Source-linked AI summary
Multimodal Injury Risk and Performance Prediction in Tennis Using Weighted Ensemble Learning
Weihao Qu, Dongyang Wang, Ling Zheng, Francisco E. Alvarez, Shobharani Polasa, Jiacun Wang
TL;DR
Tennis injury and performance prediction is limited by subjective assessments and underexplored multimodal integration. PART combines heterogeneous athlete data through specialized models and supervised weighted ensembles, achieving strong monitoring and near-term risk-estimation performance in a nine-player collegiate pilot, while requiring larger validation.
Problem
Multimodal approaches for tennis remain underexplored, while existing prediction methods often depend on subjective observations and expert assessments.
Method
PART uses specialized models to extract wellness, injury risk, physical capability, and playing style from multimodal data, then combines them with supervised weighted ensemble learning into an Athlete Readiness Score.
Results
PART demonstrates effective athlete-condition monitoring and near-term injury-risk estimation in case studies of nine collegiate tennis players; its overall wellness XGBoost model achieved an R2 score of 0.65.
Takeaways & Limitations
Within the collegiate pilot scope, PART provides personalized readiness and injury-risk insights, including identification of injury-prone areas and support for actionable interventions.
Takeaways & Limitations
The nine-player cohort limits statistical power and injury-pattern diversity, and playing-style weights are heuristic rather than learned from data.
Abstract
from arXiv · showhide
Machine learning has had a positive impact on the sports industry, with one of its most promising applications being the prediction of athlete performance and injury risk. Recent advances have employed state-of-the-art models to improve prediction accuracy, yet progress remains limited by data availability and the reliance on subjective observations or expert assessments. To address these limitations, researchers in sports such as soccer, basketball, and wrestling have begun integrating heterogeneous data sources, such as wearable device readings, with traditional subjective assessments. However, similar multimodal approaches remain underexplored in tennis. In this work, we propose a multimodal weighted ensemble learning framework, Predictive Athlete Readiness for Tennis (PART), to monitor athlete wellness and estimate near-term injury risk in tennis players. PART processes a wide range of inputs, including physiological metrics, training and match data, sleep information from wearable devices, self-reported questionnaires, vertical jump assessments, and motion analysis from match-play videos. From these modalities, specialized machine learning and deep learning models independently extract four athlete-specific characteristics: overall wellness, injury risk, physical capability, and playing style. To overcome the complexity of combining these diverse modalities, PART employs a supervised weighted ensemble integration strategy, assigning adaptive weights to each predictive model based on its reliability. Evaluation of multimodal data collected from nine collegiate tennis players demonstrates that PART achieves strong performance in monitoring athlete wellness and estimating near-term injury susceptibility. Beyond collegiate athletes, the framework also shows promise for recreational tennis players, offering personalized insights to mitigate injury risk and optimize performance.
I. INTRODUCTION
Tennis performance and injury prediction must account for interacting physiological, psychological, training, and lifestyle factors. PART addresses limitations of subjective, single-source approaches by combining multimodal data through weighted ensemble learning.
- Tennis combines athleticism, strategy, and mental resilience, making performance and injury prediction especially challenging.
- Overuse and repetitive-motion injuries are common among professional and recreational tennis players, motivating tools that address injury risk and performance.
- Existing tennis models often rely on official match statistics, subjective observations, and expert assessments, limiting their capture of interacting factors.
- Wearables and other richer data sources can provide objective physiological, workload, and recovery information alongside athlete self-reports.
- Naively combining heterogeneous modalities can dilute signal quality or overweight less reliable sources, motivating weighted ensemble learning.
- PART uses specialized models for modality-specific outcomes and supervised weighted integration to assess performance and injury risk in tennis players.
A. System Architecture
PART uses a modular two-stage architecture: specialized models first learn athlete characteristics from diverse inputs, then supervised weighted integration produces an Athlete Readiness Score. The framework draws on wearable, questionnaire, testing, and video data from collegiate players.
- PART’s architecture contains input, preprocessing, separating, and integration layers for processing diverse data types.
- The input data come from WHOOP sleep and workout records, questionnaires, professional vertical-jump testing, and match-play videos.
- Four specialized models separately learn physical capability, injury risk, overall wellness, and playing style.
- The integration stage combines the four model outputs with supervised weighted ensemble learning to calculate the Athlete Readiness Score.
- The study collected multimodal data from nine collegiate tennis players over 16 weeks.
B. Data Collection
PART collects four complementary data sources spanning self-reports, wearable measurements, physical testing, and match-play video. Textual inputs undergo cleaning and feature engineering before modeling.
