S18 Abstracts / Journal of Science and Medicine in Sport 21S (2018) S5–S76 Methods: Workload data were pooled from Australian football (n = 2550) and soccer (n = 23,742) cohorts to create a representa- tive sample of ACWR observations for team sports. One-hundred datasets were simulated, with sample sizes of 5000 and 1000 obser- vations. Injuries were simulated in the data using two pre-specified risk profiles; U-shaped and flat. The U-shaped profile aligns with the currently hypothesised relationship and a flat injury profile represents the scenario that injury risk is independent of ACWR. Simulating multiple data sets with known injury risk allows us to estimate the false positive and negative rates for each modelling approach and compare their ability to fit the true relationship in the data. Three commonly used discrete modelling strategies (z-score categories, percentile splits and arbitrary cut-points) were com- pared to two continuous approaches that avoided discretisation (restricted cubic splines and fractional polynomials). Results: Discrete models were inferior to continuous methods for fitting the true injury risk profiles in the data and fit unrealistic step-shaped profiles. Discrete methods had higher false discovery (16–21% vs. 3–7%) and false rejection rates (5–59% vs. 12–19%) than continuous methods. Evaluating models using the area under the receiver operator characteristic (ROC) curve incorrectly identified discrete models as superior in over 30% of simulations. Probabilistic scoring was better suited to assessing model performance with less than 6% incorrect model selection rate. Discussion: Modelling methods that discretise continuous risk factors are inappropriate for studying the relationship between training loads and injuries. Discrete models have inflated false discovery and false rejection rates and are unsuited to fitting non- linear risk profiles. Strong justification is required for research that chooses a discrete approach and we suggest avoiding discreti- zation and modelling relationships with continuous methods such as spline regression or fractional polynomials. Evaluating injury risk models using ROC curves may not reflect their practical use and may lead to inferior model selection. Probabilistic scoring methods and calibration curves may be more informative when assessing injury models. https://doi.org/10.1016/j.jsams.2018.09.041 O27 Predictive modelling of non-contact lower limb injuries in elite Australian footballers J. Ruddy ∗ , N. Maniar, S. Cormack, R. Timmins, D. Opar Australian Catholic University, Moorabbin, Australia Introduction: Previous research has demonstrated a limited ability to predict injuries using workload data in elite Australian footballers. Aside from global position systems (GPS) data, a wide array of data are routinely collected in Australian football, includ- ing lower limb strength data, weekly screening data and wellness data. However, no research has considered how these additional variables may mediate the impact of external workload on injury risk. Accordingly, the aim of the current study was to assess the ability of models built using GPS data and a variety of other rou- tinely collected variables to predict non-contact lower limb injuries in elite Australian footballers. Methods: Data were collected from one elite Australian football team (45 athletes) during the 2017 Australian Football League sea- son. Demographics and playing/injury history data were collected prior to pre-season. Measures of lower limb strength (isometric mid-thigh pull [IMTP], countermovement jump, squat jump, and Nordic hamstring force), as well as biceps femoris long head (BF LH ) fascicle lengths, were collected four times throughout the sea- son. Screening data (knee-to-wall, sit and reach, and isometric groin/hamstring squeezes) were collected weekly. Wellness ques- tionnaires, GPS data and prospective injury data were collected every session/match that athletes completed. Area under the curve (AUC) was calculated to measure the level of association between each individual variable and the outcome (injury/no injury). Via various machine learning techniques, predictive models were built using all data. AUC was used to assess the predictive performance of each model by comparing the predicted outcomes to the actual outcomes. Results: The minimum, maximum and mean AUC values for the predictive models built using all data were 0.71, 0.87 and 0.79 respectively. The five variables with the highest AUC, excluding the workload variables, were days missed in the prior season due to injury (0.68), mass (0.63), stature (0.60), BF LH fascicle lengths (0.60) and IMTP relative force (0.60). Discussion: Previous research has observed an average AUC of 0.65 when attempting to predict non-contact injuries using work- load data. The results of the current study suggest that the inclusion of additional data (i.e. injury history, anthropometrics and lower limb strength) may improve the ability to identify injury risk. How- ever, it is still unknown how the additional variables included this study may mediate the impact of workloads on the risk of injury. Despite this, examining additional variables in concert with GPS data may improve the ability to identify risk and predict injuries at an individual level. https://doi.org/10.1016/j.jsams.2018.09.042 O28 Monitoring players thigh injury risk in response to game loading during an elite U20s basketball camp T. McGann 1,∗ , M. Drew 2 , T. Pizzari 3 , K. Dooley 1 , S. Snodgrass 1 , E. Rio 3 , A. Schultz 4 , L. Donnan 5 , S. Edwards 4 1 University of Newcastle, Callaghan, Australia 2 Australian Institute of Sport, Bruce, Australia 3 La Trobe University, Melbourne, Australia 4 University of Newcastle, Ourimbah, Australia 5 Charles Sturt University, Albury-Wodonga, Australia Introduction: Match induced fatigue effects have been asso- ciated with reduced hamstring and groin strength, and may be related to a concomitant increase in risk of injury. The correla- tion in high performing basketball athletes however, is presently unclear. Therefore, the purpose of this study is to investigate fatigue effects following basketball match-play on hamstring and adductor strength in elite U20’s basketball players. Methods: Thirty-five elite U20 male basketball athletes (age 16.8 ± 1.1 yrs) attending a national scouting camp at the Australian Institute of Sport were recruited to participate in this study. Par- ticipants underwent strength testing via handheld dynamometry pre- and post-game for two officiated games played two days apart. Three maximal voluntary isometric contraction trials were con- ducted unilaterally using the following protocols: Hamstring: prone position on a 45 ◦ wedge placed on an electronic examination bench, using a calibrated load cell attached to the bottom of the bench. Adductors: Copenhagen five-second squeeze test in a supine with a hand-held dynamometer. Grip strength: elbow at 90 ◦ of flexion in neutral supination with a hand grip dynamometer. Fatigue effects were assessed with a mixed-effects restricted maximum likelihood regression in Stata 15 (p < 0.05).