Low-carbohydrate and ketogenic diets in trained athletes: across 13 small studies (5 per outcome) single maximal efforts held up (d +0.29, 95% CI −0.08 to 0.66, not significant), repeated sprint and interval performance tended to fall (d −0.33, −0.80 to 0.14, not significant; two elite race-walker trials 8% slower in interval sessions) and blood lactate was lower (d −0.89, −1.20 to −0.58, 3 studies); the authors call the estimates exploratory, and quality was rated with a tool that does not assess randomisation or carry-over
Main caveat: The authors': small studies with mostly fewer than 12-15 participants; crossover designs with possible carry-over.
Gawelczyk and colleagues (Katowice) pooled controlled studies of low-carbohydrate and ketogenic diets in trained athletes aged 18-45 and asked what they do to short, hard efforts; it is the anaerobic half of a project whose aerobic half is in the base as REF-1009.
They searched five databases with Boolean strings and added a semantic search through Elicit, which returned 500 records; 190 were screened, 49 read in full and 13 included. Two studies contributed to two outcomes, giving 15 entries split evenly over three meta-analyses of 5 each.
For single maximal power (Wingate, countermovement jump, 1RM), two studies found improvements, three no difference, and the pooled effect was small and not significant (d +0.29, 95% CI −0.08 to 0.66; I² 55.5%).
For repeated sprints and interval sessions, three studies found no difference and two, both 3-week ketogenic trials in elite race walkers, found interval speed 8% slower; the pooled effect leaned toward harm but was not significant (d −0.33, −0.80 to 0.14; I² 65.1%). Blood lactate was lower in all three studies with usable numbers (d −0.89, −1.20 to −0.58; I² 0%).
The authors read the pattern as phosphagen-dominated single efforts being spared while glycolysis-dependent repeated efforts suffer, treat the lower lactate as reduced glycolytic flux rather than a benefit, call every pooled estimate exploratory, and recommend carbohydrate periodisation over chronic restriction.
They also note that impairments came mostly from elite or highly trained athletes and null or positive results mostly from moderately trained ones.
Participants
13 studies; 273 participants where sample size was reported (range 5 to 65); one study (McKay 2023) did not report its enrolment, so the true total is larger (the authors), though Tables 5 and 6 list it with n = 21.
Populations: endurance-trained cyclists, runners and triathletes of both sexes (6 studies), elite race walkers (3), weightlifters and powerlifters (1), semi-professional male soccer players (1), male taekwondo athletes (1), moderately trained men (1).
Inclusion asked for competitive athletes with at least 6 months of structured training, aged 18-45; untrained and recreational samples were to be excluded.
Overall sex split: not reported.
Protocol
Search: Boolean strings in five databases to October 2025, English only, plus an Elicit semantic search of Semantic Scholar and OpenAlex (capped at 500 records) that also screened bioRxiv and medRxiv; reference lists back-screened. Screening by two reviewers (κ 0.78).
Diets: low-carbohydrate ≤130 g/day or ≤25% of energy; ketogenic <50 g/day or <10% with blood β-hydroxybutyrate ≥0.5 mmol/L; at least 3 days (acute) or 4 weeks (chronic). By entry: 8 ketogenic, 2 moderate low-carbohydrate, 4 high-fat without a stated carbohydrate ceiling (classed as low-carbohydrate), 1 calorie-restricted low- vs moderate-carbohydrate.
Comparators: high-carbohydrate (>60% of energy or >300 g/day), mixed (45-55%) or habitual diets.
Designs: 9 crossover and 4 parallel studies by Table 1 (Tables 4-6 label 8 crossover and 5 parallel).
Outcomes and analysis: three domains of 5 entries each (power; repeated sprint and interval; blood lactate, 3 with numbers); Cohen's d, random effects, Q, I², τ², 95% prediction intervals; leave-one-out, quality-stratified, design and duration sensitivity analyses; Egger, Begg and trim-and-fill.