- Questionnaires track training load, discomfort and pain, mental stress, nutrition, hydration, and reported discomfort locations.
- WHOOP devices provide objective sleep, continuous physiological, and detailed workout data for training sessions and matches.
- Vertical-jump measurements are collected through standardized weekly or bi-weekly Sparta testing.
- Match-play videos collected during the spring 2024 season support video-based analysis of player movement and style.
- Questionnaire, WHOOP, and Sparta data are cleaned through missing-value standardization, type conversion, and imputation.
- Feature engineering selects physiological measures including HRV, RHR, sleep efficiency, and activity strain, with transformations and polynomial features used to model complex relationships.
IV. SEPARATING STAGE: MODEL DEVELOPMENT
The separating stage develops task-specific models for wellness, injury risk, physical capability, injury area, and playing style using textual and video data. For overall wellness, XGBoost regression provides the selected model after outperforming linear alternatives.
- The separating stage uses different models for textual and video data to predict wellness, injury risk, injury area, physical capability, and playing style.
- Overall wellness prediction estimates a Recovery Score from 0 to 100 using physiological, strain, and sleep inputs.
- XGBoost Regressor achieved an R2 score of 0.65 and outperformed the evaluated linear regression models for overall wellness.
- Player-normalized HRV, sleep performance, and raw HRV were the three most important features in the overall wellness model.
B. Injury Risk Classification (Classification Models)
The injury-risk classification stage predicts high- versus low-risk status from physiological and self-reported data, while also evaluating body-region-specific risk. Class imbalance limits correct identification of injury cases despite useful discrimination by several classifiers.
- Classification objective: Injury risk is a binary high- or low-risk label derived from daily questionnaires, physiological metrics, and self-reported responses.Inputs include Activity Strain, HRV, and Resting Heart Rate.
- Classification models: XGBoost, Decision Tree, and Random Forest classifiers showed similar AUC-ROC and F1 performance in distinguishing athletes with and without injury risk.The passage reports comparative performance but does not provide the individual metric values.
- Class imbalance: Approximately 24% of observations represented injury cases, and low injury-class F1 scores indicate difficulty predicting the minority class.The authors addressed imbalance using class weights and stratified sampling.
- Feature importance: HRV (0.334), Resting Heart Rate (0.184), and Recovery Score (0.112) were the three most important features for injury-risk prediction.The XGBoost confusion matrix highlights stronger identification of non-injury cases than injury occurrences.
- Body-region models: 0.712 AUC-ROC was achieved for upper-body injury prediction, compared with 0.703 for lower-body injury prediction.Separate XGBoost models enabled more targeted injury-risk profiling.
C. Physical Capability Prediction
Physical capability is represented by a score combining readiness indicators and predicted from seven-day physiological and activity sequences. The LSTM substantially outperformed the MLP and captured day-to-day variation in athlete readiness.
- Capability measure: The Physical Capability Score combines jump performance, sleep, heart-rate variability, and questionnaire-based quality indicators.Its components include Reactive Strength, Vertical Jump Performance Score, Sleep Efficiency, HRVscore, Sleep Score, and Quality Score.
- Prediction models: MLP and LSTM models predicted physical capability from seven-day sequences of physiological and activity metrics.Training used dropout, early stopping, seven-day sequences, and player-based group cross-validation.
- Model comparison: The LSTM achieved an RMSE of 0.749 versus 2.506 for the MLP and an R² of 0.993 versus 0.922.These results demonstrate stronger performance for the LSTM on the reported prediction task.
- Readiness variation: Predicted physical capability scores on the held-out test set spanned approximately 60–78%.The range reflects readiness variation across different days and conditions.
- Playing style: Average Shots per Point classified players as highly aggressive, aggressive, neutral, or defensive based on rally length.The thresholds were ASP ≤5, 5 < ASP ≤8, 8 < ASP ≤12, and ASP >12, respectively.
E. Discussion
PART selects predictive models according to the factor being analyzed and relates physiological indicators to injury probability. Lower HRV, elevated resting heart rate, and higher activity strain are associated with stress, fatigue, inadequate recovery, or greater physical stress.
- Model choice: PART uses XGBoost models for nonlinear physiological regression and injury classification, while LSTM and MLP models process temporal training data for physical capability.The model choices are tied to computational efficiency, imbalanced classification, and temporal-sequence processing.
- Physiological indicators: Lower HRV indicates higher stress and inadequate recovery, while elevated Resting Heart Rate signals fatigue or overtraining.Higher activity strain is also correlated with greater physical stress.
