Abstract

Objective:

This systematic review and meta-analysis evaluated the effects of caffeine (CAF) on high−intensity interval exercise (HIIE) performance and examined potential moderators.

Methods:

Several databases were searched for studies of CAF on HIIE performance. Pooled effects were calculated using Hedge’s g (g) via a three-level random-effects meta-analysis. Subgroup analyses were performed based on sex, training status, CAF dose, CAF form, and interval type. A meta-regression analysis was conducted to investigate the potential moderating effect of the rest/work ratio on HIIE performance.

Results:

Twenty studies were included (n = 320; 57 females). CAF significantly improved HIIE performance (g = 0.28, 95% CI = 0.14 to 0.43), concurrently elevating blood lactate (g = 0.50, 95% CI = 0.22 to 0.79) and glucose (g = 0.56, 95% CI = 0.17 to 0.96). Subgroup analyses demonstrated significant improvements across all sexes (g = 0.30–0.34), trained athletes (g = 0.44), CAF dose (low and moderate) (g = 0.28–0.33), CAF form (capsule and beverage) (g = 0.32–0.33), and HIIE protocol (repeated short sprints and short intervals) (g = 0.16–0.36), with no subgroup differences (all p > 0.05). The rest/work ratio was a significant moderator of HIIE performance (β2 = 0.003, p = 0.017).

Conclusion:

CAF ingestion significantly enhances HIIE performance, with ergogenic benefits observed across both sexes and trained athletes. Effective ergogenic benefits can be achieved with a low dose (~3 mg/kg), administered as a capsule or beverage. Notably, meta-regression indicates that the rest/work ratio is a critical moderator of HIIE performance, with evidence of a nonlinear association.

1 Introduction

High-intensity interval exercise (HIIE) refers to repeated bouts of near-maximal to “all-out” effort interspersed with recovery periods of low-intensity work or rest (Coates et al., 2023). The capacity to perform HIIE is a critical determinant of performance in intermittent sports (e.g., team, racket, and combat sports) (Buchan et al., 2013; Salinero et al., 2019; Diaz-Lara et al., 2023), as high-intensity actions (e.g., jumping, tackling, and repeated approach spike jumps) directly influence scoring opportunities and can ultimately influence match outcomes (Spencer et al., 2005). However, HIIE performance progressively declines over time due to central fatigue (e.g., adenosine accumulation) and peripheral mechanisms (e.g., phosphocreatine depletion) (Nehlig et al., 1992; Gaitanos et al., 1993). To optimize athletic performance, effective nutritional strategies are required to sustain HIIE performance, ultimately translating to enhanced competitive match-play.

Caffeine (CAF) is one of the most commonly used ergogenic aids among athletes across various sports (Del Coso et al., 2011). Its efficacy in enhancing endurance (Southward et al., 2018; Chen et al., 2024) and resistance (Grgic and Del Coso, 2021; Xiao et al., 2025) exercise performance is well-established. In recent years, increasing attention has focused on the effects of CAF on HIIE performance. However, given that HIIE encompasses a highly diverse and complex range of protocols (Tschakert and Hofmann, 2013), findings in this area remain inconclusive. For instance, a previous systematic review by Lopes-Silva et al. (2019) concluded that CAF ingestion fails to enhance repeated sprint ability (RSA, a form of HIIE). Nevertheless, this evidence may not be generalizable to other HIIE modalities and is further constrained by several methodological limitations. First, their analysis was limited by incomplete study inclusion. The omission of three relevant trials could potentially alter their pooled effect estimate (Paton et al., 2001; Del Coso et al., 2012; Lara et al., 2014). For example, Lara et al. (2014) reported that 3 mg/kg of CAF significantly enhanced running speed during the sprint test (24.2 ± 1.6 vs 24.5 ± 1.7 km/h; P < 0.05). Second, the previous review did not investigate whether participant characteristics, supplementation strategies, and HIIE protocols moderate the effects of CAF on HIIE. This gap is particularly noteworthy given that primary studies suggest greater benefits are often observed in males, trained athletes, moderate doses (4–6 mg/kg), and critically, under specific interval designs such as varying rest/work ratios (Carrillo and Benitez, 1996; Davis and Green, 2009; Fiorenza et al., 2019; Khodadadi et al., 2025; Xue et al., 2025). Consequently, a comprehensive meta-analysis is warranted to systematically evaluate the effects of CAF on HIIE performance and to examine these potential moderating factors.

