Development of an algorithm-based approach using neuromuscular test results to indicate an increased risk for non-contact lower limb injuries in elite football players

Participants

We carried out a retrospective cohort study following the ethical principles for medical research involving human subjects. We used data from 77 professional players (26.16±4.63 years) of a team competing in the highest Belgian league from 2020–2021 season to November 2022, containing both injured and healthy cases. All players volunteered to participate and gave written consent.

We included players based on the following eligibility criteria: (1) the injury must be located in the lower limb or the lower back; (2) the injury must be due to a non-contact mechanism; (3) the player had to be unfit to play/train for at least 1 day, and he should have received medical attention. Healthy players were included based on the following criterium: no present injury; no data can be used from 2 months prior and after a previous injury occurred—to prevent the influence of recent injuries.

Data acquisition

We extracted results of various neuromuscular assessments from Smartabase (product of Teamworks, Durham, North Carolina, USA), a data-driven platform used by the team to store their gathered data. The baseline was measured for each player at the start of the season. When an injury occurred, four measurements in time were noted: T4, test result 4 weeks before the injury; T3, 3 weeks before the injury; T2, 2 weeks before the injury; T1, 1 week before the injury. This sequence of the injury event as starting point and the four preceding weeks were defined as a ‘case’. We used the same principle with healthy cases; one value of the included neuromuscular tests for each week of the 4 weeks before a specified time.

The following neuromuscular tests were performed for each player weekly: eccentric hamstring strength test, isometric adduction/abduction strength test and countermovement jump test. The neuromuscular tests were evaluated with the following testing systems: NordBord (NB), ForceFrame (FF) and ForceDecks (FD) (VALD Performance, Newstead, Australia). Table 1 depicts the measured outcome variables. For each measured variable, an average of three measurements were automatically calculated by the Smartabase software. Additionally, we calculated between limbs and the relative value of each variable concerning the baseline.

Test protocols

Data management

After collecting all data, we calculated the missing data percentages for the different neuromuscular testing variables. Hereafter, we imputed the missing values using three different methods.

1. Mean method: the average of the available data of that period per test was taken, and this outcome was entered into the missing values.

2. Forward–backward method: depending on which week was missing, the data from 1 week before or 1 week after the given date was duplicated. When T1 and T3 were given, the backward method was used for both missing values. So, T2 is completed by taking T1 and T4 by using the values from T3. When T2 and T3 are known, the forward method is used for T1 and the backward method is used for T4. When periods T1 and T4 were given, T2 was completed using the backward method and T3 was completed using the forward method. When only one period was not given, we used the backward method for periods T2 and T4 and the forward method for T1 and T3.

3. Linear regression method: when data are missing between 2 weeks, the average of the two surrounding periods is taken, and this average is calculated. This value is used for the remaining 3 weeks when only 1 week’s data are available. If only T4 is missing, a backward method from T3 is performed. If T1 is missing, a forward method is performed from T2. If T1 and T2 are missing, a forward method from T3 is performed and if T3 and T4 are missing, a backward method from T2 is performed. An exception can be found when missing T2 and T3, and then a linear evolution is determined using these formulas:

   • T3=T4+(T1–T4)/3

   • T2=T1–(T1–T4)/3

Each method resulted in a complete dataset; therefore, we have three separate datasets used for further analysis.

Data analysis

We used the data mining technique Subgroup Discovery, a descriptive, exploratory technique that can handle relatively small datasets, as considered here. With this technique, we can extract easy-interpretable findings that have proven useful in sport-specific settings.13–17

This technique aims to find subsets with a different distribution of the outcome variable than the distribution in the entire data collection. In this case, the outcome variable is binary and equals 1 or 0, corresponding to an injured or healthy person, respectively. The obtained subgroups are subsets of the entire dataset that contain the instances that satisfy single or multiple conditions on the independent variables or so-called features, for example, FF abduction strength imbalance or FD peak landing force at T4. The results are considered significant if there is less than a 5% probability that the result is a consequence of considering too many features.18

We used the subgroup discovery tool Cortana for our analysis (https://datamining.liacs.nl/cortana.html). We present all significant subgroups described by a restriction on a single feature. Moreover, we also performed a more in-depth analysis by allowing a combination of restrictions on two features.

Subgroup discovery with one restriction

Table 3 depicts the statistically significant results for the three datasets. Datasets are distinguished based on the three missing data-filling techniques. The largest proportion of missing data was found in the FD assessments (34.67%), while 18.06% of data was missing in FF variables and 17.90% in NB variables, respectively. The dataset of missing data filled by the mean method contains six statistically significant results, while the two other datasets have three statistically significant findings. Only restrictions on two variables have a statistically significant injury risk score across all three datasets: increased abduction strength imbalance between limbs 3 weeks before injury compared with baseline (threshold: 0.97); decreased right leg adduction strength 1 week before injury compared with baseline (threshold: 1.01). There was one feature with a statistically significant injury risk score across two datasets—mean method and linear regression: Decreased left leg adduction strength 4 weeks before injury compared with baseline (threshold: 1.08).

