We performed comprehensive literature searches in PubMed/MEDLINE, Embase, Web of Science, and the Cochrane Central Register of Controlled Trials (CENTRAL) from their inception until June 30, 2025. All databases were updated to this same cut-off date to ensure consistency across the search process. Searches were also extended to clinical trial registries, including ClinicalTrials.gov and the WHO International Clinical Trials Registry Platform (ICTRP). To ensure thorough coverage, reference lists of all eligible studies and relevant systematic reviews were manually screened for additional records.

The search strategy incorporated both controlled vocabulary (MeSH, Emtree) and free-text keywords encompassing three key domains: tuberculosis, vaccination/exposure cohorts, and antigen-specific T-cell cytokine readouts. The core search structure was: (“tuberculosis” OR “Mycobacterium tuberculosis”) AND (“vaccine” OR “vaccination” OR “trial” OR “cohort” OR “prospective”) AND (“IFN-gamma” OR “interferon-gamma” OR “IL-2” OR “interleukin-2” OR “TNF” OR “TNF-alpha” OR “polyfunctional” OR “intracellular cytokine staining” OR “ELISpot”) AND (“correlate*” OR “risk” OR “protection” OR “progression” OR “incident TB” OR “IGRA conversion” OR “QuantiFERON”). Complete search strings for each database are provided in Supplementary Table S1.

Studies were included if they met the following criteria: (1) Population: Human participants of any age, including vaccinated cohorts (BCG or investigational TB vaccines) or longitudinal cohorts with documented exposure to or infection with M. tuberculosis. (2) Exposure/Biomarker: Measurement of antigen-specific IFN-γ, IL-2, and/or TNF-α responses using validated assays such as ELISpot, intracellular cytokine staining (ICS) with flow cytometry, whole-blood stimulation assays, or equivalent methods, with explicit specification of the stimulating antigen (e.g., PPD, Ag85A, ESAT-6/CFP-10, M72). (3) Study Design: Randomized controlled trials (RCTs) with immunology substudies, prospective cohort studies, or nested case-control studies. (4) Outcomes: The primary outcome was progression to microbiologically or clinically confirmed active TB disease. Secondary outcomes included sustained IGRA conversion or other trial-defined infection endpoints, such as persistent QuantiFERON positivity. (5) Data Availability: Studies must have reported sufficient data—such as hazard ratios (HR), odds ratios (OR), risk ratios (RR), or group-level counts enabling their calculation—to assess the association between cytokine measures and clinical outcomes. Narrative conclusions on such associations from prospective analyses were also considered.

We excluded animal studies, purely cross-sectional studies without longitudinal outcome data, studies reporting only immunologic correlations without linkage to disease or infection endpoints (unless used for contextual narrative synthesis), and studies lacking adequate details on antigen stimulation protocols or cytokine measurement methodology.

Study selection was performed independently by two reviewers, who first screened titles and abstracts, followed by full-text assessment. Any disagreements were resolved through discussion or, if necessary, by a third reviewer. Data were extracted using a standardized, pilot-tested form. The extraction process was carried out in duplicate by two independent reviewers, with discrepancies reconciled by consensus.

Extracted data items included: study identifiers (author, year, setting); study design; population demographics and clinical context; vaccine or exposure details; specific antigens used; assay platform; timepoints of immunologic measurement; definitions of cytokine-positive or polyfunctional T cells (including gating strategies where available); outcome definitions; reported effect estimates (HR/OR/RR) with covariate adjustments; and key narrative conclusions when direct effect estimates were unavailable.

For randomized controlled trials, we employed a domain-based assessment aligned with the Cochrane Risk of Bias tool (RoB 2), evaluating randomization, deviations from intended interventions, missing outcome data, outcome measurement, and selective reporting. For observational studies (cohort and nested case-control designs), we used a structured approach analogous to ROBINS-I or the Newcastle-Ottawa Scale, focusing on confounding control, participant selection, exposure measurement, outcome ascertainment, missing data, and selective reporting. Two reviewers independently conducted risk-of-bias assessments, with final judgments reached by consensus.