Therefore, this study used a three-level meta-analysis to evaluate the effects of CAF on HIIE while accounting for dependence among multiple performance outcomes reported within the same study. Meta-regression was further performed to examine the moderating role of the rest/work ratio, providing a more nuanced understanding of CAF’s ergogenic effects and their practical implications for intermittent sports nutrition.

2 Methods

The systematic review adhered to the 2020 Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) statement (Page et al., 2021). Additionally, this review was registered in the International Prospective Register of Systematic Reviews (PROSPERO) database (CRD420261337845).

2.1 Literature search

The literature review was conducted across PubMed, Web of Science, Cochrane Database, and Scopus from inception to October 18, 2025. We used combinations of subject headings and keywords with Boolean operators (“AND,” “OR,” and “NOT”) and truncation, including terms such as “CAF”, “coffee”, “energy drink”, “high-intensity interval exercise”, “interval exercise” (detailed search strategy is presented in Supplementary Appendix S1). Additionally, searches of PROSPERO of Systematic Reviews were conducted to determine whether protocols for related systematic reviews had already been published. Finally, to avoid missing any relevant literature, we also conducted hand searching in reference citations of identified reviews.

2.2 Selection process

Deduplication of retrieved records was performed manually by an independent reviewer (P.Y.) using EndNote X9 (Clarivate Analytics, Philadelphia, PA, USA). Then, two independent researchers (P.Y. and H.L.) screened the titles and abstracts. If consensus could not be reached, a third independent researcher (L.B.) was consulted. Finally, two researchers (P.Y. and H.L.) independently reviewed the full texts of selected articles for final inclusion. Any discrepancies were resolved based on the predefined inclusion and exclusion criteria.

2.3 Selection criteria

A set of a priori inclusion and exclusion criteria were used to evaluate study eligibility according to the PICOS framework: (1) P: healthy adult human participants (≥ 18 years of age); (2) I: studies examining the effects of CAF ingestion on HIIE performance; (3) C: included a placebo group as the comparator; (4) O: eligible studies reporting at least one outcome of overall HIIE performance [i.e., fastest time, average time, total time, mean power output (MPO), peak power output (PPO), and total work] or physiological responses [i.e., heart rate (HR), blood lactate concentration (BLC), rating of perceived exertion (RPE), glucose]; (5) S: utilized a double-blind randomized crossover design.

Exclusion criteria were: (1) non-English articles; (2) studies examining long-term CAF interventions during HIIE; (3) systematic reviews or meta-analyses, case studies, non-peer-reviewed manuscripts, and conference proceedings; (4) animal studies; (5) studies on therapeutic or disease-related outcomes.

2.4 Assessment of methodological quality

Risk of bias was assessed using the Cochrane Risk of Bias Assessment Tool via Review Manager 5.4 software (The Cochrane Collaboration, 2020; Cohen, 1988). The tool evaluates (1) random sequence generation; (2) allocation concealment; (3) blinding of participants and personnel; (4) blinding of outcome assessment; (5) incomplete outcome data; (6) selective reporting; (7) other bias. Ratings for each category are denoted as either “low risk” (“+”), “high risk” (“-”) or “unclear risk” (“?”). Quality assessment was performed independently by two investigators (P.Y. and H.L.), with discrepancies resolved by discussion or through consultation with a third reviewer (L.B.).

Additionally, the physiotherapy evidence database (PEDro) (de Morton, 2009) scale was used to assess the risk of bias and methodological quality of included studies. Studies were scored on a scale of 0–10, with scores ≥ 6 indicating high quality, 4–5 moderate quality, and ≤ 3 low quality.