Subgroup discovery with two restrictions

Additionally, to a single restriction, we also investigated subgroups described by restrictions on two features. If we compare the results of the three datasets, we found that the subgroup with the largest deviation from the distribution of the injury score in the entire data collection is the combination of increased abduction imbalance with respect to the baseline 3 weeks before injury (threshold: 0.97) and decreased peak landing force in the left leg 4 weeks before injury (threshold: 1.24). The subgroup contained, on average, 148 cases over the three datasets. When these conditions were met, we found roughly 50% injured players compared with only 33% in the data collection.

Results and clinical implications

Hitherto, published research regarding ML to predict injuries in professional football is scarce.19 Hecksteden et al ascertained that screenings and data collection yields promising results for predicting non-contact lower limb injuries.20 Our findings align with this study—we prove that ML can be used for the above-mentioned purposes. Bahr criticised preseason screening as unfit for injury prediction due to lack of substantiation.21 Nevertheless, screening can still highlight key risk factors for sports injuries.22 The current literature comprises similar proofs of concept in the NBA11 12 or applied to a specific injury—ACL injury.23 The study of Cohan et al predicted injury based on the injury mechanism, player’s characteristics and game statistics,11 while Lu et al focused mostly on injury history and past concussions.12 Jauhiainen et al included an extensive screening protocol comprising neuromuscular and functional tests. Their results show that they could not predict ACL injuries in clinical practice.23

We found that increased abduction imbalance for the baseline 3 weeks before the injury and decreased right adduction for the baseline 1 week before injury were significant predictors of risk of injury. The analysis with two restrictions showed significant predictive value for the combination of increased abduction imbalance for the baseline 3 weeks before the injury and decreased peak landing force in the left leg 4 weeks before the injury for risk of injury. Concretely, if the test results of a player would highlight one of the thresholds mentioned above, a red flag should be raised that the player is more prone to injury. This would enable the technical and medical staff to alter training and rehabilitation programmes following the respective red flag.

The increase between-leg abduction imbalance and decreased adduction muscle strength follow the proportion of groin injuries in the current sample. Lower adductor strength shows to yield a higher risk of a groin injury.24 Considering the number of hamstring injuries in the current sample, one would expect a significant finding regarding the eccentric hamstring strength test. In the recent literature, the Nordic hamstring test has come out as one of the best predictors of a hamstring strain injury.25 Another article executed by various Premier League teams found that more than 15% asymmetry is a significant predictor for a hamstring injury.26 Both articles found that decreased eccentric hamstring strength is a predictor for an injury as opposed to our results.

Subgroup discovery

Our data mining approach obtained results that can be applied clinically. More specifically, this analysis provides specific thresholds that can be used as a red flag for increased risk of injury. The medical and technical staff can include these values in their daily programmes. Second, using a subgroup discovery algorithm, we obtained that a combination of features is even more relevant. This follows previous studies that demonstrated that injuries result from complex interactions of features that interrelate.7 Determining combinations of features, and recognising patterns, can help with injury prevention.

Limitations

Our study is not without limitations. First, we had a considerable amount of missing data and recommendations regarding which imputation method should be used ambiguously. We chose three single imputation methods within our study: mean value substitution, last observation carried forward/backwards and linear regression method.27 After our research, neither one of those methods proved to be superior or inferior. The main statistical finding remained the same over the three datasets. It has to be noted that single imputation methods could implement biased estimates even if data are missing completely at random. In case of fewer missing data, we recommend using multiple imputation procedures. These will create more reliable results and a better understanding of the impact of missing data.27

Second, the included non-contact injuries represented various injuries—from muscle strains to fractures. Even though groin injuries and hamstring injuries account for almost half of the included injuries, such equivocality can underpin confounding factors. Furthermore, it would be interesting to distinguish injuries based on anatomical area and perform the subgroup discovery analysis for each type of injury separately. Unfortunately, our dataset was too small for this kind of analysis.

Third, we included only a selection of potential predictors in agreement with the medical and technical staff of the club. More salient outcome variables, such as rate of force development, reactive strength index, power and stiffness, were not included. Considering the aim of this study—to develop a proof of concept—this is only a minor limitation. However, further development of this model should include evidence-based outcome variables.

Lastly, these results should be considered cautiously: our analysis was performed on a small dataset, our ascertained thresholds are based on average data over all participants instead of individually, we included only a small number of variables and we limited our time frame to 4 weeks before the injury; as such, the current results are not clinically applicable. Nevertheless, this dataset served our purposes, since we could determine that our model can identify red flags that indicate an increased risk of injury. We need a more robust dataset with more variables for further development to a clinically applicable model.