The principal summary measures were odds ratios (ORs) for dichotomous outcomes and mean differences (MDs) for continuous cytokine levels, each reported with 95% confidence intervals. It is important to note that the thresholds for dichotomizing cytokine responses (e.g., positive vs. negative IFN-γ or IL-2 responses) were defined in each individual study based on the assay platform and methodological conventions. These thresholds were not standardized across all studies, and variability in their application could influence the comparability of binary results. As such, comparisons across studies must consider these potential differences in dichotomization criteria.

We conducted random-effects meta-analyses when at least three clinically comparable studies assessed a similar cytokine construct and reported compatible effect measures. Random-effects models were fit using a restricted maximum likelihood (REML) estimator with Hartung–Knapp–Sidik–Jonkman adjustment for uncertainty in the pooled estimate. Statistical heterogeneity was summarized using the I² statistic.

To evaluate robustness and explore heterogeneity for the sustained IGRA conversion endpoint, we performed: (i) a leave-one-out influence analysis for the pooled IFN-γ association; (ii) subgroup meta-analysis by study design/population categories; and (iii) graphical heterogeneity/outlier diagnostics using Baujat and radial (Galbraith) plots. Small-study effects/publication bias were assessed descriptively by funnel plot inspection, recognizing that the number of contributing studies was limited and no formal asymmetry tests were emphasized.

Given heterogeneity in antigens, assay platforms, timepoints, and reporting metrics, outcomes not suitable for quantitative pooling were summarized using a structured narrative synthesis emphasizing direction, consistency, and clinical context (e.g., disease progression vs. infection endpoints; vaccine efficacy vs. immunogenicity).

Database searches identified 1, 268 records. After removing duplicates, 978 unique records were screened based on titles and abstracts. Of these, 82 full-text articles were assessed for eligibility, resulting in the inclusion of 10 studies that met all predefined criteria. A PRISMA 2020 flow diagram detailing the selection process is presented in Figure 1.

The ten included studies encompassed a range of designs and populations: two prospective infant cohorts evaluating post-BCG cytokine profiles in relation to subsequent TB disease; two prevention-of-infection randomized trials using sustained IGRA conversion as their primary endpoint (assessing H4:IC31 and BCG revaccination, including a later BCG revaccination trial); three vaccine efficacy RCTs (MVA85A in BCG-vaccinated infants, MVA85A in adults living with HIV-1, and M72/AS01E in adults); one exposure/cohort study focused on IFN-γ–independent immune signatures of M. tuberculosis exposure; and two contextual immunology studies. The latter two—one examining the relationship between polyfunctional PPD-specific T-cell frequencies and QuantiFERON magnitude, and another investigating trained-immunity pathways following adult BCG revaccination—were retained to aid interpretation but were not included in quantitative endpoint syntheses.

Risk-of-bias assessments for individual studies are displayed in Figure 2, with a domain-level summary provided in Figure 3. Most studies were rated as having a low risk of bias across most domains (typically 6 out of 10 studies per domain). The remaining studies were largely judged as having “some concerns” (commonly 2–4 studies per domain, depending on the domain). High-risk ratings were infrequent and primarily pertained to missing data (2 studies) and confounding (1 study), indicating that incomplete outcome data and residual confounding represent the principal threats to internal validity within this evidence base. Key characteristics of the included studies—including antigens used (e.g., PPD, Ag85A, M72), assay platforms (e.g., ICS, ELISpot, whole-blood), and immunology sampling timepoints—are summarized in Table 1.

In the primary analysis, which used progression to active TB disease as the clinical endpoint, associations between various antigen-specific Th1 immune markers and disease risk were generally inconsistent and close to the null. For IFN-γ, the random-effects pooled OR was 0.97 (95% CI 0.79–1.21), with negligible between-study heterogeneity (I² = 0%) (Figure 4A). Similarly, the pooled OR for IL-2 was 1.11 (95% CI 0.91–1.36; I² = 0%) (Figure 4B), and for TNF-α it was 1.01 (95% CI 0.84–1.22; I² = 0%) (Figure 4C). Analysis of polyfunctional T-cell responses (e.g., co-expressing IFN-γ, IL-2, and TNF-α) also yielded a pooled estimate near the null (OR = 1.04, 95% CI 0.85–1.27), with low heterogeneity (I² = 9.8%) (Figure 4D).