2.5 Data extraction and study coding

Data extraction was independently conducted by two researchers (P.Y. and W.X.) using Excel (Version 16.93, Microsoft, Redmond, WA, USA). Extracted data included author details, sample characteristics (e.g., number, sex, training status), exercise protocol, overall HIIE performance outcomes (e.g., MPO, PPO, total work, fastest time, average time, and total time), main physiological responses (e.g., HR, BLC, glucose) and RPE. Discrepancies in data extraction were solved by consensus. Relevant data were extracted using WebPlotDigitizer 4.7 (Drevon et al., 2017) if data were missing or presented only in graphical form. When a study reported multiple sets of target data, they were combined into a single mean value. We were converted standard errors (SE) to standard deviations (SD) using the formula recommended in the Cochrane guideline (Cumpston et al., 2019) if the study provide SE. Some studies were included multiple times in the analysis due to independent comparisons across different doses to ensure that data related to HIIE and CAF were fully analyzed.

2.6 Statistical analyses2.6.1 Calculation of effect size and variance

The effect of CAF ingestion on HIIE performance was analyzed by comparing CAF with the placebo condition. The mean difference (MD) and the SD of the change in means () were calculated according to the recommendations of the Cochrane handbook for the evaluation of intervention systems (version 6.5, 2024), using the following formula (Higgins et al., 2019). The first step involved calculating the difference in means:

where is the reported mean values of CAF group and is the reported mean values of placebo group.

Then the for crossover studies was calculated as follows (Hedges and Shymansky, 1989):

where is the SD from CAF group and is the SD from placebo group.

Considering the relatively small sample sizes of most included studies, in each analysis we used Hedge’s g (g) as the point estimate of the mean effect size, which for crossover studies is as follows (Hedges, 1983):

where is the total sample size. g was classified as trivial (SMD < 0.20), small (0.20 ≤ SMD < 0.50), moderate (0.50 ≤ SMD < 0.80), or large (SMD ≥ 0.80) (Cumpston et al., 2019).

For the crossover experimental design, the SE of g was calculated using the following formula (Hedges, 1983):

where r is the correlation coefficient between the CAF and placebo conditions. Few included studies reported the within-subject correlation coefficient (r). We also reviewed prior meta−analyses in this field, neither of which reported r values. Therefore, in accordance with the Cochrane Handbook (Cumpston et al., 2019), we assumed a correlation of r = 0.50 for the primary analysis. To assess the robustness of our findings, we conducted sensitivity analyses using alternative r values, specifically r = 0.20 (lower bound) and r = 0.80 (upper bound), for the overall HIIE performance outcomes.

2.6.2 Meta-analysis and heterogeneity

Traditional two-level meta-analysis was conducted to aggregate the physiological response with the meta and metafor packages in R (V.4.2.0, R Core Team, Vienna, Austria) (Viechtbauer, 2010). For two-level meta-analysis, we used random-effects model and standardized mean differences (SMDs), which estimated as Hedges’ g using the restricted maximum likelihood method (REML) to ensure stability. Hedge’s g was classified as trivial (SMD < 0.20), small (0.20 ≤ SMD < 0.50), moderate (0.50 ≤ SMD < 0.80), or large (SMD ≥ 0.80) (Cumpston et al., 2019).

We also conducted a three-level meta-analysis following recommendations of Assink and Wibbelink (Assink and Wibbelink, 2016) to address the double counting or missed correlations in HIIE performance with nested or multiple effect sizes (e.g., the MPO and PPO) (Kadlec et al., 2023). By preserving valuable information from multiple effects within each study, the three-level meta-analysis enhances statistical power and provides a more accurate representation of effect sizes (Assink and Wibbelink, 2016). This approach decomposes variance into sampling variance (Level 1), within-study variance (Level 2), and between-study variance (Level 3), accounting for correlated and hierarchical effects (Cheung, 2019). For the three-level model, parameters were estimated using the REML, and results were cross-verified using the maximum likelihood (ML) method to ensure stability.

Tests of individual coefficients and their corresponding 95% confidence intervals (95% CIs), were based on a t-distribution (Jukic et al., 2023). Additionally, prediction interval (PI) were calculated based on the t-distribution, which provides useful additional information compared to the 95% CI, especially considering the use of a random-effects model (Borg et al., 2024). Heterogeneity was assessed using I² and PI, with I² categorized as low (0–25%), moderate (25–50%), substantial (51–75%), or considerable (76–100%) (Higgins and Thompson, 2002). The statistical significance threshold was set at p < 0.05.

