Participants

Data were collected from football players (n=33) from one English Premier League club (age: 25.4±3.1 years, stature: 182.0±6.9 cm, body mass: 79.9±7.7 kg). All players trained on a full-time basis and played competitive fixtures within the Premier League during the 2014–2015, 2015–2016 and 2016–2017 seasons. Ten (30%) participants competed in all three seasons, 8 (24%) participants competed in two seasons and the remaining 15 competed in one season, resulting in 61 individual football seasons, analysed as independent data points. Goalkeepers were excluded from the study due to the different nature of their activity.

Quantifying workload

A global positioning system (GPS) was used to quantify workload data collected from all on-pitch training sessions and friendly matches. The GPS units (Viper 2, StatSports, Ireland), were placed between the scapulae of the players in bespoke vests. These units sampled at 10 Hz and the accelerometers at 100 Hz. Following each session, the data were downloaded into the specialised analysis software (Viper, 2.1.3.0). Competitive match data were recorded using a semiautomated camera system (video (VID)) (TRACAB; ChyronHego, New York, USA) provided, as standard by the English Premier League. The raw data files were then imported into the GPS software and analysed in an identical manner.

The validity and reliability of both GPS and VID for quantifying the physical demands of team sports has been demonstrated by numerous studies. The interchangeability of the two systems has also been established.13 In addition, as part of internal club research, the two systems used in this study were tested during a number of discrete tasks within a football-specific circuit, as well as during match play. Strong correlations (r=0.7–0.9, p<0.05) and small percentage differences (<10%) were found between GPS and VID for distances covered at all speeds except sprint distance, in line with previous work.13 Furthermore, both systems demonstrated good within player reliability (coefficient of variation (CV) <5%) and low percentage bias (<10%) when compared with the criterion during the circuit.

For sessions when data were unavailable for a participant (n=1149 of 10 221; 11%) as a result of them not wearing a unit, not having match data, not completing the entire session or the data being deemed unreliable due to intermittent satellite signal, estimations were made as follows:

Main training session data: estimated by calculating squad averages for drills completed (n=607 of 10 221; 6%).

International data: estimated by calculating the squad average of the other international players during the period of the international breaks14 (n=306 of 10 211; 3%).

Game data: matches were only monitored using VID from 2015/2016 onwards. Prior to this, match data were estimated using individual season game averages (from a minimum of three matches) from the data collected in 2015/2016 and 2016/2017. For players where this was not available (n=6), estimations were made based on friendly matches in which GPS was worn. Game averages were extrapolated according to individual game time, as per previous work12 (n=236 of 10 221; 2%).

The variables defined in table 1 were selected for use in this study due to their relevance to running loads (and potential injury). All variables were taken from the StatSports software (Viper).

Definition of injury

Injury information was classified by the club doctor and senior chartered physiotherapists. A recordable injury was defined as one that caused any absence from future football participation, that is, a time loss injury.15 Injuries were classified as being either: minimal (1–3 days of football activity missed), mild (4–7 days of football activity missed), moderate (1–4 weeks of football activity missed) or severe (4+ weeks of football activity missed).15 Injuries were also categorised by injury type and body site. The mechanism in which a participant acquired an injury was also classified as being non-contact or contact in nature.

Data analyses

Data were categorised in weekly blocks from Monday to Sunday. Every time a player participated in a training session or match, data were analysed in two ways. First, the previous 1-weekly, 2-weekly, 3-weekly and 4-weekly loads were calculated. The loads were then classified into discrete ranges from very low through to very high using z-scores16 (table 2). The relationships between these weekly cumulative loads and subsequent injury were investigated. Injuries that occurred within the next 7 days were included for analysis.17

Second, the acute workload for the current week was calculated as the 1-week load and chronic workload as the previous 4-week rolling average acute workload. The acute and chronic workloads were uncoupled to prevent them being falsely correlated. The ACWR was calculated by dividing the acute workload by the chronic workload.18 Only acute workloads that were preceded by four complete weeks were included in the ratio calculations. A value of greater than 1 represents an acute workload greater than the chronic workload. Chronic workloads were also separated into high and low categories by the median score for each variable.19 From this, workload–injury relationships between ACWR ratios combined with high and low chronic workloads were analysed. As with accumulated workloads, the ratios were categorised based on z-scores (table 3). Only conditions that contained 20 or more injuries were included in the statistical analysis to allow for moderate to strong associations to be made.20 Consequently, data were excluded for incidences when the chronic workloads were high for both non-contact and contact injuries. This was also the case for contact injuries when the chronic loads were low.

