Global Positioning System–Derived Workload Metrics and Injury Risk in Team-Based Field Sports: A Systematic Review

Methodologic Quality

Two independent raters graded each study for level of evidence according to the National Institute of Health Quality Assessment Tool for Observational Cohort and Cross-Sectional Studies⁸ (Supplementary Table 1).

Data Extraction and Synthesis

For all included articles, the following data were extracted: study design, participants (sample size, sex, sport), GPS–derived metrics collected, and key findings, including quantitative point estimates and confidence intervals (CIs) of increased or decreased injury risk. We provided CIs when they were reported in the original article. When CIs are not presented in our “Results,” they were not reported in the original article. Additionally, we supplied exact P values or dichotomous findings of statistical significance (eg, P < .05) in our “Results” based on how they were reported in the included articles. Similarly, if the authors used magnitude-based inference descriptors (eg, likely harmful), we used the same terminology (in quotation marks) in our summary. Meta-analysis was not possible because of the heterogeneity of study methods.

Search Results and Methodologic Quality

We identified 499 articles in the initial search and selected 22 original research articles for inclusion (Figure). The study design characteristics are detailed in Table 1. Injury risk was assessed in 10 studies of Australian rules football, 6 of soccer, and 5 of rugby and 1 study each of Gaelic football and American football. The methodologic quality assessments and individual study findings are summarized in Supplementary Tables 1 and 2, respectively.

View Figure

• Download as Powerpoint
• Download Figure

Total Distance

Total distance (TD) of player movement during a game or practice session was evaluated as an injury risk factor in 8 articles.^9–16 The authors most often divided the data into 1-, 2-, 3-, and 4-week timeframes; however, within these timeframes, TD was reported differently, using either total cumulative distance^11,13,14,16 or team-based Z scores.^9,12

High-Speed Running Distance

Fifteen groups of authors^{9–11,13–19,21–25} reported on relationships between the quantity of high-speed running distance (HSD) and injury risk. The minimum velocity threshold for “high speed” varied from 14.4 to 24 km/h across these studies.

Sprint-Running Distance

A handful of author groups^{10,11,15,18,25,26} specifically evaluated sprint-running distance (SRD) by further classifying high-speed running. Due to methodologic differences among articles, sprint running was sometimes grouped with high-speed running for analysis, and other times, it was analyzed separately. The definition of sprint-running speed differed among studies; speeds varied from 19.8^11,26 to 25.3 km/h¹⁰ and also as speed above an athlete's 75% maximum performance.^18,20

Accelerations and Decelerations

Taking acceleration distance into consideration, Gabbett and Ullah²⁴ indicated that the distances covered in mild (0.55–1.11 m/s²), moderate (1.12–2.78 m/s²), and maximal accelerations (≥2.79 m/s²) were related to non–time-loss injuries; the higher the number of accelerations, the lower the risk of non–time-loss injury (RR = 0.2; 95% CI = 0.1, 0.4 for mild accelerations; RR = 0.3; 95% CI = 0.1, 0.6 for moderate accelerations; RR = 0.4; 95% CI = 0.2, 0.8 for maximal accelerations).

Bowen et al^9,10 evaluated 1-, 2-, 3-, and 4-week timeframes by dividing the number of accelerations into 5 groups based on Z scores. Several relationships were identified in 1 season of soccer data.⁸ At the 1-week mark, the overall relative risk of injury (RR = 0.35, P < .05) was decreased in athletes who produced a “low” number of accelerations (Z score = −1.99 to −1.00). A “high” number of accelerations (Z score = 1.00–1.99) increased the overall injury risk in weeks 1 (RR = 1.83, P < .05) and 4 (RR = 1.66, P < .05), whereas for a “very high” number of accelerations (Z score >2.00), the relative risk of overall injury increased across 1 (RR = 3.06, P < .05), 2 (RR = 3.19, P < .05), and 3 (RR = 3.84, P < .05) weeks versus all other groups. In addition, the noncontact injury risk was also elevated for a “very high” number of accelerations in weeks 3 (RR = 5.11; 95% CI = 1.75, 14.96; P = .003) and 4 (RR = 4.25, 95% CI not reported; P < .05). Conversely, a “low” number of accelerations over 3 weeks (744–2861) reduced the noncontact (RR = 0.21; 95% CI = 0.05, 0.87; P = .03) and overall (RR = 0.31; 95% CI = 0.13, 0.76; P = .01) injury risks. Analysis of 3 seasons of data¹⁰ revealed no differences between the number of accelerations and injuries identified across any of the weekly timeframes. The same researchers¹⁰ looked at deceleration counts in the same manner. The only significant result was an increase in contact injury risk for the 1-week “moderate-low” (Z score = −0.99 to 0) number of decelerations (RR = 2.04; 95% CI = 1.0, 4.0; P = .04).