To explore sources of heterogeneity among studies and assess potentially moderating factors, this meta-analysis employed subgroup analysis and meta-regression analysis, conducting statistical analyses on continuous variables (Hopkins and Batterham, 2018). Moderator analyses were conducted when there were at least 3 studies (Cumpston et al., 2019) and at least 10 studies available for each meta-regression (Ruppar, 2020) to explore potential sources of heterogeneity. The following variables were included in the subgroup analysis: (a) sex group; (b) training status; (c) CAF dose; (d) CAF form; (e) interval type.

Based on previous participant categorization frameworks, training status was classified as recreationally active, trained, and well-trained (McKay et al., 2022). HIIE interval time were categorized as follows based on Buchheit and Laursen (Buchheit and Laursen, 2013) and Girard (Girard et al., 2011): short intervals (duration < 60s per repetition), intermittent-sprint exercise (ISE, repeated sprints with duration < 10s for each sprint, recovery > 60s), RSA (repeated sprints with duration < 10s for each sprint, recovery < 60s).

We conducted a mix-effects meta-regression analysis with REML estimation, recognized for its robustness (Zuur et al., 2009). To assess the shape of relationships, we fitted linear and quadratic functions and compared the models, selecting the model with the lowest bias-corrected Akaike information criterion (Harrell, 2015). All analyses were conducted using the ‘metafor’ package, and results were visualized using ‘ggplot2’ (Gustavsson et al., 2022). Statistical power was calculated for each subgroup and the overall pooled effect to assess the risk of false negatives. These calculations were performed using the ‘metameta’ package (Quintana, 2023).

2.6.3 Publication bias and sensitivity analysis

Funnel plots (Peters et al., 2008), Egger’s asymmetry test (Egger et al., 1997), and trim-and-fill tests were used to assess publication bias (with analyses conducted only when k > 10 (Sterne et al., 2011)). A p-value > 0.05 was considered indicative of no publication bias.

We conducted sensitivity analyses using a leave-one-out method to assess the robustness of the primary pooled effects in two-level and three-level meta-analyses. For three-level meta-analysis, Cook’s distance (Viechtbauer and Cheung, 2010) and studentized residuals (Atkinson et al., 1983) were used at both the within-study level (level 2) and the between-study level (level 3). Observations were flagged as potentially influential if their hat values or Cook’s distances exceeded three times their respective means, or if the absolute value of studentized residuals exceeded three. The three-level model was then re-estimated with outliers excluded to examine the robustness of the findings.

2.7 Certainty of the evidence

The risk of bias was considered in the interpretation of the results by applying the Grading of Recommendations Assessment, Development, and Evaluation (GRADE) methodology. Evidence was rated as “high” (further research is very unlikely to change our confidence in the estimate of effect), “moderate” (further research is likely to have an important impact on our confidence in the estimate of effect and may change the estimate), “low” (further research is very likely to have an important impact on our confidence in the estimate of effect and is likely to change the estimate) or “very low” (any estimate of effect is very uncertain) (Schünemann et al., 2019). GRADE assessment was evaluated by one reviewer (P.Y.) and independently reviewed by a second reviewer (X.K.).

3 Results3.1 Studies retrieved

The initial search yielded 1926 publications, comprising 1923 records identified through the primary database search and an additional 3 from google research. After screening and assessment of eligibility, 20 studies were included in the meta-analysis (Figure 1).

Flow diagram of study selection.

3.2 Characteristics of included studies

Table 1 shows the main characteristics of all included studies. The total participant pool comprised 320 individuals, including 263 males and 57 females. Individual study sample sizes ranged from 8 to 52 participants. CAF withdrawal periods ranged from 8 to 72 h. The HIIE recovery time ranged from 10s to 300s and sprint durations from 4s to 60s. CAF forms included capsules, beverages, and gum. Participant training status included recreational active, trained, and well-trained athletes.

