PaperlayerView exact revision on arXiv

Source-linked AI summary

Effective injury forecasting in soccer with GPS training data and machine learning

Alessio Rossi, Luca Pappalardo, Paolo Cintia, Marcello Iaia, Javier Fernandez, Daniel Medina

arXiv:1705.08079v2stat.MLstat.AP

TL;DR

Injuries affect professional soccer, while evidence about injury-risk factors is preliminary and statistical injury forecasting remains poorly evaluated. The paper develops a GPS-based training-workload forecaster and reports that it detects around 80% of injuries with about 50% precision.

  • Problem

    Injuries affect the sports industry, while studies offer only preliminary understanding of injury-risk variables and statistical models’ potential for injury forecasting remains poorly evaluated.

    Source: Introduction · p. 3

  • Method

    The paper develops a multidimensional, data-driven injury forecaster using GPS training-workload data collected from professional soccer players during a season.

    Source: Introduction · p. 4, Data collection and feature extraction · p. 6

  • Results

    Around 80% of injuries were detected by DT, with about 50% precision.

    Source: Discussion · p. 17

  • Takeaways & Limitations

    The approach offers interpretable, practical rules for evaluating and interpreting relations between injury risk and training performance in professional soccer.

    Source: Introduction · p. 4, Conclusion · p. 19

    Reliable forecasting requires an initial period of data collection to gather adequate data before a model can be trained.

    Source: Discussion · p. 17

Abstract

from arXiv · show

Injuries have a great impact on professional soccer, due to their large influence on team performance and the considerable costs of rehabilitation for players. Existing studies in the literature provide just a preliminary understanding of which factors mostly affect injury risk, while an evaluation of the potential of statistical models in forecasting injuries is still missing. In this paper, we propose a multi-dimensional approach to injury forecasting in professional soccer that is based on GPS measurements and machine learning. By using GPS tracking technology, we collect data describing the training workload of players in a professional soccer club during a season. We then construct an injury forecaster and show that it is both accurate and interpretable by providing a set of case studies of interest to soccer practitioners. Our approach opens a novel perspective on injury prevention, providing a set of simple and practical rules for evaluating and interpreting the complex relations between injury risk and training performance in professional soccer.

1 Department of Computer Science, University of Pisa, Pisa, Italy … Introduction

The paper proposes a data-driven, multidimensional and interpretable approach to forecasting soccer injuries from recent training workload, addressing limitations in existing forecasting and aiming to support club staff decisions.

Introduction

  • Injuries impose substantial costs and absences: in Spain they account for about 16% of professional soccer players’ season absence and around 188 million euros per season.The introduction also notes effects on athletes’ mental state and team performance, and considerable recovery and rehabilitation costs.Source: Introduction · p. 3
  • Existing soccer injury research offers preliminary insight into risk factors, but statistical forecasting remains poorly evaluated and prior approaches use one variable at a time.The paper argues that such approaches do not fully exploit complex patterns in available data.Source: Introduction · p. 3
  • The forecaster predicts injury in the next training session or official game from recent workload, using a multidimensional, interpretable, data-driven approach based on GPS data.The paper presents it as a potential decision-support tool for soccer club staff.Source: Introduction · p. 4
  • 50% precision, versus less than 5% for existing mono-dimensional approaches, was reported for the proposed injury forecaster.The paper says this result was validated by simulating use of the forecaster over a season as new training data became available.Source: Introduction · p. 4

Related work

  • Rugby studies report high injury risk when workloads exceed certain thresholds, while cricket players with acute/chronic ratios above 2 face 2–4 times higher injury risk than the other group.Source: Related work · p. 4
  • When skating session load outweighs recovery capacity before the next session, overtraining syndrome can occur and cause injury; in basketball, injury risk strongly correlates with 7-day session-load monotony.Monotony is the ratio of mean session load to its standard deviation, recorded over the past 7 days.Source: Related work · p. 5
  • Studies of young soccer players associated jump height, body size and prior injuries with thigh-strain probability, while an NBA classifier predicted 19% of injuries.The NBA study identified average speed, past competitions, distance covered, minutes played and average field goals attempted as important features.Source: Related work · p. 5
  • A prior soccer forecasting attempt using a reduced feature set achieved best-case accuracy not significantly better than.The supplied passage does not state the accuracy value or comparison baseline.Source: Related work · p. 5

Material and Method

This section describes a decision-tree injury forecaster built from season-long GPS workload and player-history data, using player-session examples. It outlines the stratified cross-validation, 10,000 repetitions, and comparisons with workload-based forecasters, baselines, and alternative classifiers.

