PaperlayerView exact revision on arXiv

Source-linked AI summary

Multimodal Injury Risk Prediction in Tennis

Francisco Erramuspe Alvarez, Shobharani Polasa, Weihao Qu, Jay Wang, Ling Zheng

arXiv:2608.25126v1cs.LG

TL;DR

Existing tennis injury and performance prediction often relies on subjective information, while multimodal approaches remain underexplored. PART integrates diverse athlete data to assess readiness and injury risk, achieving strong wellness and injury prediction performance in nine collegiate players.

  • Problem

    Tennis prediction research has relied heavily on subjective data, motivating more comprehensive approaches that combine diverse objective and subjective sources.

    Source: §I · pp. 1–2, §VI · p. 6

  • Method

    PART integrates video, physiological, workout, sleep, self-reported, and jump data to predict wellness, injury risk, physical capability, playing style, and injury-prone body areas.

    Source: §I · p. 2, §II · p. 2

  • Results

    XGBoost Regressor achieved an MAE of 3.82 and an R² of 0.838 for recovery-score prediction, while XGBoost Classifier achieved the highest F1 score and AUC-ROC for injury-risk prediction.

    Source: A. Overall Wellness Prediction… · p. 3, B. Injury Risk Classification… · p. 4

  • Takeaways & Limitations

    PART provides a holistic assessment of collegiate tennis athletes’ condition and identifies injury-prone body areas by combining multimodal data.

    Source: §II · p. 2, §VII · p. 6

    Future work targets finer-grained body-part predictions and a more robust method for incorporating video results.

    Source: §VII · p. 6

Abstract

from arXiv · show

Machine learning has had a significant positive impact on the prediction of athlete performance and injury risk. Most works in this field rely on subjective observations and expert assessments, which restrict their effectiveness. In sports like soccer, basketball, and wrestling, some studies attempt to address this challenge by integrating data from alternative sources, such as readings from wearable devices, alongside traditional subjective observations and expert assessments to enhance accuracy. However, similar research in tennis remains largely unexplored. In this paper, we propose a multimodal Predictive Athlete Readiness framework for Tennis (PART) to assess both performance and injury risk in tennis players. By leveraging machine learning and deep learning techniques, PART processes multiple sources of data collected from nine collegiate tennis players, including physiological metrics, training and match data, sleep data from wearable devices, self-reported information via daily questionnaires, jump assessments, and motion analysis from match play videos. PART captures four characteristics of tennis players: overall wellness, injury risk, physical capability, and playing style. By integrating these four characteristics by supervised learning, it is capable of providing a holistic assessment of the tennis athlete's condition, along with advanced forecasts of specific body areas at risk such as the upper body (e.g., elbows) or lower body (e.g., knees). Our evaluation, conducted with data from nine collegiate tennis players, shows that PART achieves strong performance in predicting both overall wellness and injury risk. Additionally, our framework also shows promise for recreational tennis players, who often suffer from injuries due to incorrect playing techniques.

I. INTRODUCTION

Tennis injury and performance prediction remains challenging because existing approaches often depend on limited statistical, subjective, and expert-assessed information. PART addresses this gap by integrating multimodal athlete data into readiness, wellness, and injury-risk predictions.

  • Tennis prediction is difficult because performance combines athleticism, strategy, and mental resilience, while injuries commonly arise from overuse and repetitive motions.Source: §I · p. 1
  • Existing tennis models often rely on official game statistics, subjective observations, and expert assessments, limiting coverage of interacting risk factors.Source: §I · p. 1
  • Wearables, self-reports, and athlete videos provide complementary physiological, lifestyle, and sport-specific information for prediction.Source: §I · pp. 1–2
  • Integrating more data sources improves coverage but substantially increases framework-design complexity, especially for wearable, survey, and video inputs.Source: §I · p. 2
  • PART integrates videos, physiological data, workouts, sleep, and surveys to predict wellness while considering physical capability and injury probabilities.Source: §I · p. 2
  • The framework collects multimodal data from nine collegiate athletes and combines models for physical capability, injury risk, wellness, and playing style through supervised learning.Source: §I · p. 2

II. SYSTEM ARCHITECTURE

PART uses a modular architecture that transforms multimodal athlete data into four specialized predictions and a composite Athlete Readiness Score. The study collected data from nine collegiate players over 16 weeks, including questionnaires, testing, videos, and WHOOP measurements.

  • PART contains input and preprocessing layers, four predictive models, and an integration model that generates athlete-condition outputs.Source: §II · p. 2
  • The four models estimate physical readiness, injury probability, overall wellness, and playing style using deep learning, XGBoost, and computer-vision methods.Source: §II · p. 2
  • The integration model synthesizes model outputs into an Athlete Readiness Score ranging from 0 to 100, with higher scores indicating better readiness.Source: §II · p. 2
  • PART also predicts injury-prone body areas from the injury-risk model, alongside the composite readiness assessment.Source: §II · p. 2
  • Data came from nine collegiate players over 16 weeks, with five male and four female participants.Source: A. Data Collection · p. 2
  • The dataset combines daily questionnaires, vertical-jump testing, match videos, sleep, physiological, training, and match data, totaling 85 attributes and 82,804 entries.Source: A. Data Collection · p. 3

B. Data Preprocessing and Feature Engineering

PART preprocesses heterogeneous textual and video data before modeling athlete wellness, injury risk, physical capability, injury area, and playing style. Feature engineering transforms, expands, and encodes predictors to support model learning.

A. Overall Wellness Prediction (Regression Models)

Overall wellness prediction estimates Recovery Score (%) from physiological and activity-related indicators using regression models. XGBoost Regressor performed best among the evaluated models.

