Zhao, Xiong2025-11-182025-11-182025-11-18http://hdl.handle.net/10393/51058https://doi.org/10.20381/ruor-31524Background: Non‑contact lower limb injuries are a leading source of loss time and performance impairment in female soccer athletes. Laboratory‑based screening methods offer detailed biomechanical insights but lack scalability and field applicability. Recent advances in markerless motion capture and data‑driven modeling promise more accessible and interpretable tools for longitudinal athlete monitoring. Purpose: The overarching goal of this thesis is to develop a comprehensive framework that integrates biomechanical data, machine learning models, and markerless motion capture systems to enhance injury prediction and movement health monitoring in female varsity soccer athletes. Methods: A single cohort of 25 female varsity soccer players (mean age 20.40 ± 1.98 years) was assessed at pre‑, mid‑, and post‑season during a suite of standardized tasks: countermovement jump, broad jump, lateral hop, Y‑Balance, T‑Balance, metronome-paced bodyweight squats, reaction time tests, and on‑field sprint tests. Four interrelated studies were conducted: 1) Seasonal changes in dynamic tasks and sprints assessed via in-lab and field measures to identify performance fluctuations and recovery patterns. 2) A 3D Goal Equivalent Manifold (GEM) analysis of metronome‑paced squats to quantify task‑relevant and task‑irrelevant motor variability and detect persistent motor control deficits in athletes with concussion history. 3) An evaluation of four machine learning pipelines, Support Vector Machines (SVM), eXtreme Gradient Boosting (XGBoost), Long Short-Term Memory (LSTM), and a Hybrid Transformer, for interpretable injury prediction in sports. 4) Validation of 28 two‑camera configurations for lower limb joint kinematics accuracy and agreement within configurations and against an 8‑camera markerless system (Theia3D). Results: In Study 1a, significant midseason decrements were observed in sprint times (10 m: χ² = 15.56, p < .001; 20 m: χ² = 31.21, p < .001; 40 m: χ² = 7.59, p = .023) and countermovement jump height (bilateral: χ² = 18.89, p < .001; unilateral: χ² = 8.78, p = .012), with partial recovery post‑season, while broad jump and L‑hop remained stable. Left lower‑limb coordination amplitude (MARP, χ² = 8.09, p = .018) and composite reach score decreased across the season. Study 1b revealed that athletes with concussion history had significantly lower 3D GEM performance indices across all timepoints, indicating persistent motor control deficits. In Study 2, the Hybrid Transformer outperformed LSTM (AUC = 0.680), XGBoost (AUC = 0.619), and SVM (AUC = 0.342), achieving AUC = 0.740 and 67.6 % accuracy; attention maps highlighted early landing and push‑off phases as critical risk periods. Study 3 demonstrated that two‑camera setups (i.e. front-back configurations) estimated sagittal‑plane hip and knee flexion/extension with RMSE < 8° and CMC ≈ 1.00, while coronal and transverse planes showed larger errors (RMSE > 10°, CMC < 0.60). Conclusion: This thesis demonstrates that a unified framework, combining longitudinal biomechanical screening, transformer‑based injury prediction, GEM‑based motor variability analysis, and field‑deployable markerless motion capture, can accurately identify performance decrements and injury risk in female soccer athletes. The findings support scalable, interpretable, and clinically relevant tools for season‑long athlete monitoring and targeted injury prevention strategies.enAttribution 4.0 Internationalhttp://creativecommons.org/licenses/by/4.0/women's soccerbiomechanicsinjury preventionmachine learningmarkerlessThesis