During the systematic review process, we adhered to the PRISMA guidelines (12) (Supplementary Table 1). Four electronic databases (PubMed, Embase, Web of Science, and the Cochrane Library) were searched on December 18, 2024, without applying any time restrictions. The search strategy employed was: ((“Carcinoma, Non-Small-Cell Lung” [Mesh] OR “non-small cell lung cancer” [Title/Abstract] OR NSCLC [Title/Abstract]) AND (“Glucocorticoids” [Mesh] OR glucocorticoid * [Title/Abstract] OR corticosteroid * [Title/Abstract]) AND (“Immune Checkpoint Inhibitors” [Mesh] OR “immune checkpoint inhibitor*” [Title/Abstract] OR “PD-1 inhibitor*” [Title/Abstract] OR “PD-L1 inhibitor*” [Title/Abstract] OR “CTLA-4 inhibitor*” [Title/Abstract])). These keywords were selected based on the PICO framework to ensure a comprehensive retrieval of studies pertinent to the meta-analysis. No language restrictions were applied, and the reference lists of pertinent articles were manually screened for additional records.

To ensure transparency and reproducibility, the study selection process was conducted in three stages. First, all records retrieved from the databases were imported into EndNote X9, and duplicate entries were removed. Second, two reviewers independently screened titles and abstracts to exclude irrelevant studies, reviews, case reports, editorials, conference abstracts, and studies lacking original data. Third, the full texts of the remaining articles were assessed against the predefined inclusion and exclusion criteria. Studies were excluded at this stage if they involved mixed cancer types without extractable NSCLC-specific data, failed to provide sufficient information on glucocorticoid timing, dosage, or indication, or presented overlapping patient cohorts. In cases of overlapping reports, the study with the most comprehensive and updated dataset was retained. Any discrepancies were resolved through discussion, with a third reviewer consulted when necessary. The data extracted included the first author, publication year, country, sex ratio, age range, sample size, details of the ICI treatment regimen, details of the glucocorticoid treatment regimen, and outcome measures. When hazard ratios (HRs) and 95% confidence intervals (95%CIs) could not be directly obtained, they were calculated from survival curves using the method described by Tierney et al. (13). Additionally, when the published report did not provide the data of interest, the investigators of the original study were contacted via email to request the unpublished information.

The quality of the studies included in the meta-analysis was assessed independently by two reviewers using the Newcastle-Ottawa Scale (NOS) (14). This established tool evaluated each study across nine items distributed among three key domains: selection, comparability, and outcome. These domains were employed to identify potential biases inherent in the studies. In the scoring process, one point was assigned for each asterisk noted in the evaluation table, resulting in a total score ranging from 0 to 9. Studies with scores between 0 and 3 were classified as low quality, those with scores from 4 to 6 were considered moderate quality, and studies scoring between 7 and 9 were deemed high quality.

Statistical analyses were performed using Stata version 17 (StataCorp, College Station, TX, United States). Initially, heterogeneity among the studies was evaluated using chi-square statistics and quantified by the I² value. When the I² was less than 50% and the corresponding p-value was ≥0.10, it was considered that no significant heterogeneity existed, and a fixed-effect model was employed to calculate the pooled effect size. Conversely, when the I² reached or exceeded 50% or the p-value was <0.10, significant heterogeneity was inferred, leading to the use of a random-effects model for effect size estimation (15). Subsequently, sensitivity analyses were conducted by sequentially omitting each study to assess the robustness of the overall effect size and to identify potential sources of heterogeneity. Publication bias was assessed using funnel plots and Egger’s regression test. In addition, exploratory meta-regression analyses were conducted using a random-effects model with restricted maximum likelihood (REML) estimation and Knapp–Hartung adjustment to examine whether study-level factors, such as timing of corticosteroid exposure, ICI class, or analytic methodology, contributed to between-study heterogeneity. All statistical tests were two-sided, and a p-value <0.05 was regarded as statistically significant.

To assess the quality of evidence, the GRADE approach was employed, evaluating key domains: Risk of Bias (study design, blinding, follow-up adequacy); Inconsistency (heterogeneity across studies); Indirectness (relevance to the research question); Imprecision (width of confidence intervals); Publication Bias (symmetry of funnel plots); Large Effect (significant effect sizes); and Dose–Response Gradient (association between dose and effect). Evidence was categorized based on these factors to determine confidence in the effect estimates and guide the interpretation of results.