B. Injury Risk Classification (Classification Models)

Injury risk classification uses physiological and self-reported features to predict overall injury risk and classify injury location. Multiple classifiers are compared, with XGBoost reported as the strongest overall model.

Classification task

  • The task predicts whether an athlete faces high or low injury risk within a specific time frame.Inputs include Activity Strain, HRV, Resting Heart Rate, and self-reported survey responses.Source: B. Injury Risk Classification… · p. 4

Classification models

  • The evaluated classifiers are Logistic Regression, XGBoost, Decision Tree, and Random Forest.The models estimate injury likelihood and support injury-location predictions for personalized interventions.Source: B. Injury Risk Classification… · p. 4

Injury-type prediction

  • Injury locations are categorized by upper or lower body, left or right side, and four body quadrants.The quadrants are upper left, upper right, lower left, and lower right.Source: B. Injury Risk Classification… · p. 4

Results

C. Physical Capability Prediction

Physical capability prediction combines physiological and subjective measures into a Physical Capability Score and uses recent seven-day history to forecast capability percentage.

Prediction approach

  • MLP and LSTM models predict physical capability percentage from seven-day sequences of physiological and activity metrics.A one-week interval was selected for LSTM because it performed best.Source: C. Physical Capability Prediction · p. 4

Score construction

  • The Physical Capability Score combines physiological and subjective metrics through a weighted formula.The supplied formulation assigns weights to Reactive Strength, Sleep Efficiency, and HRVscore.Source: C. Physical Capability Prediction · p. 4

Score components

  • Reactive Strength, Sleep Score, Quality Score, and Vertical Jump Performance Score provide the component measures used to characterize capability.They capture explosive strength, sleep quality, perceived wellness, and lower-body power, respectively.Source: C. Physical Capability Prediction · pp. 4–5, Table IV · p. 4

Evaluation setup

  • Historical PCS values over seven days are used to forecast physical capability percentage, with an 80-20 training-validation split.The historical window is intended to capture recent physiological and activity trends.Source: C. Physical Capability Prediction · p. 5, Table V · p. 5

D. Playing Style Analysis

Playing style analysis uses average shots per point from match videos to characterize rally and exertion patterns across four categories.

Playing style metric

  • The video-analysis system calculates average shots per point across all points played by each athlete.The metric is used to analyze playing patterns and physical exertion.Source: D. Playing Style Analysis · p. 5

Style categories

  • Players are classified as Highly Aggressive when ASP ≤5, Aggressive when 5 < ASP ≤8, Neutral when 8 < ASP ≤12, or Defensive when ASP > 12.The categories progress from quickly ending points to relying on consistency and extended rallies.Source: D. Playing Style Analysis · p. 5

V. MODEL INTEGRATION AND OUTPUT GENERATION

The Athlete Readiness Score is labeled using expert assessments aligned with corresponding objective and subjective features for supervised learning.

Readiness labeling

  • Coaches and sports scientists assign each athlete an overall readiness score on a comprehensive 0-to-100 scale.Assessments consider physical performance, recovery status, training load, and subjective wellness reports.Source: A. Athlete Readiness Score (ARS)… · p. 5

Label construction

  • The labeling process gathers objective metrics and subjective reports before expert evaluation.Examples include wearable physiology, performance statistics, self-reported wellness, and stress levels.Source: A. Athlete Readiness Score (ARS)… · p. 5
  • Multiple experts’ assessments are averaged to reduce individual bias, then aligned with features for supervised learning.The resulting labeled ARS data is paired with corresponding inputs at each time point.Source: A. Athlete Readiness Score (ARS)… · p. 5

B. Determining Weights via Supervised Learning

The framework learns weights for its Athlete Readiness Score through supervised learning, combining normalized component scores and style-based adjustments. Multiple Linear Regression is selected for interpretable coefficients, and the final adjustment is intended to personalize readiness and injury-risk assessments.

D. Code Availability

The paper makes its preprocessing, model-training, and evaluation code publicly available through a GitHub repository with replication instructions.

  • A public GitHub repository contains the preprocessing, model-training, and evaluation code.It also includes environment setup and instructions for running the experiments.Source: D. Code Availability · p. 6

VI. RELATED WORK

Related work spans multimodal sports data, time-series modeling, and tennis-specific performance and injury prediction. The paper positions PART as a data-driven framework addressing tennis research’s continued reliance on subjective information, while identifying future extensions for finer injury localization and video integration.

Integration of Wearable Technology and Self-Reported Data

  • Sports analytics research has combined wearable technology and self-reported metrics to study athlete stress, illness susceptibility, sleep, and training-related information.Source: §VI · p. 6

Time-Series Data

  • MLP and LSTM models are used in sports time-series research to capture temporal dependencies and nonlinear relationships in longitudinal athlete data.Source: §VI · p. 6

Specific to Tennis

  • Tennis research includes workload-based injury-risk studies, qualitative technique evaluations, and match-outcome models using historical performance statistics.Source: §VI · p. 6
  • Reliance on subjective data is identified as limiting tennis performance and injury prediction, motivating more comprehensive data-driven approaches.Source: §VI · p. 6
  • PART collects questionnaire, wearable, professional-testing, and video data from nine college athletes to predict athlete condition and injury-prone body parts.Source: §VII · p. 6

Future Work

  • Future work targets finer-grained body-part prediction, more robust video-data integration, and deployment as Intelligent Tennis Coach software for players at all levels.Source: §VII · p. 6

Related papers

Loading 2608.25126v1…