Jaspers et al¹⁶ assessed the numbers of accelerations and decelerations in soccer athletes. “Likely harmful effects” existed for “high” 2-week decelerations (>1462, OR = 1.49; 90% CI = 0.92, 2.42), “high” 3-week decelerations (>2140, OR = 1.68; 90% CI = 1.08, 2.63), and “high” 4-week decelerations (>2813, OR = 1.73; 90% CI = 1.00, 2.99) when compared with the “low” deceleration group (<2227).

Acute : Chronic Workload Ratio

Thirteen groups* explored acute : chronic workload ratio (ACWR) metrics in their analyses of injury risk factors. The ACWR divides the acute workload (typically 1 week) by the chronic workload (typically 3 to 6 weeks). The theory^1,30 is that chronic training loads are analogous to a state of fitness, and acute training loads are analogous to a state of fatigue. Currently, 2 methods are available for calculating the ACWR. First is a rolling average ACWR, which is simply the rolling average of the acute phase divided by the rolling average of the chronic phase. Importantly, the workload metrics from all days within a phase are equally weighted in this model. The second method, proposed by Williams et al,³¹ is an exponentially weighted moving average (EWMA) ACWR, which includes a time-decay constant that weighs workload values closer to the end of the chronic phase (more recently) more heavily than training load values at the beginning of the chronic timeframe (longer ago). In the next paragraphs, we highlight the results of studies with significant findings.

Malone et al²⁶ used a 3- to 21-day ratio when calculating a rolling average ACWR for HSD and found that players with ratios above 0.85 were at increased risk of injury. When the ACWR was between 0.86 and 1.00, the OR of lower extremity injury was 1.20 (90% CI = 1.10, 2.03; P = .21); between 1.00 and 1.25, the OR was 2.27 (90% CI = 2.13, 3.04; P = .001); and for ACWR <1.25, the OR was 3-fold greater (3.02; 90% CI = 2.53, 4.98; P = .001). When the same analysis was conducted on SRD, the OR decreased for ACWR between 0.71 and 0.85 (0.85; 90% CI = 0.33, 0.95; P = .04). However, for ACWRs between 0.86 and 1.25 and >1.35, the OR increased (OR = 1.14; 90% CI = 1.11, 2.14; P = .02 and OR = 5.00; 90% CI = 3.01, 7.38; P = .02, respectively).

Using absolute TD to calculate the 1 : 4-week rolling average ACWRs and interpreting them through a lens of magnitude-based inferences, Hulin et al¹² determined that, in the current week, a “very high” ACWR (≥2.11) was associated with an injury risk 6.9 times greater than a “very low” ACWR of ≤0.30 (RR = 6.9; 90% CI = 5.2, 8.6), 3.4 times greater than a “low” ACWR of 0.31 to 0.66 (RR = 3.4; 90% CI = 1.4, 5.4), 2.3 times greater than a “moderate” ACWR of 1.03 to 1.38 (RR = 2.3; 90% CI = −2.3, 5.7), and 2 times that of a “high” ACWR of 1.75 to 2.10 (RR = 2.0; 90% CI = −15.2, 37.2). In addition, a “very high” 2-week ACWR (≥1.88) was associated with a risk of injury that was 2.2 times greater than a “low” ACWR of 0.46 to 0.74 (RR = 2.2; 90% CI = −2.7, 7.1), 1.9 times greater than a “moderate-low” ACWR of 0.75 to 1.01 (RR = 1.9; 90% CI = −3.6, 7.4), and 2.4 times greater than a “moderate” ACWR of 1.02 to 1.30 (RR = 2.4; 90% CI = −0.6, 5.4). For the subsequent week, a “very high” ACWR had a 10-fold increase in injury risk compared with a “very low” ratio (RR = 9.8; 90% CI = 6.2, 13.4).