StudySample + age (years) + body mass + training statusHabitual caffeine intake (mg/day) +
caffeine withdrawal (h)Caffeine dose (mg/kg)Timing (min)Caffeine formHIIE protocolRest/work ratioOutcomesPaton et al., 200116 M; 22 ± 3; 79 ± 9; recreational activeN.A.; 48660capsules10 × 10s running (maximal-effort sprints) with 10s active recovery1⑥Crowe et al., 200612 M, 5 F; 21.1 ± 3.0; 73.3 ± 10.5; recreational active80–200; 24672beverage2 × 60s cycling (maximal-effort sprints) with 180s recovery3②③❶❷❸❹Schneiker et al., 200610 M; 20 ± 3; 77.7 ± 13.9; recreational activeN.A.; 48660capsules2 × (18 × 4s cycling [maximal-effort sprints] with 100s [35% VO 2peak]+ 20s recovery)30②③❷Carr et al., 200810 M; 25 ± 5; 85.25 ± 13.98; trainedN.A.; 48660capsules3 × (6 × 20 m running [maximal-effort sprints] with 25s recovery) + 2 × (6 × 20 m running [maximal-effort sprints] with 60s recovery)5.4④⑤❷Glaister et al., 200821 M; 21 ± 3; 77.7 ± 13.5; recreational active88.4 ± 87.3; 48560capsules12 × 30 m running (maximal-effort sprints) with 35s active recovery7.8⑤⑥❶❷❸Paton et al., 20109 M; 24 ± 7; 62.5 ± 5.4; well-trained< 300; 72240 mg5gum2 × (10 × 30s cycling [maximal-effort sprints] with 30s recovery)1①Lee et al., 2012a14 M; 18.7 ± 0.8; 67.7 ± 6.2; recreational active> 200; 72660capsules2 × (12 × 4s cycling [maximal-effort sprints] with 20s recovery)5①②③❷❸Lee et al., 2012b14 M; 18.7 ± 0.8; 67.7 ± 6.2; recreational active> 200; 72660capsules2 × (12 × 4s cycling [maximal-effort sprints] with 90s recovery)22.5①②③❷❸Del Coso et al., 201219 M; 21 ± 2; 67 ± 2; trained< 60; 48360beverage7 × 30 m running (maximal-effort sprints) with 30s active recovery6.7⑦Lee et al., 2014a12 M; 20.4 ± 1.1; 75 ± 9; recreational activeN.A.; 72670capsules10 × (5 × 4s cycling [maximal-effort sprints] with 20s recovery)5①②③❶❷❸❹Lee et al., 2014b8 F; 21.3 ± 1.2; 58.6 ± 7.3; well-trained50–100; 48660capsules10 × (5 × 4s cycling [maximal-effort sprints] with 20s recovery)5①②③❸StudySample + Age (years) + Body mass + Training statusHabitual caffeine intake (mg/day) +
Caffeine withdrawal (h)Caffeine dose (mg/kg)Timing (min)Caffeine formHIIE protocolRest/work ratioOutcomesLara et al., 201418 F; 21 ± 2; 57.8 ± 7.7; trainedN.A.; 48360beverage7 × 30 m running (maximal-effort sprints) with 30s active recovery6.7⑦Evans et al., 201818 M; 21.2 ± 1.1; 80.4 ± 6.6; recreational activeN.A.; 24200 mg5gum10 × 40 m running [maximal-effort sprints] with 30s active recovery3.4⑤⑥❷Fowles et al., 202113 M; 32 ± 11; 65.7 ± 5.9; trained190 ± 134; 122.0860beverage4 × 60s cycling (maximal-effort sprints) with 5 min active recovery5①②③❶❸Kopec et al., 201611 M; 20 ± 2; 74.5 ± 8.2; trained< 160; 24660capsules3 × (6 × 20 m running maximal-effort sprints with 25s recovery) + 2 × (30 × 1min [1 × agility + 1 × sprints + 2 × jogging + 3 × walking])7.1④⑤❶❷Dos Santos Quaresma et al., 202110 M; 26.9 ± 4.0; non-athlete200–400; 8660beverage60s cycling (90% W max) with 120s (50% W max) recovery2❸❹Bezerra et al., 202210 M; 3.6 ± 3.3; 71.2 ± 8.7 recreational activeN.A.; 24660capsules6 × IAR (maximal-effort sprints) with 60s active recovery3.4④⑤⑥❶❸Salgueiro et al., 202210 M; 18.2 ± 1.7; 72.7 ± 10.8; trainedN.A.; 48660capsules10 × 400 m swimming (maximal-effort sprints) with 60s active recovery0.2⑦❷❸Bernardo et al., 202412 M; 26 ± 4; 80.7 ± 7.6 recreational active198 ± 114; 24660capsules12 × 6s cycling (maximal-effort) with 60s active recovery10②③❷❸❹Pérez-López et al., 2025a26 M; 24.6 ± 4.5; 76.3 ± 10.7; trained68 ± 61; 72360beverage4 × 30s cycling (maximal-effort sprints) with 90s recovery3①②❷Pérez-López et al., 2025b26 F; 24.6 ± 4.5; 76.3 ± 10.7; trained68 ± 61; 72360beverage4 × 30s cycling (maximal-effort sprints) with 90s recovery3①②❷Matsumura et al., 202516 M; 20 ± 1; 66.0 ± 4.3; well-trained226 ± 168; 72660capsules3 × 30s cycling (maximal-effort sprints) with 120s recovery4①②