In the secondary analysis using sustained IGRA conversion as the endpoint, continuous measures of antigen-stimulated IFN-γ responses were slightly elevated among IGRA converters. The pooled MD was 0.07 (95% CI 0.03–0.10), with very low heterogeneity (I² = 1.2%) (Figure 4B). While the effect sizes observed are modest, these small differences in cytokine levels may still have biological relevance. Even slight elevations in cytokine responses, such as IFN-γ and IL-2, could reflect nuanced variations in immune activation associated with recent antigen exposure or infection status, particularly in the context of ongoing immune surveillance. When analyzed as a binary measure, IFN-γ responses showed a similar, albeit modest, trend toward association with infection (pooled OR = 1.13, 95% CI 0.94–1.36; I² = 0%) (Figure 4B). For IL-2, continuous measures were also marginally higher in converters (pooled MD = 0.06, 95% CI 0.01–0.11), though with moderate heterogeneity (I² = 60.0%) (Figure 4B). The binary analysis for IL-2 yielded a pooled OR of 1.07 (95% CI 0.85–1.33; I² = 0%) (Figure 4B).

Analyzed as continuous variables in relation to active TB disease progression, IFN-γ levels showed a pooled mean difference of 0.10 (95% CI 0.02–0.17), albeit with substantial heterogeneity (I² = 76.3%) (Figure 4C). For IL-2, the pooled MD was 0.06 (95% CI 0.02–0.09), also with considerable heterogeneity (I² = 62.9%) (Figure 4C). In contrast, the association for TNF-α was smaller (pooled MD = 0.02, 95% CI −0.02–0.07) and more consistent across studies (I² = 2.7%) (Figure 4C). Polyfunctional T-cell responses showed a pooled MD close to zero (0.00, 95% CI −0.06–0.06), with moderate heterogeneity (I² = 60.3%) (Figure 4C).

All analyses use inverse-variance random-effects models with 95% confidence intervals (CI) for each pooled estimate. Horizontal lines represent study-specific 95% CIs, squares represent study point estimates (size proportional to study weight), and diamonds represent pooled estimates. Heterogeneity is quantified using I² (reported within each panel). Across sustained IGRA conversion analyses, heterogeneity was low for IFN-γ (I² ≈ 0%) and moderate for IL-2 continuous measures (I² ≈ 60%). For progression to active TB disease, heterogeneity was low for binary markers (I² ≈ 0% across most panels), while substantial heterogeneity was observed for continuous IFN-γ (I² ≈ 76%) and IL-2 (I² ≈ 63%).

When assessing binary outcomes for sustained IGRA conversion, Th1 cytokine markers collectively showed a slight trend toward positive association, though confidence intervals for most pooled estimates crossed the null. For IFN-γ, the pooled OR was 1.13 (95% CI 0.94–1.36; I² = 0%) (Figure 5A). The corresponding estimates for IL-2 and TNF-α were 1.07 (95% CI 0.85–1.33; I² = 0%) and 1.08 (95% CI 0.89–1.31; I² = 0%), respectively (Figures 5B, C). Polyfunctional responses yielded a pooled OR of 1.12 (95% CI 0.92–1.38; I² = 0%) (Figure 5D).