Extraction: first drafts by Elicit's language models, every field re-checked by two reviewers; 3 of 325 fields disagreed; reviewer agreement κ 0.94.
Quality: Newcastle-Ottawa Scale, cohort-style adaptation; the authors acknowledge RoB 2 would have been the fitting tool.
Funding and interests: no external funding; no conflicts declared.
Key Results
| Domain (low-carbohydrate or ketogenic vs control) | k | Pooled d (95% CI) | Heterogeneity | 95% prediction interval |
|---|---|---|---|---|
| Single maximal power (Wingate, CMJ, 1RM) | 5 | +0.29 (−0.08 to 0.66), not significant | I² 55.5%, τ 0.30, Q p = 0.061 | −0.84 to 1.41 |
| Repeated sprint and interval performance | 5 | −0.33 (−0.80 to 0.14), not significant | I² 65.1%, τ 0.39, Q p = 0.022 | −1.74 to 1.07 |
| Blood lactate during or after hard efforts | 3 | −0.89 (−1.20 to −0.58) | I² 0%, Q p = 0.952 | −2.87 to 1.04 |
Studies behind each domain (Tables 4-6, as the review reports them):
| Study | Design, n | Sport | Diet, duration | Result |
|---|---|---|---|---|
| Paoli 2021 | parallel, 16 | soccer | ketogenic, 30 days | CMJ improved vs Western diet (p = 0.0021) |
| McSwiney 2018 | crossover, 47 | mixed | ketogenic, 12 days (Table 4) | sprint peak power increased vs high-carbohydrate (p = 0.025) |
| Sawyer 2013 | crossover, 31 | mixed, both sexes | low-carbohydrate, 7 days | Wingate, 1RM and jump maintained |
| Greene 2018 | crossover, 14 | weightlifting, powerlifting | ketogenic, 3 months | 1RM no difference |
| Lambert 1994 | crossover, 5 | cycling | high-fat, 2 weeks | Wingate no difference |
| Prins 2023 | crossover, 10 | running | low-carbohydrate, 6 weeks | 6 × 800 m equivalent |
| Hsu 2023 | crossover, 12 | taekwondo | low/moderate carbohydrate, 7 days | repeated sprints no difference |
| Ramonas 2023 | crossover, 9 | running | low-carbohydrate, 2 days | 50-3000 m time trials no difference |
| McKay 2023 | parallel, 21 | race walking | ketogenic, 3 weeks | interval speed 8% slower (p = 0.001); lactate lower (p < 0.001) |
| McKay 2022 | parallel, 28 | race walking | ketogenic, 3 weeks | interval speed 8% slower (p < 0.001) |
| Moitzi 2024 | parallel, 65 | moderately trained men | ketogenic, 10 weeks | lactate lower (p < 0.001) |
| Burke 2020 | parallel, 28 | race walking | low-carbohydrate, 25 days | lactate lower (p < 0.001) |
Leckey 2018 and Prins 2023 are listed in the lactate domain with no lactate data.
Sensitivity: leave-one-out estimates overlapped the primary ones; the quality, design and duration strata are described as directionally consistent, with no numbers given; restricting to the 4 studies with ≥15 per condition gave overlapping intervals (no numbers given).
Publication bias: Egger p = 0.42 (power) and p = 0.18 (repeated sprint); trim-and-fill added no studies for power.
Quality as rated: mean Newcastle-Ottawa 7.2 ± 0.8 (range 6-9); 12 of 13 at 7 or more; 6 of 13 with complete follow-up and low dropout; outcome-assessor blinding in about half.
Practical Application
A low-carbohydrate or ketogenic diet left single maximal efforts unchanged on average but tended to cost repeated hard efforts, and the clearest losses were in elite race walkers. Across 13 small studies the pooled effects were d +0.29 for single maximal power and d −0.33 for repeated sprint and interval performance, neither significant; the two significant losses, 8% slower interval sessions after 3 weeks of a ketogenic diet, both came from elite race walkers.Lactate was lower, which the authors read as less glycolysis, not better fitness.
