External training load metrics have become fixtures of modern sport, streaming from GPS vests, accelerometers, and velocity trackers in every training session. A systematic review now argues that the numbers practitioners rely on are far more limited than their labels suggest. After examining 46 studies published between 2011 and 2026, researchers concluded that most external training load metrics are supported as descriptions of external movement, work, or exposure rather than direct indicators of metabolic demand, neuromuscular load, tissue stress, fatigue, or adaptation. The review, published in Sports Medicine – Open, mapped how each study connected a recorded signal to the biological meaning claimed for it, and found that this chain is usually incomplete.
The authors, Houchen Li, Qingqiong Yang, and Bin Wang, searched MEDLINE, Embase, Web of Science, Scopus, and Academic Search Ultimate from database inception to March 2026, with final searches run on 1 March except for Embase on 7 March. They included peer-reviewed English-language primary studies that evaluated an external training load indicator and its interpreted construct in athletes or physically active participants. Reports limited to device accuracy, reliability, measurement error, or descriptive monitoring were excluded, because the review asked a different question: not whether a device measures distance accurately, but whether the evidence supports the meaning practitioners attach to that distance.
Each of the 46 included studies contributed one primary evidence chain, linking the focal indicator to its intended construct, the reference or contrast used, and the study finding. This study-level rule prevented articles reporting many metrics or statistics from counting as independent replications. Construct attribution was explicit in 14 studies and explicit but qualified in 26; four were implicit but qualified, and two were reverse-target, meaning the external load indicator mainly served as an exposure or comparator while another construct was the actual validation target. Only 13 of the 46 primary claims, about 28 percent, used a direct construct-matched criterion, meaning a reference that measured the target quantity itself or a construct-matched physiological, biomechanical, or performance standard.
The remaining claims leaned on weaker routes. Alternative proxies, such as heart rate, training impulse, or session rating of perceived exertion, supported 16 studies, roughly 35 percent. Condition or group contrasts, which test whether expected differences appear between sessions or groups, supported 11 studies, about 24 percent. Mixed reference or contrast evidence accounted for six studies, 13 percent. The authors stress that these classes describe the evidential relation to the construct rather than study quality, but the pattern is clear: most biological interpretations rest on indirect comparisons rather than direct measurement of the claimed construct.
General exercise intensity was the most frequent construct family, appearing in 15 of 46 studies, or about 33 percent. Cardiorespiratory or metabolic demand and mechanical or movement demand each accounted for 10 studies, roughly 22 percent apiece. Nine studies, about 20 percent, examined neuromuscular or fatigue states, and two addressed other recovery or adaptation outcomes. The evidence routes differed sharply by construct. Eight of the 10 cardiorespiratory or metabolic claims used a direct construct-matched criterion, whereas alternative proxies supported 10 of the 15 general-intensity claims, and condition or group contrasts supported five of the 10 mechanical or movement-demand claims. In one representative study estimating total energy expenditure, doubly labelled water served as a direct criterion; in another, an analysis relating Running Load Index to leg stiffness relied on an alternative proxy; and a third compared sessions with and without tackles as a condition contrast.
The review’s central argument is that technical accuracy and construct validity answer different questions. Reliability and measurement error indicate whether a signal can be reproduced under specified conditions. Construct validity asks whether the available evidence supports the assigned meaning. A stable association alone does not show that an external training load metric measures a particular physiological or biomechanical construct. The authors illustrate the point with GNSS-derived high-speed-running distance: such a metric may accurately record distance above a threshold, but that accuracy does not establish metabolic stress, hamstring tissue load, or neuromuscular fatigue. Each interpretation has a different mechanism and requires different evidence. Mathematical stability likewise does not establish the physiological meaning of a proprietary composite score.
Methodological appraisal reinforced the caution. The authors routed analyses to review-adapted COSMIN worksheets, PROBAST+AI for prediction models, and Joanna Briggs Institute tools for other designs, covering 124 analysis units and 1,082 item judgements. Hypotheses-testing modules were rated Adequate in only 3 of 46 studies, Doubtful in 39, and Inadequate in 4. Criterion-validity modules were Adequate in 3 of 13 and Doubtful in 10, reliability modules were Adequate in 2 of 13 and Doubtful in 11, and all 17 measurement-error modules were Doubtful. Responsiveness modules were Adequate in 1 of 10 and Doubtful in 9. PROBAST+AI covered 18 model phases across nine studies; development quality was rated High concern in three phases and Unclear in six, while evaluation risk of bias was High in four, Unclear in four, and Low in one.
The appraisal also surfaced concrete examples of weak inference. One study used backward selection from multiple candidate predictors in only ten players; another applied forward selection on repeated records from 13 players without accounting for athlete effects; and athlete- and session-specific kayaking models showed apparent fit on their own data. In a soccer session-classification study, data splitting and stratified cross-validation were reported, but athlete and session separation and uncertainty quantification were not fully established. In a carbohydrate and energy expenditure study, models developed in 15 cyclists were evaluated without recalibration in 59 additional participants, and subsequent recalibration on those data amounted to model updating rather than independent validation. Even independently evaluated prediction, the authors note, does not validate a broader biological interpretation.
Reporting practices added another layer of concern. Nineteen studies, about 41 percent, explicitly distinguished technical performance from construct interpretation, conceptually or through separately operationalised tests; the remaining 27 did not meet this reporting criterion. Only 12 studies, about 26 percent, explicitly justified their comparator selection. In basketball testing, relatively good reliability coexisted with limited construct validity for detecting fatigue-related changes, showing that an explicit distinction does not necessarily imply a direct reference or a strong construct test. The authors also found that most studies included only men: 27 of 46 studies included only male participants, one included only females, 12 had mixed-sex samples, and six did not explicitly report sex, which limits how far conclusions can transfer to female athletes.
The practical implications are direct. Practitioners, coaches, and sport scientists can use external training load metrics most defensibly as measures of external exposure: distance, speed, acceleration, power, bar velocity, and total-load outputs can describe what an athlete did and support within-athlete monitoring. Given the identified appraisal limitations, the authors conclude that model outputs may inform research or complement other assessments, but current evidence does not support their use alone to diagnose individual fatigue, measure tissue loading, or determine training adjustments. Such uses require construct-matched validation and evidence for the intended decision. The authors also call on device manufacturers and software platforms to avoid naming or marketing proprietary composite scores as biological load without construct-matched evidence, and to publish transparent equations, units, preprocessing rules, and version histories so that evidence can accumulate across devices and studies. Metric selection, they argue, should begin with the monitoring decision, and the proposed interpretation should then be supported by an explicit construct, a plausible mechanism, a testable validity hypothesis, and an appropriately matched reference.
Subject of Research: Construct validity of external training load metrics used in sport
Article Title: From Signal to Biological Meaning: A Systematic Review of Construct-Validity Evidence for External Training Load Metrics in Sport
Article References: Li, H., Yang, Q., & Wang, B. (2026). From Signal to Biological Meaning: A Systematic Review of Construct-Validity Evidence for External Training Load Metrics in Sport. Sports Medicine – Open, 12(1), Article 163. https://doi.org/10.1186/s40798-026-01133-7
Image Credits: AI Generated
DOI: 10.1186/s40798-026-01133-7
Keywords: external training load, construct validity, systematic review, GPS, accelerometry, athlete monitoring, fatigue, metabolic demand, neuromuscular load, sports science, measurement validity, training load metrics
News Source: Ophelia Keating. (November 4, 2026). Training Load Metrics Mostly Measure Movement, Not Biological Strain, Review Finds. Scienmag.
