REF-1231 · Tanner, Millet & Bourdillon, 2024 · Sports Medicine - Open · Paper · PubMed · Free full text
Can an HRV ramp test (such as DFA a1) replace a lab threshold test?
As a group estimate: across 50 studies HRV thresholds sat near lab thresholds on average (low certainty). Individual error was not pooled; treat yours as an estimate.
Tanner, Millet and Bourdillon pooled every study that determined heart-rate-variability thresholds (HRVT1, HRVT2) and lactate or ventilatory thresholds (LT1-VT1, LT2-VT2) in the same incremental test. From 952 records they included 50 studies (1160 subjects, 8 to 116 per study) with 314 agreement and 246 correlation assessments.
Random-effects meta-analyses found trivial standardized mean differences between HRV and reference thresholds (HRVT1 SMD 0.08, 95% CI −0.04 to 0.19, 22 studies; HRVT2 −0.06, −0.15 to 0.03, 42 studies) and very strong correlations (r = 0.84 and 0.85). Heterogeneity was high and prediction intervals wide (HRVT1 SMD −0.43 to 0.59; correlation 0.06 to 0.99).
Age, sex, weight class, training status, health status, ergometer and starting or increment workload did not change agreement; the reference used did (HRVT1 sat above VT1 and close to LT1, HRVT2 below LT2 and close to VT2), as did, for HRVT2, the HRV index (time-domain RMSSD worse than frequency-domain or non-linear indices) and increments of 3 min or more (worse).
Simple, visual determinations did as well as or better than computed or algorithmic ones. GRADE certainty was low for agreement and moderate for correlation, and limits of agreement could not be pooled because fewer than half of the comparisons reported them. The authors conclude that HRV thresholds might serve as surrogates for traditional thresholds when taken as a whole.
Our reading: good evidence that the method finds the right place on average, much weaker evidence on how close it lands for one person; it replaces single-study support (REF-873, one of the 50) for a no-lab heart-rate anchor at group level.
Protocol
Search: March to August 2023, eleven databases (Cochrane Library, EBSCO, Embase, Google Scholar, Ovid, ProQuest, PubMed, Scopus, SPORTDiscus, Virtual Health Library, Web of Science), strategies peer-reviewed by a health information specialist, plus the reference lists of three earlier reviews; no date limit; English, French, Italian and German. No registration is mentioned.
Eligibility: humans of any age, sex, health or training status; HRVT1 and/or HRVT2 and LT/VT determined in the same incremental test, with methods detailed; high-quality RR series from a validated HRV device; a statistical comparison of HRVT with LT/VT.
Screening: titles and abstracts by one author, checked with DistillerSR's machine-learning screening-error tool; full texts by two authors independently.
Included: 952 records, 852 excluded on title and abstract, 50 at full text, 50 included; 1160 subjects (on average 23 per study, range 8-116; age 32 (13-70) years; BMI 25 (18-39); VO2max 44 (10-79) mL/kg/min).
Effect sizes: SMD for agreement (positive = HRVT higher; trivial < 0.2), Pearson r for correlation (very strong > 0.8); random effects (DerSimonian-Laird), adjustment for multiple effects within a study; Q test, I² and prediction intervals; subgroup analyses and meta-regression for subjects, methods and protocols; funnel plots, Egger, Begg-Mazumdar, fail-safe N, leave-one-out; GRADE.
Quality: QUADAS-2 low risk of bias in all four domains for 21 of 50 studies, high in at least one domain for 4; STARD-HRV 78 ± 8% (62-94%); sample-size justification in 16% of studies, a stabilization period before HRV sampling in 40%, breathing control stated in 30%.
Funding and interests: open-access funding from the University of Lausanne; the authors declare no conflicts of interest.
