TL;DR
This study explores using a single trunk-mounted accelerometer and machine learning to predict lower-limb overuse injuries (LLOIs) in young adults. The key finding is that gender-specific models significantly outperform mixed-gender models, achieving higher prediction accuracy, and that combining statistical and sports-specific features yields the best results.
What Problem Does This Paper Address?
Overuse injuries, particularly in the lower limbs, are a common and costly consequence of physical activity. While injury prevention is crucial, accurately predicting the risk of these injuries remains a significant challenge. Previous research has often focused on a limited set of risk factors and has not consistently accounted for individual differences like gender, which can influence biomechanics and injury susceptibility. This paper aims to develop a more effective machine learning model for LLOI prediction by investigating the impact of feature selection and gender-specific modeling, using data from a readily available sensor.
Methods
The researchers employed a machine learning approach for LLOI prediction using a nested cross-validation (CV) framework to ensure robust performance estimation. The core steps involved:
- Data Preprocessing: Raw accelerometer data from a Cooper test was cleaned, including tilt correction and removal of initial/final segments of running stages to focus on steady-state running.
- Feature Engineering: Two main sets of features were extracted:
- Statistical Features: Automatically generated from the accelerometer time series using the TSFuse package, aiming to capture dynamic patterns without domain-specific knowledge.
- Sports-Specific Features: Derived from domain expertise, focusing on measures related to dynamic stability, loading, and spatio-temporal characteristics, often requiring step detection.
- Questionnaire Features: Included demographic and injury history data (gender, weight, height, previous injuries, insole use).
- Model Training: Three machine learning algorithms were evaluated: L1-regularized logistic regression (LR), random forest trees (RFT), and support vector machines (SVM).
- Hyperparameter Tuning: A grid search within the internal CV loop was used to optimize hyperparameters for PCA components, feature selection, and model-specific parameters.
- Performance Evaluation: Area Under the Receiver Operating Characteristic Curve (AUC) and Brier Score were used as primary metrics.
A novel aspect was the investigation into the impact of gender by training both mixed-gender and gender-specific models.
Datasets
The study utilized data collected from 161 first-year undergraduate students (109 males, 52 females) from a movement science program at KU Leuven, Belgium. Data was collected during a single 12-minute Cooper test using a trunk-mounted 3D accelerometer. Participants were followed for six months to record the occurrence of lower-limb overuse injuries (LLOIs). This dataset is a newly collected one for this specific research purpose, focusing on a general active population rather than elite athletes.
Key Insights and Ideas
- Gender Matters Significantly: The most striking finding is that gender-specific models consistently outperformed mixed-gender models. This highlights that injury risk factors and their influence can differ substantially between males and females, necessitating tailored predictive approaches.
- Feature Combination is Key: Combining statistical (automatically derived) and sports-specific (domain-expert-driven) features generally led to better prediction performance than using either set alone. This suggests these feature types are complementary, with statistical features often proving more informative.
- Statistical Features Show Promise: While sports-specific features are intuitive, the study found that statistical features, derived automatically, were often more informative for LLOI prediction, particularly in female-specific models. This opens avenues for leveraging data-driven feature extraction.
- Directionality of Features Varies by Gender: The relative importance of acceleration signal directions (e.g., vertical vs. medial-lateral) differed between genders, with vertical acceleration features being more influential for males and medial-lateral for females. This further supports the need for gender-specific analysis.
Key Results
- Best Performing Models:
- Female-specific: Logistic Regression (LR) with all features achieved the highest AUC of 0.645 ± 0.056.
- Male-specific: Logistic Regression (LR) with all features achieved the highest AUC of 0.615 ± 0.063.
- Mixed-gender: Support Vector Machine (SVM) with all features had the highest AUC of 0.557 ± 0.091, which is only slightly better than random guessing.
- Feature Set Performance: For both genders, models using statistical features generally performed better than those using only sports-specific features. The combination of statistical and sports-specific features often yielded the best results, especially for female models, suggesting greater complementarity.
- Feature Importance: For females, statistical features in the medial-lateral direction were highly influential. For males, statistical features in the vertical direction were most important. Non-directional features (often questionnaire-based) generally had low importance.
Limitations and Future Work
- Generalizability: The study participants were primarily undergraduate students in a movement science program, with similar ages and training loads. This limits the direct generalizability of the findings to other populations, such as elite athletes or different age groups.
- Model Performance: While improved by gender-specific modeling, the AUC scores (around 0.61-0.64) are still considered moderate, indicating room for improvement. The authors acknowledge that the models are not "high-performing" and could be enhanced.
- Missing Data: Participants with missing values or non-LLOI injuries were excluded, which could introduce bias.
- Limited Feature Scope: The study relied on a single Cooper test and a limited set of intrinsic factors. Future work could incorporate more risk factors (e.g., previous sports participation, footwear, detailed anthropometrics) and longitudinal data from multiple testing sessions to capture changes over time and fatigue more comprehensively.
- Statistical Feature Interpretability: While statistically informative, the automatic generation of statistical features makes them less interpretable for practitioners, posing a challenge for translating them into actionable interventions. Further research is needed to bridge this gap.
- Transferability: Future studies should explore the transferability of these models to different sports, settings, and populations.
Summary updated