- Practical implication: Integrating physiological indicators into predictive models supports targeted interventions intended to mitigate injury risks.The paper connects monitoring these metrics with proactive identification of athletes at risk.
V. INTEGRATION STAGE: MODEL INTEGRATION AND OUTPUT GENERATION
The integration stage combines specialized predictions into an Athlete Readiness Score using supervised linear weighting. The initial weights assign equal importance to overall wellness and physical capability, with slightly lower weight for injury risk.
- Supervision: Expert evaluations from coaches and sports scientists supplied 0–100 readiness labels incorporating performance, recovery, training load, and subjective wellness.These labels enabled supervised determination of ensemble weights.
- Weighting design: The integration design uses a linear weighting scheme for Overall Wellness Score, Injury Risk, and Physical Capability Score.The scheme was chosen for simplicity, transparency, interpretability, and representation of modality importance.
- Training data: Historical data from September to December 2023 determined the weights using WHOOP, vertical-jump, questionnaire, coach-evaluation, and match-performance data.Weekly collection combined objective, self-reported, expert, and match-based information.
- Learned weights: The initial Athlete Readiness Score weights were w1 = 0.35, w2 = 0.30, and w3 = 0.35.The weights correspond to Overall Wellness Score, Injury Risk, and Physical Capability Score, respectively.
C. Incorporating Video Analysis Results
PART uses video-derived playing style to adjust Athlete Readiness Scores for differing physical demands, producing more personalized wellness and injury-risk assessments.
- Playing style from video analysis serves as a weighting factor in Athlete Readiness Score calculations.The framework assigns higher recovery-metric weights to defensive players, who typically experience greater physical demands during matches.
- Style-specific weights are 1.0 for highly aggressive, 1.2 for aggressive, 1.4 for neutral, and 1.6 for defensive players.The final ARS is adjusted by dividing the initial score by the style weight.
- The style adjustment is intended to account for varying physical demands and personalize readiness and injury-risk assessments.
- ARS scores above 80 indicate good condition and low injury risk, whereas scores below 60 indicate suboptimal condition and high injury risk.Scores from 60 to 80 indicate standard physical condition; lower-scoring players may need evaluation, reduced training loads, or adjusted match schedules.
- P7 recorded a final ARS of 93.25, while P2, P4, P6, and P8 scored below 60; P8 had the lowest score at 37.50.The reported scores distinguish excellent wellness and minimal injury risk from accumulated fatigue and need for recovery.
- The case-study results support monitoring and near-term risk estimation, but not validated prospective injury forecasting.The evidence comes from a pilot study involving nine players and is presented as qualitative feasibility evidence for future larger-scale validation.
A. Case Study 1: Non-Fatigue Injury Analysis
The first case study shows that PART can track recovery after an acute ankle sprain, while distinguishing sudden mechanical injuries from fatigue-related risk patterns.
- The April 2024 case study captures an ankle sprain injury event and its aftermath.
- Ankle sprains can arise from sudden biomechanical events rather than accumulated stress, making them inherently difficult to predict from gradual ARS decline.
- An ARS of approximately 60 on April 11 placed the athlete in the standard condition range.
- The post-injury ARS rose steadily to 95, entering the good-condition threshold as physical strain and injury risk declined.The reported improvement was associated with sufficient rest and improved wellness.
- The recovery trend demonstrates PART’s ability to capture post-injury recovery dynamics and signal a return to stable condition.
- A separate February 2024 case showed ARS falling from about 70 to roughly 40 after three consecutive matches, indicating cumulative fatigue.The score shifted from the standard condition range into the high-risk zone.
VII. CONCLUSIONS
PART is a two-stage multimodal framework that uses ARS to monitor tennis-athlete wellness and estimate near-term injury risk. Its conclusions are preliminary because the study used nine collegiate players and heuristic style weights.
- PART separates multimodal inputs into four predictive characteristics and integrates them into an Athlete Readiness Score.The four characteristics are physical capability, injury risk, overall wellness, and playing style.
- The study collected data from nine collegiate tennis players over one semester using questionnaires, wearable devices, professional testing protocols, and video recordings.The reported results should therefore be interpreted within this cohort and data-collection scope.
- Playing-style weighting factors are heuristic values informed by domain expertise rather than learned from data.The authors describe these factors as preliminary.
- The cohort size of nine players limits statistical power and the ability to capture the full diversity of injury patterns.The authors call for validation with larger and more diverse populations.
- Future work proposes implementing PART as intelligent tennis-coach software that provides personalized advice using large language models.