This fits [SYNTH-69]: fuel the key sessions, and use low availability only on easy ones. The authors themselves recommend carbohydrate periodisation over chronic restriction for athletes whose sport has repeated surges or sprint finishes; nothing here argues for doing intervals on a ketogenic diet.
Do not read a lower lactate on a low-carbohydrate diet as fitness. Inference, not tested here: a lactate test or a lactate-based zone set on a ketogenic diet may sit lower than it would on a normal diet, so test on the diet the athlete normally eats.
Against the opposing view: a commentary in the base argues that habitual diet makes no difference to performance after adaptation (REF-1701, C); a 2023 Prins 6 × 800-m trial, apparently the one it cites, is among the null studies here (equivalent performance), while the race-walker trials it calls too short carry the impairments. The two non-significant pooled estimates leave both readings open.
Not settled by this design: whether a longer adaptation removes the repeated-sprint cost (the authors suggest it may; no pooled test of duration is given), whether apparent gains in relative power reflect weight loss rather than more absolute power, and anything about women, races or cycling-specific sprint work beyond one 5-person Wingate study.
Limitations
The authors': small studies with mostly fewer than 12-15 participants; crossover designs with possible carry-over; diets from 2 days to 12 weeks, strict and moderate restriction pooled; laboratory tests with limited link to competition; English only; endurance athletes over-represented (9 of 13 studies counting race walkers); a semantic search that cannot be re-run exactly; the Newcastle-Ottawa Scale used instead of RoB 2; only k = 3-5 per domain, so the estimates are exploratory and the prediction intervals cross zero.
Risk of bias not assessed in the trial sense: the cohort-style Newcastle-Ottawa items (representativeness, exposure ascertainment, comparability, follow-up) do not rate randomisation, allocation concealment or carry-over; the "high quality" label for 12 of 13 studies rests on that tool.The methods list randomisation adequacy and concealment, blinding, selective reporting, power and conflicts of interest as additional quality considerations, but the results report only outcome-assessor blinding (about half of the studies). No GRADE or other certainty rating.
One domain pools different things: 1RM, a 30-s Wingate and a countermovement jump, which the authors themselves say are distinct constructs; the repeated-sprint domain mixes 6 × 800 m, race-walk interval sessions, 50-3000 m time trials and taekwondo sprints.
Counts that disagree inside the paper: designs 9 crossover / 4 parallel in Table 1 but 8 / 5 by the study tables; McKay 2023 "did not report total enrolment" yet appears with n = 21; McSwiney 2018 is a 12-day crossover in Table 4 and the text but among the 6-week to 3-month interventions in the duration count, and the team-sport study is attributed to it in the participant section while Table 4 gives soccer to Paoli 2021; Prins 2023 lasts 6 weeks in Tables 5-6 and 31 days in the note to Table 2; durations run 2 days to 12 weeks in the abstract, 7 days to 12 weeks in the methods, and the inclusion minimum was 3 days, yet a 2-day study is included; the population criterion excluded non-competitive samples, yet the largest study is of moderately trained men without competitive status; the conclusions speak of 15 studies; the lactate domain is called "statistically robust, and homogeneous" while its prediction interval (−2.87 to 1.04) also crosses zero.
Not given: per-study effect sizes (forest plots not available), the numbers behind the subgroup and sensitivity statements, and the overall sex split.
Extraction by language model: first drafts by Elicit, re-checked by two reviewers (3 of 325 fields corrected).
How this entry was checked and changed
Why this level
a systematic review and meta-analysis read in full, held below A- because its quality tool (Newcastle-Ottawa, built for observational studies) does not rate randomisation, allocation concealment, blinding or carry-over, so the trials' risk of bias in the rubric's sense was not assessed and the rule-1 cap applies as if it were unread (with mostly small studies and I² of 55.5-65.1% in the performance domains it could reach A- at most even with a trial tool); the one significant pooled result, lower blood lactate, is a marker (rule 3) and on its own would sit at B