Statistical analyses

The analysis was performed in a similar manner to the previous work of Hulin et al.18 Injury incidence was determined by dividing total number of injuries by the ‘on-legs’ exposure time and reported as rates per 1000 hours. Injury risks were calculated as the number of injuries sustained relative to the number of exposures to each workload classification.18 Exposure data were recorded as per the procedures outlined by the Fédération de Football Association Medical Assessment Research Centre.15 A binary logistic regression model was used to compare workloads between injured and non-injured players for all workload variables independently. Accumulated workload and ACWR were independently modelled as predictor variables. RR was then calculated using to determine the magnitude of the injury risk above and below given workloads or ratios (MedCalc Software, Ostend, Belgium). When an RR was greater than 1.00, an increased risk of injury was reported (ie, RR=1.50 is indicative of a 50% increased risk) and vice versa. For an RR to be significant, 95% CIs did not contain the null RR of 1.00. Data were analysed using IBM SPSS Statistics V.25.0 and reported as means and 95% CI. Significance was accepted at p<0.05.

Injury incidence

For the duration of the study, 132 injuries (13.3/1000 hours ‘on-legs’ exposure time) were recorded (2014–2015 season, 17.6/1000 hours; 2015–2016 season, 10.2/1000 hours; 2016–2017 season, 12.4/1000 hours), including contact and non-contact injuries (see online supplementary appendix A). The knee was the most common site of injury across the three seasons (2.9/1000 hours), 69% of which were non-contact injuries (2.0/1000 hours), predominantly meniscal or cartilage lesions and ligament sprains (0.9 and 0.7/1000 hours, respectively). The ankle was the most common site of contact injury (1.9/1000 hours), with the most common type being ligament sprains (1.6/1000 hours). The injury incidence in competition was over five times that of training (33.7/1000 hours vs 5.8/1000 hours). In particular, contact injuries were considerably greater in competition than in training (16.9/1000 hours vs 1.3/1000 hours). Despite a lower exposure to competition, 80% of contact injuries occurred in matches. The total number of days missed through injury was 4820 (36.5±62.7 [mean±SD] days per injury).

Overall injuries

A low chronic workload of ACC (<1881), decelerations (DEC; <1731) and low-intensity distance (LID; <19 222 m) combined with a very high ACWR (>2.0) elicited the greatest overall injury risk (RR=3.2, 95% CI 1.3 to 7.6, p=0.01; RR=3.5, 95% CI 1.5 to 8.2, p=0.01; and RR=2.76, 95% CI 1.2 to6.6, p=0.02, respectively). The risk was also significant for very high ACWR of the same metrics, plus total distance (TD), combined with all chronic loads (RR=2.4–2.6) (table 5). Conversely, a low ACWR of TD (0.4–0.7) for all chronic loads was associated with a decreased injury risk (RR=0.2, 95% CI 0.1 to 0.8, p=0.02).

Non-contact injuries

Low chronic workloads combined with very high ACWRs for TD (>2.14), LID (>2.15), ACC (>2.30) and DEC (>2.32) resulted in a non-contact injury risk 5–7 times greater than ACWRs below this (RR=4.5 (TD), –6.6 (DEC), p<0.05). Additionally, a low amount of TD accumulated over 4 weeks (45 844–67 519 m) also resulted in an increased risk (RR=2.2, 95% CI 1.0 to 4.6, p=0.04) (table 4). Significant risks were also found for TD, ACC, DEC and LID for all chronic loads when the ACWR was very high (RR=3.7–3.9) (table 5).

Contact injuries

Moderate to high ACWR (1.1–1.5) for TD, DEC and LID produced the largest contact injury risk (RR=2.0, 95% CI 1.0 to 4.0, p=0.04; RR=2.0, 95% CI 1.0 to 4.0, p=0.04; and RR=2.6, 95% CI 1.3 to 5.2, p=0.01, respectively). A moderate to high amount of TD (24 042–30 549 m) and a low to moderate amount of DEC (1288–1782) accumulated over a week also showed a heightened risk of contact injury (RR=2.1, 95% CI 1.1 to 4.0, p=0.03 and RR=2.0, 95% CI 1.1 to 3.9, p=0.03, respectively).

Non-contact injuries and workload

A very high ACWR combined with low chronic workload categories demonstrated the greatest non-contact injury risk for most metrics (except high-speed distance [HSD] and sprint distance [SD]), with DEC being most strongly associated with RR (ACWR >2.3, RR=6.6). When all chronic workloads were analysed, a very high ACWR demonstrated a lesser but still significant risk for the same metrics (RR=3.7–3.9). These findings are in line with studies in cricket,18 rugby,19 22 Australian football,23 Gaelic football24 and football5 where high ACWR, referred to as ‘spikes’ in workload, have been associated with heightened injury risk. Due to the inevitable increased exposure to risk with greater workloads, previous research has focused on the higher load–higher injury risk relationship.7 25 However, the above findings, alongside the lack of significant risks associated with high accumulated loads in this study, support a growing body of literature suggesting that acute, excessive, rapid increases in loads may be responsible for a large proportion of non-contact injuries, rather than chronic exposure to higher loads.10