Data collection and feature extraction

  • Twenty-six Italian professional male players were monitored for 23 weeks in 2013/2014, yielding 931 individual training sessions recorded with 10 Hz GPS devices.The club’s medical staff recorded 23 non-contact injuries during the study.Source: Data collection and feature extraction · p. 6
  • Nineteen of the 23 injuries involved players with a previous injury, and each session example was labeled for injury in the next game or training session.The forecaster also incorporated player characteristics, recent workload, and previous-injury information.Source: Data collection and feature extraction · pp. 6, 8

Phase 1: Training dataset construction

  • The training dataset comprised 55 features across 952 individual training sessions: 18 daily, 12 EWMA, 12 ACWR, 12 MSWR, and one previous-injury feature.Source: Phase 1: Training dataset construction · pp. 8–9
  • RFECV selected features using 30% of the dataset, and ADASYN oversampled the seven injury examples in that training subset before decision-tree fitting.Feature selection aimed to reduce dimensionality and overfitting while improving interpretability.Source: Phase 1: Training dataset construction · p. 9

Phase 2: Model construction and validation

Results

The decision-tree forecaster achieved high injury recall with fewer false alarms than the baselines and ACWR/MSWR forecasters; in a week-by-week scenario, its performance improved as seasonal data accumulated.

Results

  • Across 10,000 tests, forecasting quality varied with the injury types in each training and test split; RF had higher recall but lower precision than DT.RF recorded recall=0.87±0.05 and precision=0.41±0.08; LR performed much worse than DT.Source: Results · p. 12, Figure 2 · p. 13
  • On the injury class, DT achieved recall=0.80±0.07 and precision=0.50±0.11, significantly improving on baselines B1–B4, whose maximum precision was about 6%, and ACWR/MSWR-based injury forecasters.Typically, DT drastically reduces false alarms.Source: Results · p. 12

week, how many injuries the club can actually prevent throughout the season?

  • 9 of 14 injuries were detected from w6 onward (F1-score=0.60; precision=0.56), and DT outperformed all other models from w14 as seasonal data accumulated.DT performed poorly in the first weeks but predicted most injuries in the season’s second half.Source: Figure 3 · p. 14, week, how many injuries the club can… · p. 14

Interpretation of the injury forecaster

  • DT rules offer coaches and trainers a practical handbook; RFECV selected PI(EWMA), dHSR(EWMA), and dTOT(MSWR) from 55 features, with Gini importances 0.71, 0.23, and 0.06.The rules can help coaches modify training schedules and improve players’ fitness.Source: Interpretation of the injury forecaster · p. 15
  • After prior injury, HSR(EWMA)<112.35 describes 42% of injuries with 100% correctness; with HSR(EWMA)>112.35 and Dtot(MSWR)<1.78, rule 2 describes 30% at 100% accuracy.Source: Interpretation of the injury forecaster · pp. 15–16
  • After prior injury, HSR(EWMA)>112.35 with Dtot(MSWR) two and a half times higher than the player’s average describes 28% of cases, at mean accuracy 75±5%.The authors advise monitoring total distance and high-speed running for players recently returned after injury.Source: Interpretation of the injury forecaster · p. 16

Discussion

The decision-tree forecaster detects around 80% of injuries with about 50% precision, and its evolving use during a season can support injury prevention and reduce injury-related costs.

Discussion

  • The decision tree detects around 80% of injuries with about 50% precision, has a small false-positive rate and moderate false-negative rate, and outperforms baselines and state-of-the-art approaches.The authors note that false alarms can affect team performance when player availability is limited.Source: Discussion · p. 17
  • Cumulative injury-class F1-score reaches 0.60 as the season progresses, outperforming RF and LR and enabling prevention of more than half of injuries.The discussion frames this as a real-world setting in which a club begins collecting data and retrains the forecaster during the season.Source: Discussion · p. 17
  • Because scarce initial data affect forecasting, the authors recommend an initial collection period; in their dataset, classifier performance stabilizes after 14 weeks.The required period depends on club needs and strategy, including training and game frequency, injury frequency, player availability, and tolerated false alarms.Source: Discussion · p. 17
  • Feature selection changes over the season, then stabilizes after 14 weeks at three of 55 features: PI(EWMA), dHSR(EWMA), and dTOT(MSWR).PI(EWMA) is the most important and only always-selected feature; changing selections may reflect scarce early data and the injuries that have occurred.Source: Discussion · pp. 17–18
  • Among detected injuries, fewer than half occurred immediately after return to regular training, while 60% occurred long after a prior injury and involved specific dHSR(EWMA) and dTOT(MSWR) values.The authors advise particular care during players’ first sessions back and later monitoring of kinematic workloads.Source: Discussion · p. 18
  • PI(EWMA) alone has little predictive power, whereas combining it with dHSR(EWMA) and dTOT(MSWR) enables injury prediction.The PI-based baseline B4 has much lower accuracy than the decision tree.Source: Discussion · p. 18
  • The club recorded 139 injury-related absence days, estimated to cost at least 11,583 euros, or 3.81% of salary costs; using the model could have prevented 9 of 14 injuries and saved 8,881 euros, a 77% cost decrease.The savings estimate assumes the club stops players predicted to be at risk as the season progresses.Source: Discussion · p. 18