Murray et al¹³ used the ACWRs for TD, HSD, and player load (operationally defined in Supplementary Table 2) to examine injury risk. For injuries occurring in the current week, players with an ACWR of >2.0 for TD were 5 to 8 times more likely to sustain an injury than players with an ACWR < 0.49 (RR = 7.98; 90% CI = 5.86, 10.88; P = .015) and between 0.5 and 0.99 (RR = 5.04; 90% CI = 4.16, 6.11; P = .012). For HSD, an ACWR of >2.0 was associated with a 6 to 12 times greater injury risk than ACWRs of <0.49 (RR = 11.62; 90% CI = 10.04, 13.45;, P = .006), 0.50 to 0.99 (RR = 9.63; 90% CI = 9.21, 10.07; P = .002), and 1.0 to 1.49 (RR = 6.54; 90% CI = 6.19, 6.92; P = .003). Similarly, athletes with an ACWR of >2.0 for player load had a greater risk of injury than those with an ACWR of 0.50 to 0.99 (RR = 6.27; 90% CI = 5.62, 6.00; P = .006) and 1.0 to 1.49 (RR = 7.72; 90% CI = 7.57, 7.88; P = .001). In this final result, we recognize that the point estimate lies outside the CI and contacted the authors about this discrepancy, but they did not respond to our query.

For injuries occurring in the subsequent week, Murray et al¹³ indicated that, during the preseason, athletes with an ACWR of >2.0 had an increased likelihood of injury versus players with an ACWR of 1.0 to 1.49 for TD (RR = 4.87; 90% CI = 2.33, 10.21; P = .05) and player load (RR = 12.46; 90% CI = 8.35, 18.59; P = .02). Similarly, an ACWR of >2.0 for HSD compared with an ACWR ratio of 0.50 : 0.99 was associated with an increased likelihood of injury (RR = 6.46, 90% CI = 4.63, 9.02, P = .02). During the in-season period, findings were similar. Specifically, when the ACWR exceeded 2.0, compared with an ACWR between 1.0 and 1.49, the likelihood of injury increased 4-fold to 7-fold for TD (RR = 5.49; 90% CI = 4.19, 7.20; P = .02), HSD (RR = 4.36; 90% CI = 3.50, 5.43; P = .02), and player load (RR = 5.80; 90% CI = 4.62, 7.27; P = .01).

Over a single season, Bowen et al⁹ evaluated TD, HSD, number of accelerations, and total load (operationally defined in Supplementary Table 2) in the context of ACWR using rolling averages and a 1 : 4-week ratio. Relative risk was determined by comparing injury risk with all other ACWR categories. For overall ACWR, the RR of contact injury was increased to 4.98 (P < .05) for a “very high” TD ratio (≥2.0). Conversely, a decrease in overall injury risk (RR = 0.47, P < .05) was present for athletes with a “low” HSD ACWR (−1.99 to −1.00). For accelerations, an increased RR of contact injury of 4.98 occurred at “very high” ACWR (≥2.0; P < .05). Lastly, for overall total load ACWR, the RR of contact injury was 1.92 among players with “moderate-low” ACWR (P < .05) and the RR of noncontact injury was 1.87 among players with “moderate-high” ACWR (P < .05).

Another study by Bowen et al¹⁰ spanned 3 soccer seasons and demonstrated significant findings in TD, low-intensity distance, and accelerations in relation to injury when using a rolling average ACWR method. For TD, a “very high” ACWR (≥2.0) increased the RR of noncontact and overall injury (3.67, P < .05, and 2.40, P < .05, respectively), a “moderate to high” ACWR (0.00–0.99) increased the RR of contact injury (2.03, P < .05), and a “low” ACWR (−1.99 to −1.00) increased the overall injury risk (RR = 0.19, P < .05). For low-intensity distance, a “moderate to high” ACWR (0.00–0.99) was associated with increased RRs for both contact and overall injuries (2.60, P < .05, and 1.91, P < .05, respectively), and a “very high” ACWR (≥2.0) had increased RRs for both noncontact and overall injuries (3.93, P < .05, and 2.56, P < .05, respectively). For the numbers of accelerations and decelerations, both categories of “moderate to high” ACWR (0.00–0.99) and “very high” ACWR (≥2.0) had increased injury risk associations. For a “moderate to high” ACWR, the RR increased to 1.57 (P < .05) for overall injury and 1.99 (P < .05) for contact injury for accelerations and decelerations, respectively. In the “very high” ACWR category of accelerations, noncontact (RR = 3.86, P < .05) and overall (RR = 2.52, P < .05) injury risk increased, and in the same category for decelerations, the risk also increased for both noncontact and overall injuries (RR = 3.73, P < .05, and RR = 2.44, P < .05, respectively).