At the inception of this systematic review and meta-analysis, an exhaustive search of multiple electronic databases yielded an initial set of 1,179 potentially relevant articles. Duplicate records were subsequently removed to ensure that each study was uniquely represented. Thereafter, titles and abstracts were screened against pre-established inclusion and exclusion criteria, which encompassed study design, demographic characteristics of the study population, clinical outcomes measured, and overall methodological quality. This preliminary screening resulted in 42 articles being selected for full-text review. Multiple investigators independently assessed the full texts to ensure an unbiased and comprehensive evaluation. During this phase, 27 articles were excluded for the following reasons: review articles (n = 11), sequentially published works (n = 6), studies with insufficient data for analysis (n = 7), and clinical trials lacking control groups (n = 3). Ultimately, 15 articles satisfied all the stringent criteria delineated in the research protocol and were included in the final meta-analysis (16–30) (Figure 1).

A total of fifteen studies were incorporated into the meta-analysis, with individual sample sizes ranging from 67 to 1,025 patients. The median ages across these studies spanned from 63 to 77 years, and the proportion of male participants varied between studies. All studies evaluated the concomitant use of glucocorticoids in patients receiving ICIs for cancer treatment. The ICIs employed were predominantly PD-1/PD-L1 inhibitors, although several studies also included CTLA-4 inhibitors either as monotherapy or in combination with PD-1/PD-L1 agents. Glucocorticoid administration protocols differed considerably among the studies. In some instances, glucocorticoids—primarily prednisone or prednisolone—were administered at the initiation of ICI therapy, whereas other studies reported their use within a 30-day window before or after the commencement of ICI treatment. Dosages varied widely, with some studies reporting relatively low doses (e.g., 6.5 mg) and others employing much higher doses (up to 280 mg or more). Furthermore, the indications for glucocorticoid use were heterogeneous, encompassing management of cancer-related symptoms, brain metastases, palliative care needs, and immune-related adverse events (Table 1).

Quality assessment was conducted using the Newcastle-Ottawa Scale, which evaluated each study based on selection, comparability, and outcome criteria. The total scores of the included cohort studies ranged from 7 to 9, with the majority achieving a score of 9, indicative of high methodological quality. A few studies scored 7 or 8 due to minor limitations in specific domains, such as the ascertainment of exposure or the adequacy of follow-up. Overall, the studies were deemed to be of moderate to high quality, thereby supporting the robustness and reliability of the meta-analysis findings (Table 2).

Thirteen studies contributed data on progression-free survival (PFS) comparing patients who received glucocorticoids with those who did not. Owing to significant heterogeneity among these studies (I² = 77.7%, p < 0.001), a random-effects model was applied to pool the results. The combined analysis yielded a pooled HR of 1.44 (95% CI, 1.15–1.73; p < 0.001), indicating that glucocorticoid use was associated with a significantly increased risk of disease progression relative to non-use (Figure 2). To ensure the robustness of this finding, a sensitivity analysis was performed by sequentially excluding each study. This leave-one-out analysis confirmed that the overall PFS result remained stable and robust, demonstrating that the pooled effect estimate was not unduly influenced by any single study (Figure 3).

Fourteen studies were included in the evaluation of overall survival (OS) outcomes based on glucocorticoid exposure. Due to considerable heterogeneity across these studies (I² = 84.2%, p < 0.001), a random-effects model was employed to synthesize the data. The pooled analysis revealed a HR of 1.58 (95% CI, 1.24–1.93; p < 0.001), signifying those patients receiving glucocorticoids experienced a markedly higher risk of mortality compared to those who did not (Figure 4). A sensitivity analysis, conducted by sequentially removing each study from the analysis, confirmed the stability of the overall OS estimate. This rigorous evaluation demonstrated that the observed association between glucocorticoid use and reduced OS was consistently robust despite the exclusion of any individual study (Figure 5).