Key Results
| Meta-analysis | Studies | Pooled effect (95% CI) | Prediction interval | I² |
|---|---|---|---|---|
| HRVT1 vs LT1-VT1, agreement | 22 | SMD 0.08 (−0.04 to 0.19), p = 0.18 | −0.43 to 0.59 | 89% |
| HRVT1 vs LT1-VT1, correlation | 22 | r = 0.84 (0.75 to 0.91) | 0.06 to 0.99 | 93% |
| HRVT2 vs LT2-VT2, agreement | 42 | SMD −0.06 (−0.15 to 0.03), p = 0.19 | −0.61 to 0.49 | 93% |
| HRVT2 vs LT2-VT2, correlation | 41 | r = 0.85 (0.80 to 0.89) | 0.27 to 0.97 | 92% |
Subgroups most relevant to a heart-rate anchor (SMD, or r, with 95% CI and number of studies):
| Subgroup | HRVT1 | HRVT2 |
|---|---|---|
| Threshold expressed as heart rate | SMD 0.01 (−0.08 to 0.10), 15; r = 0.88 (0.79 to 0.93), 13 | SMD 0.01 (−0.09 to 0.11), 20; r = 0.86 (0.81 to 0.90) |
| DFA a1 | SMD 0.16 (−0.08 to 0.40), 6; r = 0.86 (0.71 to 0.94), 7 | SMD 0.06 (−0.24 to 0.36), 5; r = 0.80 (0.64 to 0.89), 5 |
| RMSSD | SMD 0.04 (−0.10 to 0.19), 10; r = 0.71 (0.57 to 0.81), 10 | SMD −0.25 (−0.38 to −0.13), 14 (lower than reference); r = 0.70 (0.62 to 0.76), 13 |
| Reference: ventilatory vs lactate | vs VT 0.18 (0.07 to 0.30), 15; vs LT −0.10 (−0.29 to 0.09), 8 | vs LT −0.28 (−0.40 to −0.15), 16; vs VT 0.02 (−0.07 to 0.10), 31 |
| Determination: visual vs computed | r = 0.84 (0.76 to 0.89), 12 vs 0.70 (0.55 to 0.81), 11 | SMD 0.02 (−0.06 to 0.10), 31 vs −0.31 (−0.59 to −0.03), 12 |
Increments: for HRVT2, increments of 3 min or more gave a lower threshold than the reference (SMD −0.24, −0.39 to −0.09, 16 studies) unlike 1-min or 2-min increments; increment length did not matter for HRVT1.
Devices: device type (chest strap, ECG, sport watch) did not change agreement (HRVT1: chest strap 0.10, 5 studies; ECG 0.08, 4; sport watch 0.07, 13; HRVT2 sport watch −0.09, 28); one model did (HRVT1 from the Polar RS800 lower, SMD −0.44, 4 studies).The named devices are ECG systems and Polar chest-belt monitors (H3, H7, T61 transmitters; S810, RS800 and RS800CX receivers, which take beat-to-beat intervals from a chest belt; the last point is our knowledge of the devices, not stated in the paper).
Certainty (GRADE): low for agreement, moderate for correlation (upgraded for a large, precise effect).
Publication bias: Egger's test significant for both correlation analyses (p < 0.001 and 0.002), not for agreement; the authors judge the overall risk low.
A number that does not match: the abstract gives the HRVT1 correlation as r = 0.85 (95% CI 0.75-0.91); the Results and Discussion give r = 0.84 with the same CI. Two copy slips in the Results (an HRVT2 paragraph ending with a sentence about HRVT1, a fail-safe N for HRVT2 described as for HRVT1) do not change any number.
Mechanisms
Measured: during an incremental test HRV indices change in a way that locates two thresholds near the lactate and ventilatory thresholds on average; HRVT1 sat above VT1 and slightly below LT1, HRVT2 slightly above VT2 and below LT2, and the authors decline to conclude that HRV thresholds lie between the two references.
Authors' interpretation: total spectral power falls with intensity; the frequency of the high-frequency peak follows breathing frequency, which drives the ventilation curve at high intensity, so HF-based indices carry breathing information that RMSSD lacks, and time-domain indices have a low signal-to-noise ratio near HRVT2; DFA a1 reflects the fractal correlation of beat intervals and organismic demand; long increments steady the RR series but blunt the test, lowering HRVT2.
Inference: a trivial pooled SMD is a statement about group means; with a prediction interval of −0.43 to 0.59 for HRVT1, a new study could find a mean bias of up to about half a between-subject standard deviation either way, and an individual's error is larger still. The anchor is good for setting a starting zone, not for gating decisions on a few beats.
Inference: commercial apps compute DFA a1 automatically; in the pooled data computed determinations correlated less well than visual ones (HRVT1 r = 0.70 vs 0.84), so an app's number deserves a look at the underlying curve.
Practical Application
Can an HRV ramp test (such as DFA a1) replace a lab threshold test? As a group estimate: across 50 studies HRV thresholds sat near lab thresholds on average (low certainty). Individual error was not pooled; treat yours as an estimate.
