M curves [Dina et al (19) from 2019 and Kansara et al (38) from 2015], the OS-HR was 0.56 (95% CI: 0.39–0.82, P=0.003). Detailed sensitivity analysis results are presented in Table SVI.

Table SVII provides a comprehensive summary of heterogeneity statistics and the effect model used for each pooled analysis. All meta-analyses in this study used the random-effects model, regardless of the I²-value found. Substantial heterogeneity (I²>50%) was observed for OS-HR (I²=57%), 3-year PFS (I²=58%), PFS-HR (I²=53%) and anemia (I²=63%), necessitating the use of random-effects models. Low heterogeneity (I²≤50%) was observed for 3-year OS (I²=33%), hepatic toxicity (I²=45%), CR (I²=13%), neutropenia (I²=0%), leukopenia (I²=0%), thrombocytopenia (I²=0%), infection (I²=0%) and renal toxicity (I²=0%).

For efficacy outcomes (3-year OS, OS-HR, 3-year PFS and PFS-HR), the observed between-study heterogeneity was primarily attributed to differences in study design (RCTs vs. retrospective cohorts) and variations in the ethnic composition of included populations. For the safety endpoint of grade ≥3 anemia, the substantial heterogeneity was likely driven by inconsistent adverse event grading and reporting standards across the included studies.