No interval structure gave significantly more time at or above 90% VO2max in one session than even-paced long reps (3 min or more) in a network meta-analysis of 19 crossover studies with 240 athletes (198 M / 42 F, VO2max 60.5): decreasing-duration ladders (SMD 0.87, 95% CI −0.30 to 2.04) and varied-intensity reps (0.54, −0.42 to 1.49) ranked highest, 1:1 short blocks lowest (−0.68, −1.40 to 0.04), but the network was inconsistent (I² 69.2%) and GRADE certainty very low for every comparison with even long reps; an acute in-session measure, not a training outcome. Read in full
Main caveat: The authors': varied-intensity and decreased-duration rest on two studies each.
Held and colleagues asked which way of building a high-intensity interval session keeps athletes at or above 90% of VO2max for longest, combining direct and indirect comparisons in one network.
They sorted protocols into seven mutually exclusive nodes by rule: even long intervals (3 min or more), even short intervals (1 to under 3 min with rest shorter than work), short interval blocks (bouts of 1 min or less with shorter rest, grouped into longer blocks, such as 30/15 sets), 1:1 blocks (30-120 s with equal rest), varied-intensity intervals (surges within each rep), decreased-intensity intervals (a fast start then a lower intensity within each rep) and decreased-duration intervals (reps that shorten through the session).
Nineteen crossover studies with 240 athletes contributed 20 study units. Against even long intervals, no structure differed significantly.
Point estimates were highest for decreased-duration (SMD 0.87) and varied-intensity protocols (0.54) and lowest for 1:1 blocks (−0.68, p = 0.063); P-scores ranked decreased-duration first (89.6%) and 1:1 blocks last (2.8%), with even long reps in the middle (45.2%) ahead of short interval blocks (35.1%) and even short reps (32.2%).
The network was heterogeneous and inconsistent, the direct and indirect evidence disagreed for the two comparisons of 1:1 blocks and short interval blocks with long reps, and certainty was very low throughout.
The authors call the ranking exploratory and say protocol choice should rest on feasibility, the athlete and whether sessions are matched for time or work, not on the ranking.
Participants
240 athletes (198 men, 42 women) in 19 studies; 12 ± 3 per study (8 to 20); age 25.5 ± 5.3 years (study means 19.0-41.0); VO2max 60.5 ± 8.1 mL/kg/min (44.0-73.9).
Samples (Table 3): cyclists in 10 studies (from 8 'elite' cyclists, VO2max 73.9, to 8 cyclists with VO2max 47.6 and 12 amateur cyclists aged 41), runners in 7 (including 12 highly trained middle-distance runners, VO2max 59), cross-country skiers in 2 (VO2max about 70), canoe-kayak athletes in 1; one study (Beltrami 2021) contributed separate cycling and running strata and is counted in both.
Inclusion: healthy athletes aged 18 or over without conditions affecting testing; the search excluded junior, older and patient samples.
Protocol
Search: PubMed, Web of Science and SPORTDiscus to 31 January 2026, plus backward citation tracking; English peer-reviewed articles; two reviewers independently.
Inclusion: acute endurance HIIT sessions (cycling, running, swimming, rowing, canoe/kayak) compared with a different HIIT session in a prospective two- or multi-arm controlled design; outcome accumulated time at or above 90% VO2max. All included studies were crossovers; 3 had three relevant arms, the rest two.
Nodes and number of arms: even long 8; even short 6; short interval blocks 8; 1:1 blocks 10; varied-intensity 2; decreased-intensity 4; decreased-duration 2.
Analysis: SMD (difference in means over pooled SD) in a frequentist random-effects NMA (R, netmeta), even long intervals as reference because it was the most common traditional comparator; P-scores for ranking; Q split into within-design heterogeneity and between-design inconsistency; node-splitting, league table and contribution matrix added after global inconsistency was found; comparison-adjusted funnel plot and an exploratory Egger test with protocols ordered a priori.
Quality and certainty: PEDro (11 items, 6 or more out of 10 called high quality); GRADE-NMA for comparisons with even long intervals.
Matching: all comparisons were time-matched except the two decreased-duration studies, whose sessions were longer than their comparators.
Funding and interests: no external funding; no competing interests declared.
Key Results
Against even long intervals (accumulated time at or above 90% VO2max; positive = more time):
| Node | SMD (95% CI) | P-score |
|---|---|---|
| Decreased-duration | 0.87 (−0.30 to 2.04) | 89.6% |
| Varied-intensity | 0.54 (−0.42 to 1.49) | 76.7% |
| Decreased-intensity | not significant (estimate not in the text) | 68.4% |
| Even long (reference) | — | 45.2% |
| Short interval blocks | not significant (estimate not in the text) | 35.1% |
| Even short | not significant (estimate not in the text) | 32.2% |
| 1:1 blocks | −0.68 (−1.40 to 0.04), p = 0.063 | 2.8% |
Heterogeneity and inconsistency: τ² = 0.356; I² = 69.2% (49.9% to 81.0%); total Q = 55.12, df 17, p < 0.0001; within-design Q = 34.34, df 11, p = 0.0003; between-design Q = 20.78, df 6, p = 0.002.