Jaspers et al¹⁶ investigated the average 1 : 4-week ACWR for the workload metrics of TD, HSD, accelerations, and decelerations. A “likely harmful” effect for a “high” ACWR was present for HSD (>1.18, OR = 1.71; 90% CI = 0.90, 3.26), a “very likely beneficial” effect occurred for a “medium” ACWR for decelerations (0.86–1.12, OR = 0.38; 90% CI = 0.20, 0.72), and “likely beneficial effects” existed for a “medium” ACWR for accelerations (0.87–1.12, OR = 0.49; 90% CI = 0.24, 1.02).

Multivariate Models

Bacon and Mauger¹⁷ calculated a simple linear regression to predict the incidence of overuse injuries based on TD and HSD assignment to “low,” “normal,” or “high” groups. A significant regression equation contained only the TD variable (F_1,39 = 6.482, P = .02, R² = 0.14). Injury incidence rate per 1000 hours was decreased by −5.835 times when moving upward from 1 TD loading group to the next, meaning a higher TD loading group lowered the risk of an overuse injury.

Malone et al²⁵ also assessed HSDs in combination with training loads. Players who had higher 21-day chronic training loads (≥2584 AU) were at reduced risk of injury when they covered 1 weekly HSD of 701 to 750 m compared with the reference group of <674 m (OR = 0.65; 90% CI = 0.25, 0.89; P = .024). Conversely, athletes who exerted low chronic training loads (≤2584 AU) and covered the same distance of 701 to 750 m were at greater risk of injury versus the reference group of <674 m (OR = 3.12; 90% CI = 2.99, 4.54; P = .04). Similar trends were observed for SRD with higher 21-day chronic training loads: players pursued increased high-speed and sprint running distances with reduced injury risk.²⁵

Colby et al²⁶ used a mixed-model generalized estimating equation to analyze the relationship between weekly data and injury in the subsequent week. In their multivariate model of highest predictive accuracy, they found that a “low” chronic distance coupled with a “very high” distance ACWR was associated with an increased risk (adjusted incidence rate ratio [adjusted IRR] = 2.60; 95% CI = 1.07, 6.34) compared with an above-average chronic load and moderate ACWR. In addition to workload variables, playing experience, heavy nonfootball activity, and a history of lower limb pain retained significance in the multivariate model as in the univariate models (adjusted IRR = 2.02–2.25; 95% CI = 1.02, 4.95). Colby et al²⁶ noted the predictive accuracy of the multivariate model (area under the curve [AUC] = 0.70; 95% CI = 0.64, 0.75) was better (P < .001) than in all univariate models (AUC = 0.52–0.60) when tested on in-sample data. In addition, cross-fold validation results of simulated data indicated a very similar fit (k = 10: univariate root mean square error = 0.16 ± 0.02 versus multivariate root mean square error = 0.16 ± 0.02) on out-of-sample data.

In another study, Colby et al,¹⁹ using a multivariate model across the full in-season phase of Australian rules football, demonstrated that a “very low” (<108 km) late preseason (January to mid-February) distance placed players at greater injury risk compared with moderate (125–164 km) loads (OR = 5.6; 95% CI = 1.4, 22.8; P = .02). Similarly, “low” (76–88 km) precompetition (mid-February to mid-March) distances compared with moderate (89–112 km) distances increased the injury risk (OR = 6.0; 95% CI = 1.6, 23.3; P = .01). In addition, a “very high” distance covered (>170 km) during early preseason (November and December) was associated with greater in-season injury risk (OR = 3.2; 95% CI = 1.3, 8.5; P = .02) versus a moderate distance (95–143 km).

