REF-1608 · Lea et al., 2022 · Sports Medicine - Open · Paper · PubMed · Free full text

In lifting, RPE rose with how hard the set was: a registered meta-analysis of 75 studies (99 cohorts, 2231 healthy participants) found a pooled correlation of r = 0.88 (95% CI 0.84-0.91) between RPE and load, heart rate, lactate, blood pressure or muscle activity, with very large heterogeneity (I² 96.1%); the correlation was higher when the loads compared spanned a wider range, with isometric work and when load or repetition time was varied, while age, sex, training status and the scale used were not significant moderators. Read in full

Main caveat: Very large heterogeneity; the best model explained 64% of the between-study heterogeneity, and what remained was still large (I² 86.2%).
fairly strong evidenceMeasured on non-athletes to elite athletesRead in full

Lea and colleagues asked whether ratings of perceived exertion measure intensity during resistance exercise as they are known to do in aerobic work. They pooled studies that reported RPE alongside an external intensity (load, reps, repetition time or rest) or a physiological response (heart rate, blood lactate, blood pressure, EMG).

Across 99 cohorts the weighted correlation was 0.88, with very wide heterogeneity; an EMG-only analysis gave 0.84.

The width of the load range compared explained most of the spread on its own, isometric actions gave higher correlations than dynamic ones, and studies that changed load or repetition time gave higher correlations than those that changed the number of reps or the rest.

Participant age, sex and training level, the scale, its number of points, estimation vs production mode, and active-muscle vs whole-body vs session ratings were not significant moderators.

The authors conclude that RPE is a valid, accessible way to prescribe and monitor resistance training, with correlations at least as high as those reported for aerobic exercise, and that its reliability needs more study.

Participants

99 cohorts, 2231 participants in the main analysis (outcome: intensity in 89 cohorts, heart rate 2, EMG 6, lactate 2); 8 cohorts, 340 participants in the EMG analysis. Healthy people of any age, no injury, illness or medication affecting performance.

Protocol

Search: four databases and websites to 28 February 2020, reference lists, contact with authors; English only; grey literature allowed.
Inclusion for pooling: a rating scale for perceived exertion or discomfort, resistance exercise of any type, at least one group without another intervention, and RPE with intensity, heart rate, blood pressure, EMG or lactate reported as a correlation or as means across two or more conditions.
Coding: participant age, sex and training level; muscle action, body segment, protocol, workload range (%1RM or %MVC); scale, points, fixed maximum, mode (estimation or production), rating type (active muscle, overall, session); outcome and the intensity variable changed.
Risk of bias: 9 items (eligibility, participant details, power calculation, exercise and intensity details, exact scale, RPE instructions, anchoring procedure, reliability); 0-3 high, 4-6 moderate, 7-9 low.
Analysis: random-effects models; univariate and multivariate meta-regression; leave-one-out sensitivity; funnel plot, trim-and-fill and fail-safe N.

Key Results

AnalysisResult
All outcomes, 99 cohortsr = 0.88 (95% CI 0.84 to 0.91); 95% prediction interval −0.07 to 0.99; τ² 0.526; I² 96.1%
EMG only, 8 cohortsr = 0.84 (95% CI 0.56 to 0.95); prediction interval −0.68 to 1.00; I² 97.3%
Workload range (56 cohorts)wider range, larger correlation, p < 0.001; explained 59% of the heterogeneity alone
Muscle action (98 cohorts)isometric higher than dynamic, concentric or eccentric, p = 0.004
Intensity variable changed (83 cohorts)load (p < 0.001) or repetition time (p = 0.002) above number of reps or rest (rest: two studies only)
Best multivariate model (50 cohorts: sex, rating type, workload range)R² 0.64, τ² from 0.391 to 0.142; I² still 86.2%
Not significant moderators: age, sex, resistance-training level, body segment, protocol, scale, number of points, fixed maximum, rating mode, rating type, outcome measure and risk of bias.
Publication bias: some funnel-plot asymmetry; trim-and-fill added no studies; the classic fail-safe N was 158,597.
Risk of bias: 44 studies low, 31 moderate, none high; only 13 of the 75 reported a reliability measure.
Reliability (narrative, not pooled): active-muscle ratings ICC 0.67-0.96, whole-body ratings 0.58-0.76, session RPE 0.95 (95% CI 0.90-0.97); two studies reported much lower values (0.07-0.80 and −0.05 to 0.46), the second with lower ratings on the second day, where more familiarisation days were argued to have been needed.
Anchoring (narrative): memory anchors, exercise anchors and combined anchors are all in use; one study comparing them at 40-90% MVC found no significant difference in mean ratings, with better reliability over the first two sessions in the two groups that had an exercise anchor. No study compared anchoring with no anchoring.
Training: ratings at a fixed load fell after training programmes, which the authors attribute to strength rising, so the same weight became relatively lighter.
Rating type: active-muscle ratings were consistently higher than whole-body ratings in the studies reviewed, the gap widening with intensity.

Practical Application

In 75 studies of healthy people lifting, RPE rose closely with how hard the set was (pooled r 0.88), and training status, age, sex and the scale used did not significantly change that, so validity held from novices to the trained. Correlations were higher when the sets compared spanned a wider range of loads, which the authors call expected from how a correlation works; that is not evidence about how to learn the scale, and a pooled r says little about the error of one rating.
Anchoring is not shown to be needed: in the one comparison reviewed, memory, exercise and combined anchors gave similar mean ratings, and an exercise anchor gave steadier ratings over the first two sessions only; nothing compared anchoring with none.
A non-significant moderator is not proof of no difference: novices and trained people were not shown to rate identically, only to rate validly.
When the same weight starts to feel easier over weeks, that is the expected reading as strength rises (the authors' explanation of training studies), not the scale drifting.
The base's own practice: use one kind of rating consistently. Active-muscle ratings ran higher than whole-body ratings, so switching between them moves the number without the effort changing.
It does not cover reps-in-reserve accuracy (REF-1607) or ratings on a bike (REF-1611, REF-1612).

Limitations

Very large heterogeneity; the best model explained 64% of the between-study heterogeneity, and what remained was still large (I² 86.2%).
Correlation-based validity rises with the range of loads compared, which the authors note; a pooled r says little about the error of one rating.
The risk-of-bias tool was built by the authors and mostly scores reporting.
Few studies used heart rate, blood pressure or lactate; most used external intensity.
Only 13 of 75 studies reported reliability, so test-retest consistency is thinly covered.
English-language studies only; some unusual modes (elastic bands, wall squats, simulated manual work) were pooled with gym lifts.
How this entry was checked and changed
Why this level

a registered systematic review and meta-analysis with the risk of bias read in the full text: 44 of the 75 pooled studies were rated low risk and 31 moderate, none high; held at A- because heterogeneity is very large (I² 96.1%), the risk-of-bias tool is the authors' own and scores mostly reporting, and the pooled studies are correlational validity studies, not controlled trials