Local inconsistency (node-splitting): direct and indirect estimates disagreed for 1:1 blocks vs even long (p = 0.0147) and short interval blocks vs even long (p = 0.0132).
Network shape: 26 pairwise effect sizes; the most frequent direct comparisons were even long vs decreased-intensity (4) and short interval blocks vs 1:1 blocks (4); varied-intensity and decreased-duration rest on two studies each.
Small-study effects: funnel plot without a clear pattern; Egger test 1.73 (SE 0.83), t = 2.09, df 24, p = 0.0469.
Quality and certainty: PEDro 7.1 ± 0.4 (6 to 8); GRADE-NMA very low for every comparison with even long intervals.
The authors note that some non-reference contrasts excluded zero but give no values in the text.
Practical Application
No interval shape is shown to keep an athlete near VO2max longer than plain long reps of 3 min or more. Across 19 crossover studies no structure differed significantly from even long intervals; hard-start, surging and shortening-ladder sessions ranked higher and 1:1 short blocks lower, with very low certainty.Time at or above 90% VO2max is measured inside one session: it is a proxy for the training stimulus, not a training outcome, and this review measured no VO2max gain or race result.
Where it agrees with the base: the direction of the ranking matches SYNTH-33's shaped long reps (fast start, over-under, from REF-700 and REF-194, both B-) and the decreasing ladder (REF-130, which is one of the two decreased-duration studies here); it matches REF-1363, where variable-intensity reps beat even-paced ones but gained only a little when intensity was matched, and where, in the base's own post hoc re-pooling of its forest plot, 1:1-type short sets lost to long reps; and it matches SYNTH-80's rule that total quality time is the first lever, since the top-ranked decreased-duration sessions were longer than their comparators.
Where it differs: SYNTH-33 says flat reps accumulate less time near VO2max than the same work done as 30/15, and REF-1363's re-pooling had 30/15-type sets matching or beating long reps, and SYNTH-80 says short-short work with short active recovery gives more time near VO2max than long reps at the same power (REF-1051, REF-046); in this network short interval blocks (the node holding 30/15 sets, and also 40/20 and 30/20 sets) ranked below even long reps (P-score 35.1% vs 45.2%), not significantly, and the direct and indirect evidence for exactly that comparison disagreed (p = 0.0132).The variable-intensity gain that is significant in REF-1363 (SMD 0.80) is not significant here (0.54, −0.42 to 1.49).
For the base's defaults: nothing here overturns SYNTH-33's 30/15 default for a trained cyclist, which rests on a crossover at equal mean power (REF-046) and two small training trials (REF-045, REF-1050); but this network adds no support to "30/15 beats long reps" for time near VO2max. Inference, not tested here: when an athlete prefers long reps, a 4-5-min rep with a hard opening is as defensible for time near VO2max as a 30/15 set.
1:1 short sets are the weakest shape for this measure: 30-120-s reps with equal rest gave the least time near VO2max (SMD −0.68, CI just crossing zero), consistent with SYNTH-80's point that short reps with long rests make an easier session; the authors' explanation (VO2 falls during the long rests) is not tested here.
Limitations
The authors': varied-intensity and decreased-duration rest on two studies each; substantial heterogeneity and significant global and local inconsistency limit the indirect comparisons and the ranking; possible effect modifiers differ between nodes (training status, sport, sex, VO2max, intensity anchoring, recovery type, time or work matching, how time at 90% was computed); PEDro totals summarise quality only; GRADE-NMA very low everywhere; small-study effects cannot be ruled out.
Acute proxy: every number is time at or above 90% VO2max inside one session; the link from that time to adaptation is supported in the introduction by one longitudinal study and was not tested here, and REF-1363 reports that this measure repeats only poorly to moderately between sessions.
Not matched for work: the decreased-duration sessions were longer than their comparators, so their top rank may partly reflect more time exercising; the authors argue the extra time near VO2max exceeded the extra session time, but that was not tested in a matched comparison.
Reporting: SMDs for decreased-intensity, even short intervals and short interval blocks against the reference are only in the forest plot, not in the text; the non-reference contrasts that excluded zero are not named; Table 3 lists one sample as "12 well-trained male cyclist and triathletes (8 males)"; the even long arm of Rønnestad 2022 is described as a mean of two conditions; reference [10], cited for even short intervals and 1:1 blocks, is not among the 19 studies in Table 3.
Overlap: many included studies are also in REF-1363 (for example Almquist 2020, Appelhans 2025, Bossi 2020, Beltrami 2021, Rønnestad 2022, Urianstad 2024, Vaccari 2020), so the two reviews are not independent evidence.
Small crossovers: 8 to 20 participants per study (30 in the study split by sport), mostly men.
Not read: the supplementary tables, including the direct and indirect estimates and the GRADE-NMA table.
How this entry was checked and changed
Why this level
risk of bias read in the full text (PEDro 7.1 on average) but certainty very low for every comparison with significant inconsistency, which caps it at A-; one step lower because the outcome is an acute in-session oxygen-uptake measure, not adaptation or performance (rule 3)