To advance the multivariate modeling, Colby et al²¹ published another paper in 2018 that addressed high-risk workload scenarios. Of the proposed high-risk scenarios, exposure to maximum speed (the number of times a player was at >85% of maximum speed) had the most interesting findings. The authors divided the scenario into 8- and 4-week timeframe subcategories, labeling them as “most significant” and “most practical,” respectively. These designations were selected because, although 8 weeks had the best predictive features, in real-life application, 8 weeks may be too long to collect and apply the data, depending on the setting. The exposure to maximum velocity was divided into 5 categories (“very low” to “very high”), with the “moderate” category serving as the reference group. For both the 8- and 4-week timeframes, the “low,” “moderate,” and “very high” categories had good specificity (ability to identify players who were not injured), but low sensitivity (did not identify players who were injured; 8 weeks: “low”: sensitivity = 0.13, specificity = 0.83; “moderate”: sensitivity = 0.05, specificity = 0.80; “very high”: sensitivity = 0.18, specificity = 0.86; 4 weeks “low”: sensitivity = 0.11, specificity = 0.83; “moderate”: sensitivity = 0.11, specificity = 0.84; “very high”: sensitivity = 0.11, specificity = 0.90).

Windt et al¹⁴ created 2 multivariate models to quantify the effect of preseason participation on injury risk while controlling for training-load variables. The first model evaluated the likelihood of injury in the current week, and even though it was associated with reduced odds of an injury, it was not significant (OR = 0.85; 95% CI = 0.70, 1.02). Similarly, when preseason participation and acute distance were controlled, a greater percentage of distance run at high speeds appeared to be associated with an increased injury risk (OR = 1.27; 95% CI = 0.99, 1.63). Finally, as with the univariate models, greater acute distance was associated with a reduced likelihood of injury (OR = 0.56; 95% CI = 0.3, 0.87). The second model demonstrated the likelihood of injury in the subsequent week from the metrics of preseason participation, acute distance, and acute percentage of distance run at high speeds. In this model, when distance and percentage of distance at high speed were controlled, increased preseason participation (at least 10 full preseason sessions completed) was associated with a reduced likelihood of injury (OR = 0.83; 95% CI = 0.70, 0.99). In this model, neither acute distance nor percentage of distance run at high speeds was significantly associated with injury risk in the subsequent week.

A few researchers furthered attempts at predictive modeling by using analytic methods such as random forest^32,33 and support vector machines³³ along with generalized estimating equation (GEE) modeling. Thornton et al³² used GEE and random forest models to determine which training load variables were most important in understanding injury among positional groups of a rugby team (hit-up forwards, adjustables, wide-running forwards, and outside backs). Specifically, for the adjustables group (in order of importance), the total load (TL) variables recognized were 7-day TD, 7-day high metabolic power distance, high metabolic power distance ratio, 28-day HSD, and 21-day high metabolic power distance, in which the model quasilikelihood under independence model criterion (QIC) was 566.5, and the statistical significance of these measures ranged from P = .001 to .091. For the hit-up forwards group, identified variables were session (s)RPE-TL ratio, 14-day TD, 14-day high metabolic power distance, 7-day sRPE-TL, and 28-day high metabolic power distance, in which the QIC was 441.7, and the significance of these measures ranged from P = .006 to .138. The outside-back models indicated that 21- and 28-day sRPE-TL, 7-day HSD, high metabolic power distance ratio, and TD ratio were most associated with injury risk, with a QIC of 406.6 and significance of these measures ranging from P = .092 to .225. For the wide-running forward group, sRPE-TL ratio, HSD ratio, 14-day high metabolic power distance, 14-day TD, and 7-day sRPE-TL were the most likely contributors to injury risk, with a QIC of 410.5 and significance ranging from P = .068 to .830. The random forest models of Thornton et al³² indicated that, for the adjustables group, the relative importance of TL variables was similar. The mean (± SD) receiver operating characteristic (ROC) of the random forest models was 0.74 ± 0.24. For the hit-up forwards group, 7-day sRPE-TL and 14-day high metabolic-power distance were associated with the greatest importance in injury, in which the mean model ROC was 0.65 ± 0.06. Variables recognized for the wide-running forward group varied substantially among athletes. The mean model ROC was 0.64 ± 0.05. Similarly, for the outside back group, the importance of TL variables varied between players, showing a large discrepancy for TD ratio and 28-day sRPE-TL. These models had a mean ROC of 0.64 ± 0.04.