The protective effect of high chronic loads versus low chronic loads has been reported in rugby.19 22 They concluded that the players who were capable of achieving high exposure had the enhanced physical attributes needed for decreased injury risk. This theory has recently been demonstrated in hurling, where players with well-developed lower body strength, repeated sprint ability and speed tolerated higher training loads and had a reduced risk of injury compared with lower performance groups.26 In youth football, a high ACWR combined with low chronic HSD load showed a significantly increased risk of non-contact injury (ACWR=1.4–2.0, RR=2.6), which was not evident when combined with high chronic HSD (ACWR=1.3–1.8, RR=0.5).5

In the current study, there were not enough injuries when the chronic workloads were high to determine the RRs of non-contact injury, further indicating a potential protective effect. Ultimately, training at higher workloads may cause players to develop a greater tolerance for the increasing intensity and fatigue of competition.10 Concurrently, reducing training loads, while lowering a player’s exposure to risk, may also have a negative effect on fitness and physical preparedness, potentially increasing the risk. Therefore, as per the training–injury prevention paradox model,10 optimal load management to minimise injury risk should involve appropriate, progressive exposure to higher loads while avoiding training load spikes that the player is not prepared for.

Contact injuries and workload

Very little research has investigated the relationship between contact injuries and workload, despite early workload–injury research suggesting that players with better developed physical capacity may be at less risk of contact injury.27 In the current study, 80% of all contact injuries occurred in matches, similar to previous injury incidence reports.28 This may be due to the high speed and intensity of play, resulting in more body contact such as sliding and tackling.29 The risk of contact injury was greatest when the ACWR was moderate to high for TD, LID and DEC (RR range=2.0–2.6), meaning the acute workload was very similar to the chronic workload.

Also, as load was categorised by z-scores, contact risk was highest for the most commonly occurring ratios (z=0.0–1.0). Players who are regularly in the team, and therefore more at risk of contact injury, typically have a lack of variation in their workload due to a large proportion of the weekly load being attained from matches. Therefore, when the match load was constant, variations in the training load produced very little fluctuation in the total acute load.23 Thus, it would appear that in the current study, contact injury is most likely to be related to match exposure, rather than the prescribed workload. The lack of association of the ACWR to contact injuries is highlighted further by the large RR of non-contact injury following an acute spike compared with overall injury, suggesting that including contact injuries reduces the association of the ACWR with injury risk. Consequently, these injuries should be analysed separately when establishing workload-injury relationships and determining uniform injury definitions across research.30

Potential limitations

Previous studies have highlighted the limitations of using estimated match data, as it does not account for match to match variability.14 This study has attempted to improve on this by including match data for the 2015/2016 and 2016/2017 seasons. However, for the 2014/2015 season, TRACAB data were not available, resulting in estimations being calculated as per previous work,12 emulating the aforementioned limitation. Additionally, as the match data for the latter two seasons were collected using a different system than training, the precision and sensitivity of the data may be decreased, despite it being calibrated to maximise between system agreements.13 With technological advancements, and the recent admittance of GPS in league matches, future research should aim to use a single monitoring system for both competition and training.

One potential explanation for the lack of significant non-contact injury risk for very high acute HSD and SD, despite all other metrics reporting otherwise, may be the use of absolute speed thresholds in this study. Buchheit14 recently stated that the use of fixed thresholds may reduce the sensitivity of the ACWR due to the varying locomotor profiles between players, particularly as subtle differences in speed at high intensity have been found to have important implications on injury risk. Future research could consider the use of individualised thresholds, although caution must be taken when anchoring all locomotor categories to one fitness measure.31

The present study only examined external load; however, the incorporation of the rating of perceived exertion (RPE) values, as a measure of internal workload, may provide a more complete insight into the likelihood of injury, as well as taking into consideration the athlete’s response to a given workload.17 Fanchini et al 21 recently analysed the ACWR in relation to injury risk in elite football using RPE as their load measure. Similar to the current study and the previous work of the authors,5 they demonstrated a heightened risk with acute ‘spikes’. A combination of both methods may give a more complete assessment of load-related risk, while also considering the validity and specificity of the chosen metrics to the sport and the individual.32

Calculating the ACWR using rolling averages is evidence based and supported by a large body of literature. However, future research may consider using exponentially weighted moving averages, which consider the decaying nature of fitness and fatigue over time. This method has recently been shown to have a greater sensitivity to increases in injury risk at higher ACWRs.33

The statistical power of this study was not calculated prospectively. As retrospective power analysis calculations are not appropriate,34 the power analysis was not included. However, this study included 81 injury cases, which is enough to make moderate to strong associations regarding injury risk factors.20 Future studies should ensure prospective power analysis for inclusion. Furthermore, as commonly recommended in elite sport research, future work involving multiple clubs would enhance the ability to generalise these findings, advance the statistical analysis and detect small to moderate associations (+200 injury cases).