A subgroup analysis was performed to evaluate the impact of glucocorticoid administration timing on survival outcomes in NSCLC patients treated with ICIs (Table 3). In the subgroup where glucocorticoids were administered before ICI initiation, five studies were included, showing pooled HRs of 1.21 (95% CI, 0.85–2.01; p = 0.15) for PFS and 1.31 (95% CI, 0.69–2.28; p = 0.39) for OS, indicating no significant association with adverse survival. By contrast, three studies assessing glucocorticoid administration at ICI initiation demonstrated significant associations with poorer outcomes, with pooled HRs of 1.38 (95% CI, 1.08–2.15; p < 0.001) for PFS and 1.56 (95% CI, 1.15–2.01; p < 0.001) for OS. Furthermore, in the subgroup where glucocorticoids were administered after ICI initiation, five studies consistently reported significantly worse outcomes, with pooled HRs of 1.98 (95% CI, 1.51–2.62; p < 0.001) for PFS and 2.28 (95% CI, 1.61–3.41; p < 0.001) for OS. Collectively, these findings suggest that glucocorticoid use, particularly when administered at or after ICI initiation, is significantly associated with an increased risk of disease progression and mortality, whereas pre-ICI use does not appear to exert a statistically significant impact on survival.

In addition to the leave-one-out sensitivity analyses, we further conducted subgroup-based exclusion analyses to assess the robustness of the pooled results. Specifically, we sequentially excluded all studies categorized as pre-ICI use, at-ICI initiation use, or post-ICI use. The exclusion of each subgroup did not materially alter the overall direction of the association between glucocorticoid exposure and survival outcomes. For progression-free survival, pooled HRs remained within the range of 1.39–1.51 across exclusion scenarios, and for overall survival, pooled HRs ranged from 1.52–1.63. Although minor variations in effect size were observed, the associations consistently indicated that glucocorticoid use was linked to poorer outcomes. These findings reinforce that no single subgroup of studies disproportionately influenced the overall effect estimates, thereby supporting the robustness of our conclusions.

Feasibility thresholds were met for timing of exposure and, for a subset of studies, ICI class; other moderators (analytic method, indication, dose category) were insufficiently reported across studies to support stable modeling. In timing-based models (reference: pre-ICI exposure), post-ICI corticosteroid use showed a directionally positive association with both PFS and OS (larger HRs), consistent with our subgroup findings; however, coefficients did not reach statistical significance after Knapp–Hartung adjustment, and the reduction in between-study variance was modest. Models incorporating ICI class did not materially alter pooled effects and explained little heterogeneity. Due to sparse and heterogeneous reporting, meta-regression by analytic approach, clinical indication, or dose category was underpowered and therefore not undertaken in the main analysis. Overall, exploratory meta-regression did not identify a robust study-level moderator that fully accounted for the observed heterogeneity.

Visual inspection of the plots revealed a symmetric distribution of effect sizes, indicating that no significant publication bias was detected (Figure 6). Consistently, Egger’s regression test did not demonstrate significant small-study effects for either PFS (p = 0.42) or OS (p = 0.37), further suggesting that publication bias was unlikely. Although funnel plots and Egger’s test have inherent limitations when the number of studies is moderate, the current findings indicate that publication bias is unlikely to have substantially influenced the results of this meta-analysis.

Using the GRADE approach, we systematically evaluated the certainty of evidence for key outcomes in the impact of glucocorticoid administration on OS and PFS in patients receiving ICIs (Figure 7).

The evidence for OS was assessed as Moderate due to significant concerns regarding inconsistency and imprecision despite a large effect and low risk of bias. The HR for OS in patients receiving glucocorticoids compared to those who did not was 1.58 (95% CI: 1.24 to 1.93), indicating a higher risk of mortality associated with glucocorticoid use. Although statistically significant, the broad confidence intervals reflect considerable uncertainty, necessitating cautious interpretation.

For PFS, the quality of evidence was also rated as Moderate, influenced by inconsistency across studies and imprecision in effect estimates. However, the evidence showed a clear large effect with no publication bias. The Hazard Ratio (HR) for PFS associated with glucocorticoid administration compared to non-use was 1.44 (95% CI: 1.15 to 1.73), suggesting that glucocorticoid use was significantly associated with an increased risk of disease progression.

Both outcomes were deemed to have Moderate quality evidence, with the effect sizes providing statistically significant findings. However, the broad confidence intervals and moderate heterogeneity reduce the certainty of the estimates, urging caution in the interpretation of these results.