Injury Tracking

Possibly the simplest topic to resolve is the various definitions of injury used. A mix of acute, overuse, chronic, upper body, lower body, time-loss, and match-loss injury definitions were used with minimal uniformity across studies. Given that the GPS–derived metrics monitor total workload over time and generally workload applied to the lower extremity, these metrics appear best suited to forecasting lower extremity overuse injuries. One exception to this would be when monitoring changes in HSD and its effects on acute hamstrings injury, as this injury was previously linked to the eccentric loads that occur during sprinting.^34–36

Regarding injury variables, a binary (injured or not-injured) classification based on a time-loss metric is easy to understand and collect uniformly. However, given the breadth of the GPS metrics and nuances of injury definitions in sport, especially overuse injury, these data may be best captured on a daily ranking scale of player availability based on ongoing musculoskeletal concerns, even if those concerns do not rise to the level of an injury report or time-loss injury. The current dichotomous reporting of injury status (injured, not injured) is convenient and easy to record, yet it misses the nuances of contemporary sports medicine in which athletes are often partial participants in practices due to their injury status. Our recommendation in this area is 2-fold. First, within the realm of practical sports science in which data are collected daily by sports medicine and sports performance staff and then analyzed retrospectively, we advise the use of the 4-tier injury and participation classification scheme described by Ahmun et al³⁷ in response to the recent consensus statement on injury surveillance in cricket.³⁸ The 4 tiers are (1) fully available for training and matches, with no injury or illness, (2) fully available for training and matches but with an injury or illness, (3) available for selection in a major match but with modified activity due to injury or illness, and (4) unavailable for selection in a major match due to injury or illness. These categories mirror injury reports commonly used in the clinical practice of sports medicine within organized sport and allow high-level tracking of both transient and time-loss injuries. Second, from a research and modeling perspective, knowing the injury mechanism, cause, and tissue type is important as it may lead to an understanding of how to best model different injury types using the collected data. To accomplish this, studies must be conducted in a prospective manner and may involve more daily engagement of the researchers to avoid burdening the sports medicine staff.

Statistical Analysis

Stratifying workload metrics into binned categories (ie, low, medium, high) allows for easier modeling and interpretation of results in an individual study; however, these groupings are typically based on the Z scores of 1 team. Differences in injury risk may have occurred within specific workload groupings, but the generalizability of these results to teams and athletes other than those studied is limited. For example, the difference in SRD between workload groups of approximately 200 m over 1 week or exposures to maximum velocity that vary by less than 5 m is most likely not practically meaningful; when workload metrics are grouped in this fashion, understanding the injury risk as a player moves 1 level on the scale is impossible.^20,21 Valuable information may be lost with stratification,³⁹ especially without research-based cut points. These models treat all data within the group the same and look for differences between groups when, in reality, this early in the field of GPS variables, researchers do not yet know if the upper threshold of a “low” group and the lower threshold of a “medium” group are truly different enough to be separated. The use of workload metrics on a continuous scale, rather than in stratified categories, is likely a more appropriate way to apply these measures to assess injury risk. If groupings are needed, basing them on population-based metrics (ie, “high-speed” running velocity for professional soccer players across an entire league, not just a single team) or groupings that are specific to each individual in the analysis is encouraged. Further analysis of how groupings of various levels affect injury risk calculations is indicated.

Magnitude-based inferences are another statistical choice that, in theory, attempt to make results more practical when the sample is small^40,41; however, their mathematical grounding has been refuted.^42,43 Researchers and practitioners should be cautious when considering the use of magnitude-based inferences in their work and when interpreting the results of others. In our review, the authors of 5 studies^{12,13,15,16,29} used magnitude-based inferences in their analyses. We cited their interpretation thresholds in quotation marks to signify that these are the interpretations of the original investigators. We also clearly designated this analysis approach in Supplementary Table 2. We believed it was important to include these papers because the studies were pertinent to this body of literature, albeit the original authors' interpretations may not be valid.

Other statistical variances that require attention in this body of research moving forward include appropriate sample-size calculations for the number of predictors that researchers intend to use in their models.⁴⁴ Using pseudo-R² statistics, Lolli et al,⁴⁵ in their 2019 paper on perceived exertion, session duration, and hamstrings injury, determined that sample sizes of 329, 583, or 1166 players were needed for 1, 5, or 10 load-related predictors, respectively. None of the studies in this review came close to those sample sizes.