Use it in heart-rate units: expressed as heart rate, HRVT1 agreed with the reference almost exactly (SMD 0.01, 15 studies; r = 0.88), which is the form an HR-zone athlete needs (A-, group level).
Run it well: stages under 3 min for the second threshold; prefer frequency-domain or non-linear indices to RMSSD near threshold 2; a beat-to-beat chest belt or ECG, not a wrist sensor (the base's inference from the devices studied).
Label it estimated until a hard effort or a lab test agrees with it: individual limits of agreement were not pooled, and certainty for agreement is low.
Limitations
Group-level only: limits of agreement were not analysed (reported in fewer than half of the comparisons), so how far one athlete's HRV threshold may sit from their lab threshold is not known from this review.
High heterogeneity (I² 89-93%) that the moderators "hardly explained" (the authors); wide prediction intervals.
Certainty: GRADE low for agreement, moderate for correlation; included studies small (23 subjects on average), 16% justified their sample size.
Population: mostly young healthy men; few women, older adults and patients.
Methods pooled: lactate and ventilatory thresholds treated as equivalent for the main analyses although they disagree; many HRV indices and determination methods grouped; units vary, so pooled differences are standardized, not in bpm.
Screening: titles and abstracts screened by one author with a machine-learning check; no registered protocol mentioned.
Devices: the named devices are ECG systems and Polar chest-belt monitors; the paper's "sport watch" category (13 HRVT1 and 28 HRVT2 agreement studies) is not described, so the absence of wrist optical HRV is our reading, not the paper's statement.
Bearing on [SYNTH-38]: supports an HRV-threshold row as an estimate, not as a measured anchor, and does nothing for the 20-30 min field protocol.
Reporting: abstract r = 0.85 vs body r = 0.84 for HRVT1; minor copy slips.
Funding: University of Lausanne open-access funding; no conflicts declared.
How this entry was checked and changed
Why this level
systematic review and meta-analysis of 50 validation studies (n = 1160, 23 per study on average), read from the full text (open access, CC BY 4.0). A- rather than A: heterogeneity was high (I² 89-93%) with wide prediction intervals, GRADE certainty was low for agreement and moderate for correlation, individual limits of agreement were not pooled, and the included studies are small. The level covers group-level agreement; accuracy for one athlete is not established by it
Full text (checked 2026-10-27)
Fifty included studies (1160 subjects) assessed 314 agreements (95 for HRVT1, 219 for HRVT2) and 246 correlations (82 for HRVT1, 164 for HRVT2) (full text)
Subjects: on average 23 per study, range 8–116; age 32 (13–70) years; V̇O2max 44 (10–79) (full text)
HRVT1 agreement: SMD = 0.08, 95% CI -0.04–0.19, p = 0.18; prediction interval − 0.43 to 0.59; I2 89% (full text)
HRVT1 correlation: Pearson's r = 0.84, 95% CI 0.75–0.91; prediction interval 0.06 to 0.99; I2 93%; the abstract gives r = 0.85 (full text)
HRVT2 agreement: SMD = -0.06, 95% CI -0.15–0.03, p = 0.19; prediction interval − 0.61 to 0.49 (full text)
Threshold as heart rate: HRVT1 Heart Rate (0.01, -0.08–0.10, n = 15), correlation Heart Rate (r = 0.88, 0.79–0.93, n = 13); HRVT2 Heart rate (0.01, -0.09–0.11, n = 20) (full text)
DFA-α1 (0.16, -0.08–0.40, n = 6) for HRVT1 agreement; DFA-α1 derived HRVT1 (0.86, 0.71–0.94, n = 7) for correlation (full text)
Visual determination of HRVT1 (0.84, 0.76–0.89, n = 12) correlated more strongly than computed determination (0.70, 0.55–0.81, n = 11, p = 0.04) (full text)
For HRVT2, 3 min increments or more (-0.24, -0.39 – -0.09, n = 16) gave lower thresholds (full text)
The limits of agreement were not analysed because they were available in less than half of the agreement analyses (full text)
Risk of bias was low in the four QUADAS-2 domains for 21 of the 50 included studies; STARD-HRV average 78 ± 8% (full text)
Open access funding provided by University of Lausanne; no potential conflicts of interest (full text)