a, b represents difference trials. M, male. F, female. W max, maximal power output. CAF, caffeine; PLA, placebo; HIIE, high-intense interval exercise; IAR, Illinois agility run; N.A., represents that the value is not available. The symbols ①, ②, ③, ④, ⑤, ⑥, ⑦, ❶, ❷, ❸, and ❹ represent mean power output, peak power output, total work, total time, fastest time, average time, speed, heart rate, blood lactate concentration, ratings of perceived exertion, and glucose.

3.3 Primary analysis

For overall HIIE performance, the three-level meta-analysis showed that CAF ingestion significantly improved HIIE performance (g = 0.28, 95% CI: 0.14 to 0.43 p < 0.001) (Figure 2), In addition, variance decomposition showed that there was no within-study variance (Level 2, 0%). The overall heterogeneity was driven entirely by variance between studies (Level 3, 68%), of which 32% was attributable to sampling error (Level 1). In accordance with the recommendations of Hunter and Schmidt (Raudenbush, 1991), significant heterogeneity is considered present when the total variance explained by sampling error is less than 75%. Consequently, we proceeded with a moderator analysis to investigate the sources of this heterogeneity.

Primary pooled effect sizes for acute caffeine supplementation on overall HIIE performance. K, the total number of effects included in the pooled effect size; Hedge’s g, the effect size indicators used in the pooled; 95% CI, 95% confidence interval; PI, prediction interval; p value, statistically significant p values for pooled results; I2, quantitative indicators of heterogeneity; Power, statistical power for pooled effect size; Blue circles, Grade, grading of recommendations assessment, development and evaluation, a system for evaluating the quality of evidence and strength of recommendations.

Regarding physiological responses, the two-level meta-analysis showed that CAF ingestion significantly increased glucose (g = 0.56, 95% CI: 0.17 to 0.96, p = 0.005) following CAF ingestion during HIIE. Additionally, a significant elevation in BLC was observed (g = 0.50, 95% CI: 0.22 to 0.79, p < 0.001). However, there was no significant difference in HR (g = 0.23, 95% CI: -0.07 to 0.54, p = 0.126) between CAF and placebo conditions (Figure 3).

Effect of caffeine ingestion in HIIE as compared to placebo on HR, BLC, glucose and RPE. 95% CI, 95% confidence interval; Hedge’s g, the effect size indicators used in the pooled; I2, quantitative indicators of heterogeneity; K, the total number of included studies; CAF, caffeine-supplement group; PLA, placebo-supplement group; p value, statistically significant p values for pooled results; HR, heart rate; BLC, blood lactate concentration; RPE, ratings of perceived exertion.

Regarding perceptual responses, the meta-analysis found no significant difference in RPE (g = -0.11, 95% CI: -0.37 to 0.15, p = 0.391) between CAF and placebo during HIIE (Figure 3).

3.4 Moderator analysis

We conducted moderator analysis to explore th