Conclusion

  • The proposed GPS-and-machine-learning approach balances accuracy and interpretability, reduces false alarms relative to state-of-the-art approaches, and provides rules for understanding injuries.The conclusion identifies game-performance features, transferability across clubs, and combining GPS with health data as future extensions.Source: Conclusion · p. 19

SUPPORTING INFORMATION · S1 Appendix. Descriptive statistics of the workload features.

The appendix summarizes 12 training workload features and reports that their distributions are non-normal, tending to be bimodal and right-skewed.

S1 Appendix. Descriptive statistics of the workload features.

S2 Appendix. The ACWR method

The appendix reports that sessions with ACWR below 1 had the highest observed injury risk, while ACWR predictors were generally impractical because of poor precision.

The ACWR method

  • The study computed ACWR for each of 12 workload features using 7-day acute and 28-day chronic exponentially weighted moving averages.ACWR is the ratio between acute and chronic workload.Source: S2 Appendix. The ACWR method · p. 22
  • In contrast with the literature, no individual training session resulted in ACWR > 2, while players whose individual training sessions resulted in ACWR < 1 had the highest injury risk.Source: S2 Appendix. The ACWR method · p. 22
  • Each feature-based classifier forecasts injury when that feature's ACWR is below 1; combined variants require most, all, or at least one feature below 1.Only C(vote) achieved slightly better injury-class precision than the single-feature ACWR forecasters.Source: S2 Appendix. The ACWR method · p. 23
  • 0.80 ± 0.08 recall but 0.03 ± 0.003 precision (97% false alarms) made ACWR forecasters impractical, despite significantly exceeding baselines in injury-class recall.The reported recall and precision are averages across the feature-based forecasters.Source: S2 Appendix. The ACWR method · p. 23
  • Among the combined variants, only C(vote) slightly improved injury-class precision over the single-feature ACWR forecasters.C(vote) predicts injury when a majority of workload features have ACWR below 1.Source: S2 Appendix. The ACWR method · p. 23
  • Using ACWR quintiles instead of predefined groups confirmed that low-ACWR groups had the highest injury risk, while the resulting predictors still had too-low precision for practical use.Classification results for the quantile-based predictors were similar to those for predefined groups.Source: S2 Appendix. The ACWR method · p. 24

S3 Appendix. The MSWR method … S9 Appendix. Predictions results

The appendices describe workload-risk methods, feature construction, injury-history weighting, sampling, and evaluation, and report prediction results under different training and feature-selection settings. They also present classifier and feature-selection outputs and visual analyses of workload distributions, injury likelihood across workload-ratio groups, and EWMA span performance.

S3 Appendix. The MSWR method

  • High MSWR values were associated with high injury risk for most workload features, but MSWR predictors had low average recall (0.10 ± 0.10) and precision (0.03 ± 0.03), comparable in predictive power to ACWR predictors.The risk association substantially confirmed results observed in the literature by Foster et al..Source: S3 Appendix. The MSWR method · pp. 24–25

S4 Appendix. Example of the training dataset construction

  • The example dataset assigns each training session a 42-feature vector and labels it 1 when associated with an injury, otherwise 0.In the toy example, the first three sessions are labeled 0 and the fourth 1.Source: S4 Appendix. Example of the training… · p. 25

S5 Appendix. Exponential Weighted Moving Average (EWMA)

S6 Appendix. Computation of PI(WF)

  • PI(WF) is an EWMA of prior weekly injury counts that reflects the time since return to regular training; 0 denotes no past injuries, while values above 0 denote at least one.PI(WF) = 0.50 corresponds to training 3 days after return to regular training.Source: S6 Appendix. Computation of PI(WF) · p. 26

S7 Appendix. Adaptive synthetic sampling approach

  • ADASYN adaptively generates minority-class synthetic samples according to majority-class density, interpolating between examples to produce a more balanced dataset.Each synthetic example is formed as xi+(xzi-xi)λ, with λ randomly chosen between 0 and 1.Source: S7 Appendix. Adaptive synthetic… · p. 26

S9 Appendix. Predictions results

Related papers

Loading 1705.08079v2…