An international research team led by Nina Claassen and Konstantin Warneke systematically searched PubMed/MEDLINE, Web of Science, and Scopus, ultimately including 63 studies evaluating the validity, reliability, and device agreement of commercially available velocity sensors, and 38 studies assessing velocity-based one-repetition maximum prediction models. The review was preregistered in the PROSPERO database and followed PRISMA reporting guidelines. Methodological quality was assessed using an adapted version of the COSMIN risk of bias checklist, and the team pooled results using multilevel random-effects meta-analysis, examining metrics such as the intraclass correlation coefficient (ICC), Lin’s concordance correlation coefficient (CCC), and Pearson’s correlation coefficient.

The first major finding concerns the sensors themselves. Commercially available devices, which include linear position transducers such as GymAware, Tendo, and Vitruve, as well as inertial measurement units like PUSH and EnodePro, demonstrated good-to-excellent pooled validity and device agreement, with ICCs of 0.91 to 0.92 and confidence intervals ranging from 0.83 to 0.97. Intra- and inter-day reliability were similarly strong, with pooled ICCs of 0.90 to 0.91. In plain terms, when the same lifter performs the same movement on different days, or when two devices measure the same lift simultaneously, the readings generally track each other well.

However, the averages conceal considerable variability. The researchers found substantial heterogeneity across studies, and their moderator analyses revealed that sensor technology mattered significantly. Linear position transducers, which calculate bar speed directly from displacement over time, generally showed more consistent performance, with validity correlations ranging from 0.77 to 0.99. Inertial measurement units, which estimate velocity indirectly by integrating acceleration signals, displayed a much wider spread, with correlations ranging from 0.40 to 0.99. The authors speculate that this inconsistency stems from the sensitivity of acceleration integration to signal noise, drift, and calibration assumptions, making IMU-based readings more dependent on the specific device, exercise, and conditions.

Exercise type and training intensity also shaped the results. Device agreement was significantly higher at moderate and high training intensities compared with low intensities, and certain movement patterns, such as squat-based assessments, influenced agreement estimates depending on the statistical metric used. Peak velocity measurements tended to show lower concordance than mean velocity in some analyses. These patterns suggest that no single number can summarize sensor performance; instead, practitioners must consider the specific combination of device, exercise, load, and velocity variable when interpreting readings.

The second part of the review tackled the central promise of velocity-based training: predicting the one-repetition maximum without actually testing it. Here the pooled results were again favorable. Velocity-based prediction models showed good-to-excellent interday reliability, with a pooled ICC of 0.90, and high average validity, with an ICC of 0.91 and a pooled Pearson correlation of 0.96 between predicted and actual maximal strength. On the surface, this supports the practice of building a load–velocity profile from a handful of submaximal lifts, potentially even within a warm-up, and using it to estimate maximal strength.

Yet the devil, as the authors emphasize, is in the details. Large heterogeneity in lower-body exercises significantly biased the results. Squat-based and deadlift-based predictions showed systematically lower validity and reliability than upper-body movements such as the bench press, where two-point and polynomial models achieved ICCs of 0.89 to 0.96. Individual study estimates for squat predictions ranged dramatically, from ICCs as low as 0.24 to as high as 0.99, depending on the number of load points used and the modeling approach. Polynomial regression performed significantly worse than other methods in the validity analysis. This means that a prediction that works well for a bench press may be considerably less trustworthy for a heavy back squat.

Perhaps the most sobering finding concerns what is missing from the literature. The authors highlight a dearth of systematic measurement error and agreement analyses. Most studies reported only relative metrics such as correlation coefficients, which reflect the strength of association between devices but say nothing about the absolute magnitude of differences. High correlations do not guarantee that two sensors can be used interchangeably. Where agreement data existed, the reported limits of agreement indicated potentially meaningful device-to-device differences, in some cases spanning more than 0.3 meters per second around mean velocities near 1 meter per second. The authors also note that individual studies have reported overestimations of maximal strength of up to 30 kilograms and mean prediction errors of up to 20 percent, underscoring that favorable averages do not eliminate the risk of practically relevant errors for individual lifters.

The review also calls out methodological inconsistencies in how validity and agreement have been defined and reported. Some studies treated linear position transducers as a gold standard, blurring the distinction between validity against a true criterion and simple device-to-device agreement. Others presented Bland–Altman plots without numerical agreement indices or interpreted them incorrectly. The authors argue that future validation studies must complement relative validity analyses with standardized agreement and measurement error reporting, including limits of agreement expressed relative to mean velocities, to allow meaningful practical interpretation.

The bottom line for athletes and coaches is a nuanced one. Commercial velocity sensors generally provide high relative validity and reliability, and velocity-based one-repetition maximum prediction achieves impressive average accuracy. But the evidence is sensor- and exercise-specific, results vary with intensity and modeling approach, and the scarcity of absolute error analyses prohibits final conclusions. The authors conclude that velocity-based monitoring and one-repetition maximum prediction require cautious interpretation. For now, the technology appears sound enough to inform training decisions, particularly for upper-body exercises and when using linear position transducers at moderate to high loads, but practitioners should treat predicted maximal strength values as estimates with real uncertainty rather than precise measurements, and researchers should prioritize rigorous agreement analyses before the field can issue definitive practical recommendations.

Subject of Research: Validity, reliability, and device agreement of commercial velocity sensors and velocity-based one-repetition maximum prediction models in resistance training

Article Title: Reliability, Device Agreement and Validity of Load–Velocity Profiles: A Systematic Review with Meta-analysis

Article References: Claassen, N., Siegel, S. D., Sproll, M., Lebelt, N., Bargende, A. V., Fasold, A. M., & Warneke, K. (2026). Reliability, Device Agreement and Validity of Load–Velocity Profiles: A Systematic Review with Meta-analysis. Sports Medicine – Open, 12(1), Article 131. https://doi.org/10.1186/s40798-026-01102-0

Image Credits: AI Generated

DOI: 10.1186/s40798-026-01102-0

Keywords: velocity-based training, load-velocity profile, one-repetition maximum, linear position transducer, inertial measurement unit, systematic review, meta-analysis, resistance training, measurement validity, reliability, strength monitoring, sports science