Injury risk classification from feature grouped tabular athlete data using deep neural networks: a cross sectional benchmark study.
- 1School of Marxism, Beijing Sports University, Beijing, Beijing, 100084, China.
- 2School of Physical Education, Liaoning Normal University, Dalian, Liaoning, 116029, China. [email protected].
Abstract
The classification of sports injuries is one of the most important tasks in athlete monitoring, and commonly available benchmark datasets are often small, with limited injury classes, missing data, and inadequate documentation of injury occurrence, injury labels, and data provenance. This study evaluates a feature-grouped deep neural network for binary injury-risk classification on the University Football Injury Prediction dataset from Kaggle. The 18 input variables are grouped into four feature groups in tabular format: demographics and anthropometrics, training factors, physical fitness indicators, and lifestyle habits. A separate branch of the neural network encodes each feature group, and concatenation or attention-based gates fuse them. Under a leakage-aware preprocessing and evaluation protocol, the proposed model is compared against logistic regression, random forest, gradient boosting, and a single-stream multilayer perceptron. The performance metrics are: AUROC, AUPRC, F1 score, balanced accuracy, Brier score, expected calibration error, bootstrap confidence intervals, and seed-based stability analysis. Results indicate minimal gains from feature-grouped neural fusion over classical and single-stream baselines, particularly in precision-recall and probability calibration. These results, however, must be treated with caution, as the benchmark is internally assessed, small, and lacks a complete, independently verified longitudinal follow-up or a complete, documented labeling procedure. Explanability analysis is thus discussed in terms of the descriptive interpretation of models, rather than the causal identification of injury risk factors. In summary, this work offers a reproducible reference standard for feature-grouped tabular injury classification and underscores the necessity of externally validated, prospectively collected, longitudinal athlete datasets before such classification models can be considered for real-world training and/or clinical decision-making.
Related
- Feature-Engineering Strategies for EEG-Based ADHD Classification in Children: A Controlled Benchmark of Expert, LLM-Guided and Full-Space Comparators
- Timing and Context Features in Machine Learning Classification of Inter-Patient ECG Heartbeats on the MIT-BIH Benchmark
- Multi-task deep learning framework for brain tumor classification and segmentation using ensemble feature extraction
- Feature-aware attention convolutional neural network for efficient pulmonary fibrosis classification from chest CT images with FPGA deployment
- Comparative Evaluation of Large Language Model Interfaces in Third Molar Surgery Complication Scenarios: Response Quality, Clinical Content, Potential Clinical Risk, Readability, and Externally Observable Response Latency—A Cross-Sectional Comparative Benchmark Study
BibTeX
@article{yan2026injury,
title = {{Injury risk classification from feature grouped tabular athlete data using deep neural networks: a cross sectional benchmark study.}},
author = {Han Yan and Xiaoqiao Sun},
journal = {BMC sports science, medicine \& rehabilitation},
year = {2026},
month = jul,
doi = {10.1186/s13102-026-01911-6},
url = {https://europepmc.org/article/MED/42902427},
note = {CC BY-NC-ND},
}
RIS
TY - JOUR
TI - Injury risk classification from feature grouped tabular athlete data using deep neural networks: a cross sectional benchmark study.
AU - Han Yan
AU - Xiaoqiao Sun
PY - 2026
DA - 2026/07/01
JO - BMC sports science, medicine & rehabilitation
DO - 10.1186/s13102-026-01911-6
UR - https://europepmc.org/article/MED/42902427
AB - The classification of sports injuries is one of the most important tasks in athlete monitoring, and commonly available benchmark datasets are often small, with limited injury classes, missing data, and inadequate documentation of injury occurrence, injury labels, and data provenance. This study evaluates a feature-grouped deep neural network for binary injury-risk classification on the University Football Injury Prediction dataset from Kaggle. The 18 input variables are grouped into four feature groups in tabular format: demographics and anthropometrics, training factors, physical fitness indicators, and lifestyle habits. A separate branch of the neural network encodes each feature group, and concatenation or attention-based gates fuse them. Under a leakage-aware preprocessing and evaluation protocol, the proposed model is compared against logistic regression, random forest, gradient boosting, and a single-stream multilayer perceptron. The performance metrics are: AUROC, AUPRC, F1 score, balanced accuracy, Brier score, expected calibration error, bootstrap confidence intervals, and seed-based stability analysis. Results indicate minimal gains from feature-grouped neural fusion over classical and single-stream baselines, particularly in precision-recall and probability calibration. These results, however, must be treated with caution, as the benchmark is internally assessed, small, and lacks a complete, independently verified longitudinal follow-up or a complete, documented labeling procedure. Explanability analysis is thus discussed in terms of the descriptive interpretation of models, rather than the causal identification of injury risk factors. In summary, this work offers a reproducible reference standard for feature-grouped tabular injury classification and underscores the necessity of externally validated, prospectively collected, longitudinal athlete datasets before such classification models can be considered for real-world training and/or clinical decision-making.
SN - 2052-1847
ER -