REF-1502 · Li et al., 2026 · International Journal of Sports Science & Coaching, article · Paper

Was my slow last run a fatigue problem?

Splits cannot say. In 61,541 HYROX records early splits left a wide late-run spread (P90-P10 67.8 s; 33.2 s in a second sample); the authors bar reading it as fatigue or a training threshold.

Main caveat: Abstract only; bands not read. The publisher's host refused with a bot check, not bypassed.
moderate evidenceMeasured on recreational to elite athletes

Li and colleagues used public HYROX split records to ask how much of the late-race running can be expected from the early race. The discovery sample was 61,541 athlete-event records from 89 events (2021-2023); a separately assembled secondary-source sample of 19,769 records from 237 events (2021-2025) checked direction.

The main model used runs 1-3 and stations 1-3 to explain runs 6-8. Run 8 showed the clearest terminal slowing. The record-level residual P90-P10 was 67.8 s (33.2 s in the second sample) and the final-run P90-P10 111.8 s (63.2 s). Stricter screens of 120 s and 150 s minimum run splits changed little.

Adding runs 4-5 and stations 4-5 narrowed the record-level spread to 47.6 s (23.0 s). Time-ordered models of run 8 using all prior history kept it at 67.2 s (33.0 s). Descriptive final-run reference bands were given by division, sex and early-ability tier; we could not read them.

The authors conclude that richer history explains a meaningful part, not all, and that the residual is a model-dependent post-race indicator, not longitudinal stability, a fatigue diagnosis or a direct training threshold. Our reading: a clear warning that two athletes with the same first half can finish the last run far apart, and that this spread cannot be read as a diagnosis.

Protocol

Design: retrospective observational analysis of public HYROX split records.
Discovery sample: n = 61541 athlete-event records from 89 events, 2021-2023.
Secondary-source validation sample: 19,769 records from 237 events, 2021-2025; separately assembled; directional validation.
Repeated athletes: the unit is the athlete-event record; the abstract does not give unique athletes or say how repeats were handled.
Division and sex: bands were stratified by division, sex and early-ability tier, but the mix in each sample is not in the abstract.
Primary model (E13_L68): run 1-3 plus station ("Work") 1-3 as the early window; run 6-8 as the late outcomes.
Sensitivity: stricter 120-s and 150-s minimum run-split screens; adding run 4-5 and station 4-5; time-ordered run 8 models with all prior run 6-7 and station 6-7 history.
Funding and interests: not available from the abstract. Overlapping authors with REF-1498.

Key Results

MeasureDiscovery (61,541 records, 89 events)Secondary source (19,769 records, 237 events)
Record-level residual P90-P10, primary model67.8 s33.2 s
Final-run P90-P10111.8 s63.2 s
Record-level residual P90-P10, with run 4-5 and station 4-5 added47.6 s23.0 s
Time-ordered run 8 models, residual P90-P1067.2 s33.0 s
Screens: the 120-s and 150-s minimum run screens gave essentially unchanged estimates.
Direction: the second sample showed the same narrowing with richer prior information.
Reference bands: given in the paper by division, sex and early-ability tier. Not read, and no value is quoted here. No run or station split appears in the abstract.
Does not add up, or needs the paper: 47.6 s is lower than 67.2 s even though the second model holds more history. They are likely different outcomes (a record-level late-run residual against run 8 alone); the abstract does not say so. And the abstract does not say whether the final-run P90-P10 is a residual or a raw spread.
A twofold gap between samples: the secondary-source spread is about half the discovery spread. The abstract gives no reason. Different event mix, field and data source are all possible (our reading); 19,769 records over 237 events are far fewer per event than 61,541 over 89.

Mechanisms

Measured: split times; model residuals are a statistic of those splits, not a physiological measure.
Authors' interpretation: residual late-run variability is a model-dependent post-race performance indicator; richer prior race history explains a meaningful component but does not eliminate it. They state that it should not be read as longitudinal stability, a fatigue diagnosis or a training-prescription threshold.
Inference: a residual is what a model fails to explain. It includes pacing choices, fuelling, course length, timing noise and bad days, as well as anything physiological.

Practical Application

Was my slow last run a fatigue problem? Splits cannot say. In 61,541 HYROX records early splits left a wide late-run spread (P90-P10 67.8 s; 33.2 s in a second sample); the authors bar reading it as fatigue or a training threshold.
Use it for humility in a race review: a late run far from what the first half suggested is common, not in itself a finding. One race's residual is not a stable trait, by the authors' statement.
Not for pacing targets: no reference value from this paper is given in the base, since the bands could not be read; for split reference values see REF-1496 (PRO) and REF-901 (percentiles).
Not from this paper; the base's own practice ([SYNTH-102]): review late-race runs against the athlete's own previous races and the race's conditions before calling anything a limiter. SYNTH-42 has the general rule: a profile is not shown to name the limiter.

Limitations

Abstract only; bands not read. The publisher's host refused with a bot check, not bypassed. Bands, division mix, sex split, unique athletes, modelling details, funding and interests are unknown here. If the PDF is obtained, the bands should be added and this entry regraded if needed.
Fixed order, finishers only. The late runs always follow the same stations; a residual cannot separate a station from a place in the race.
Model and sample dependent. The authors say the construct depends on the model; the two samples differ about twofold.
Records, not athletes. Repeats are not addressed in the abstract.
No physiology, no training, no intervention. Descriptive only, by the authors' own statement.
How this entry was checked and changed
Why this level

large observational sample by the rubric. Read from the abstract only: the publisher's host answers with a bot check, which was not bypassed, so the bands, division mix, handling of repeated athletes and exact model are not read. A model-derived residual, not a performance outcome; descriptive