A preprint study of PART, a multimodal tennis model, reported an accuracy of 0.852 for overall injury-risk classification, alongside an AUC-ROC of 0.695 and an F1 score of 0.100. The figures come from nine collegiate players whose data were collected over 16 weeks, so they describe performance on this dataset rather than a calibrated individual risk estimate.
Those three metrics answer different questions. AUC-ROC indicates how well a classifier separates outcomes, accuracy records the share of classifications that were correct, and F1 balances positive cases correctly identified against missed or false calls. The paper describes the injury dataset as imbalanced, so the accuracy figure cannot be read on its own.
The paper also reports an F1 score of 0 for Logistic Regression and notes dataset imbalance and model limitations in this setting. Because the supplied analysis does not report injury-event counts or class sizes, the figures alone do not show how strongly that imbalance shaped the result.
The injury figures need context
For broad injury-type prediction, the reported upper-body figures were an AUC-ROC of 0.712, accuracy of 0.865 and F1 of 0.132. The lower-body figures were 0.703, 0.858 and 0.118. These outputs cover broad upper- and lower-body categories; they do not demonstrate that the model can pinpoint a specific injury site.
Different tasks, different scores
For overall wellness, the XGBoost Regressor reported a mean absolute error of 3.82, a root mean squared error of 6.81 and an R-squared score of 0.838. Mean absolute error and root mean squared error describe the size of prediction errors, while R-squared describes how closely estimates track the target. These are model scores for the study's wellness task, not evidence of clinical benefit.
Physical capability was evaluated separately with an MLP and an LSTM. The MLP had a training loss of 0.0388 and a validation loss of 0.6391, while the LSTM had corresponding losses of 0.0360 and 0.6059. The LSTM's validation loss was lower in the reported evaluation, which used an 80-20 training/validation split.
What PART put together
PART is organized around four characteristics: overall wellness, injury risk, physical capability and playing style. The study proposes it as a multimodal framework for assessing tennis performance and injury risk.
The study drew on four sources: wearable sleep, physiological and workout records; daily questionnaires; vertical-jump tests; and match-play video. It reports 85 attributes and records collected across the 16-week period from the nine players.
Before modeling, researchers standardized the data, converted field types and imputed missing values. They created an injury-risk target, selected relevant features, addressed multicollinearity and scaled numerical features.
A readiness score built from expert labels
Readiness was represented by a composite Athlete Readiness Score, or ARS. Its ground truth used a 0-to-100 expert-readiness scale, with multiple assessments averaged into a final label. The resulting score depends on expert assessments and the averaging choices used to create it.
Match video was used to derive playing style and adjust the ARS weighting. Defensive players described as facing greater match demands received higher weights for recovery metrics. That design makes the composite score dependent on the selected weighting scheme as well as the underlying assessments.
The study's reach is limited
The study included only nine collegiate players, five male and four female, with a reported mean age of 20.3 ± 1.5 years. No larger or external validation population was reported. The findings therefore provide direct evidence about model metrics in this group, not established performance across tennis players generally.
The injury results also lack information needed to translate them into individual risk estimates. The supplied analysis reports no confidence intervals, calibration results, absolute injury risks or observed injury counts. The evaluation did not test whether PART reduces injuries or improves athletic performance, so it is a model assessment rather than an intervention result.
The physical-capability findings have a similarly narrow evaluation context because the analysis used an 80-20 training/validation split. A lower validation loss in that comparison does not establish how the two model types would perform for other players.
What comes next
The authors conclude that PART provides robust wellness and injury prediction and could identify athletes' condition and injury-prone body parts. The supplied analysis identifies finer-grained body-area prediction and more robust video integration as future work.
Supplied metadata labels the item a preprint, while its front matter identifies it as an accepted author manuscript published in the 2025 IEEE 5th International Conference on Human-Machine Systems. The authors state that preprocessing, model-training and evaluation code is available in a public GitHub repository.
Paper data and sources
Original title: Multimodal Injury Risk Prediction in Tennis
Authors: Francisco Erramuspe Alvarez, Shobharani Polasa, Weihao Qu et al.
Journal/Repository: 2025 IEEE 5th International Conference on Human-Machine Systems (ICHMS), pp. 28-34, 2025
Status: Preprint, not yet peer-reviewed
First online: 2026-08-25
DOI: 10.1109/ichms61085.2025.11154160
Original paper · Full text