Continuous cytokine readouts (reported as antigen-stimulated cytokine concentrations and/or frequencies of cytokine-positive cells) were generally higher among participants who experienced sustained IGRA conversion/infection (IGRA converters) compared with non-converters across polyfunctional responses, TNF-α, IL-2, and IFN-γ outcomes (Supplementary Figures S1A–D). Because included studies reported continuous outcomes on non-comparable scales (e.g., different units, background subtraction approaches, stimulation antigens, and assay platforms), we present these data as a structured supplemental synthesis rather than a single pooled estimate for each marker. Nevertheless, the direction of effect was broadly concordant across panels, and for IFN-γ and IL-2 this visual pattern aligns with the modest pooled mean differences observed in the main continuous analyses, supporting the interpretation that incremental increases in Th1 cytokine magnitude may track recent M. tuberculosis exposure and antigen load. In contrast, the largely null binary results for sustained IGRA conversion/infection may reflect information loss introduced by dichotomization and the additional between-study variability arising from non-standardized positivity thresholds, which can attenuate associations when the underlying signal is small and graded. Importantly, sustained IGRA conversion/infection is an infection/exposure endpoint; therefore, higher Th1 cytokine responses in converters should be interpreted as correlates of exposure/risk rather than correlates of protection against progression to active TB disease. These findings underscore the need for future studies to standardize continuous immune readouts (including units, antigen stimulation conditions, and analytic pipelines) to enable more quantitative cross-study synthesis and to test whether multi-marker signatures outperform single cytokines for distinguishing infection risk from disease protection.

A leave-one-out sensitivity analysis was performed to assess the robustness of the pooled association between IFN-γ responses and sustained IGRA conversion. Sequentially removing each study did not materially alter the overall random-effects summary estimate (pooled OR = 1.13, 95% CI 0.94–1.36). All leave-one-out confidence intervals overlapped with the main pooled estimate, indicating that the observed association was not driven by any single outlier study but represented a consistent, modest signal across the available evidence (Supplementary Figure S2).

We conducted a subgroup meta-analysis to explore potential effect modification by study design and population. Across pre-specified subgroups (infant cohorts, RCTs, adult cohorts, and an immunology-focused subgroup), pooled odds ratios remained close to unity, with substantial overlap in 95% confidence intervals. Point estimates were slightly above 1.0 in the RCT and adult cohort subgroups, closer to the null in the infant cohort subgroup, and higher but imprecise in the immunology subgroup (which contained only one study). Overall, these analyses revealed no clear or consistent subgroup-specific differences in the association between IFN-γ responses and sustained IGRA conversion (Figure 6).

To identify studies that disproportionately contributed to overall heterogeneity and influenced the pooled effect estimate, we constructed a Baujat plot (Supplementary Figure S3). The plot displays each study’s contribution to total heterogeneity (Q_i, x-axis) against its influence on the pooled log odds ratio (y-axis). Studies located in the upper-right quadrant contribute substantially to both heterogeneity and the instability of the summary estimate. In our analysis, Kagina et al. (2010) and Nemes et al. (2022) appeared in this region, suggesting they are key sources of heterogeneity and influence. Schmidt et al. (2025) and Van Der Meeren et al. (2018) also showed notable influence on the pooled effect. In contrast, studies like Tameris et al. (2013) contributed minimally. The plot thus indicates that a small number of studies likely drive both the observed heterogeneity and the sensitivity of the overall conclusion.

A Radial (Galbraith) plot was used to visually assess between-study heterogeneity and identify potential outliers (Figure 7). The plot displays the inverse of the standard error (precision) on the x-axis against the standardized effect (Z-score) on the y-axis. Most study points clustered near Z = 0, indicating general consistency in effect direction. However, a few studies, notably Nemes et al. (2022) and Kagina et al. (2010), showed greater vertical deviation, suggesting their effect estimates differ somewhat from the overall trend and may contribute to heterogeneity. Studies with higher precision, such as Lu et al. (2019) and Fletcher et al. (2016), are positioned toward the right of the plot and carry greater weight in the pooled estimate. The Radial plot corroborates findings from previous influence and heterogeneity analyses, helping to pinpoint potential outliers and inform sensitivity assessments.

For the outcome of sustained IGRA conversion, funnel plot inspection revealed a degree of asymmetry that may suggest small-study effects or publication bias. However, given the limited number of studies and narrow range of standard errors for this endpoint, the funnel plot should be interpreted cautiously and does not provide definitive evidence of bias (Figure 8).

To further aid in understanding the relationship between different endpoints and correlates, a schematic summarizing the endpoint hierarchy and their respective interpretations is provided (Supplementary Figure S4). This schematic illustrates how progression to active disease serves as the definitive endpoint for validating correlates of protection, while infection endpoints reflect antigen exposure and infection risk. The stratified approach ensures that these distinct types of immune responses are interpreted in their proper context.