In assessing injury risk, it is also important that both authors and readers understand and interpret the differences between injury rates and injury risk. From afar, the nuances of the 2 calculations can be easily overlooked, but the interpretation differences are critical, especially when comparing results. Additionally, when discussing injury risk, researchers should always report CIs to address the precision of the measurement and in the interpretation of the results.

The variations in statistical analysis among the papers reviewed were notable and reflect the larger conversation about statistics in sports medicine regarding a shift to complex approaches versus conventional, reductionist methods. It is becoming widely accepted that injuries occur as a result of complex and nonlinear interactions among multiple variables and that conventional approaches, even multivariable ones, are unlikely to capture the dynamic and complicated nature of injuries.^46,47 A reductionist approach assumes that the parts of the model can be broken down, examined individually, and then summed to represent the system as a whole.⁴⁷ Even multivariable approaches can be limited by the assumption that a system is equal to the sum of its parts.⁴⁸ Reductionist approaches should be used to inform the creation of complex modeling, yet their standalone utility in prediction (if this is the ultimate goal) is questionable. With all models, a validation study of unseen data should be completed before causation with injury is implied. Without validation, these models can only be considered to offer descriptive associations rather than true prediction. Implementing complex models may be difficult in a practical setting; however, the advantages of more accurate injury risk models should outweigh the implementation concerns. The authors of 2 studies^25,32 in this review attempted to use complex approaches when analyzing their data and may serve as exemplars as the field advances.

Assertions Made by Authors

From its inception, the novelty of GPS technology to monitor athlete workload was captivating to sports performance and sports medicine researchers. As work in this area began, many of the methods used to analyze these data and interpret results linked to injury prevention were published in editorials or commentary articles rather than in original research papers with quantitative analyses.^1,49,50 More recently, others⁵¹ have authored editorials expressing concerns regarding the use of these claims in research without further validation.

One highly debated metric that originated in research but was propagated by editorials was the ACWR.^31,52–54 This metric is prominent in the papers we reviewed, with many claims about its ability to predict injury. The relative simplicity of the ACWR has been complicated by the various ways investigators have used to calculate it (different numbers of days for acute and chronic periods, rolling averages versus EWMA, etc) and the broad array of metrics used to define player workload. Additionally, using chronic load as a modifier to the ACWR presents a potential mathematical error, as the ratio already uses the chronic load value to normalize the acute load.⁵⁵ Due to these variations, direct comparisons between articles and understanding what specific values of ACWR may be signaling present challenges. Others^56,57 have refuted the utility of ACWR in injury prediction, suggesting the metric may be signaling another confounding factor to injury risk, such as match congestion during a season.

Additionally, when interpreting the results of a univariate analysis, authors and readers should be cautious in attempting prediction, as variables may be associated with injury without meeting the threshold for direct causation. Difficulty understanding the nuances of association versus prediction may result in practitioners concluding that a factor associated with injury risk can be used to predict, and ultimately prevent, injury.⁵⁸ In turn, this may lead to incorrect inferences from spurious data. In the context of injuries, association can help identify individual factors in the overall puzzle of why injuries occur but only at a theoretical level.⁵⁹

Limitations of Existing Research and Suggestions for Future Research

We evaluated 22 original research articles that involved a total of 1136 athletes who sustained 2045 total injuries over 40 team-seasons. The vast majority of this research was based on analyses of relationships for a single team over 1 season. Troublingly, none of the investigations included female athletes. Furthermore, without standardization of injury definitions and GPS–derived workload metrics, replication and meta-analysis are impossible. A prediction model based on a single team also decreases generalizability to other populations.

A focus on hypothesis-driven research would help to overcome many of the concerns addressed in this review. Studies of GPS technology often occur when a team is collecting data for a performance-related reason, and the dataset is provided to a researcher post facto to assess the relationships of workload metrics and injury. By identifying the research questions, variables (especially injury-related variables), and the best statistical approaches a priori, data can be collected in a manner that lends itself to more complex and appropriate statistical modeling in order to identify true cause and effect.

As methods in this field become more robust and uniform in manner, using different scales to critically appraise the research will become possible. We elected to use the National Institute of Health Quality Assessment Tool for Observational Cohort and Cross-Sectional Studies; however, other methodologic quality-assessment instruments such as the Prediction Model Risk Of Bias Assessment Tool⁶⁰ and Quality in Prognosis Studies⁶¹ may be more specific to injury risk research.