Abstract
Surface electromyography (sEMG) measures electrical signals in muscles during contraction. Its established applications include medical diagnostics, prosthetics, and rehabilitation. However, in sports science, particularly professional football (soccer), sEMG’s potential remains largely unexplored. This research investigates the use of sEMG in football, which is significantly affected by both financial implications and competitive performance due to injuries. The study focuses on commonly observed football injuries: hamstring, adductor, and soleus injuries (HASI). The analysis utilises sEMG data from 45 elite athletes affiliated with two top English Premier League (EPL) clubs over 2.5 seasons. Data collection, predominantly on the day following matches (MD+1), aligns with professional sports practices, enhancing the dataset’s relevance for injury analysis.The core novelty of this thesis is a comprehensive framework that transforms raw sEMG data into actionable injury predictions. This systematic approach meticulously converts sEMG recordings into probabilities of HASI occurrence, with each stage contributing distinctly to sports science and sEMG analytics.
A significant innovation is the development of a deep learning model for accurate drill-time identification in continuous sEMG recordings. This model effectively discerns genuine drill-related muscle contractions, addressing a common challenge in continuous data segmentation within sports. Athletes in sports settings undertake various activities such as walking, jumping, or stretching during exercise interludes. This adds complexity in differentiating exercise-related muscle activations from incidental movements. The drill-time labelling model incorporates custom loss regularisation specific to football training and a kernel and depth selector attention module. This module enhances feature learning versatility and model performance, accurately identifying drill times amid diverse athletic activities.
Furthermore, the study introduces a novel approach for denoising Motion-Induced Artefacts (MIA) in sEMG data, which is common in dynamic sports environments. Athletes engaging in high-motion exercises result in unavoidable MIA inclusion. The proposed model first detects contaminated zones in the data sequence and then employs a deep learning-based denoising process to clean the data while preserving the original sEMG signals’ integrity. This dual approach outperforms existing state-of-the-art models and is particularly effective in sports settings.
In motion classification, the research surpasses existing models on public hand gesture datasets and adeptly categorises football-specific lower extremity drills. The deep learning strategy divides input sEMG data into low and high-frequency bands, processed in parallel. Channel and temporal attention mechanisms focus on significant features within each band, enhancing performance and versatility. The model is effective for both hand gesture recognition and lower extremity football drills without parameter tuning. Additionally, a novel few-shot metric learning approach enhances classification accuracy, especially in sports settings with limited data. This approach outperforms state-of-the-art few-shot learning applications, demonstrating effectiveness in sparse dataset scenarios.
Another key innovation is the injury prediction model, hypothesising that muscle recruitment pattern deviations indicate impending injuries. The study uses a meta-learning-based deep metric learning technique to detect these deviations. This technique learns a representation space where healthy athlete states are near their baseline centroid, and risky states are more distant, forming a cluster between post-injury and safe data points. The model efficiently detects anomalies and accurately predicts HASI injuries, rapidly adapting to individual athletes’ risk traits. This adaptability is vital, considering that each athlete’s risk factors related to physiological states are unique and subjective. By learning these individual characteristics effectively, the model offers a nuanced and athlete-specific understanding of injury risks, enhancing the precision and relevance of HASI predictions.
In conclusion, this research underscores the substantial potential of combining sEMG with deep learning in sports science, particularly for injury prevention in professional athletes. It integrates comprehensive datasets, novel methodologies, and practical applications, positioning itself as a pivotal reference for future developments in sports injury prediction and prevention.
| Date of Award | 1 Jul 2024 |
|---|---|
| Original language | English |
| Awarding Institution |
|
| Supervisor | Panos Kosmas (Supervisor) & Osvaldo Simeone (Supervisor) |