Current cheerleading skill assessment predominantly relies on coaches' subjective evaluations, lacking quantifiable biomechanical benchmarks. Although electromyography and ground reaction forces provide critical neuromuscular and kinetic insights, their computational integration for automated motion quality assessment remains underdeveloped in aesthetic sports. This study systematically investigates the Tour à la Seconde through explainable machine learning, establishing an interpretable human-AI collaboration framework to bridge empirical coaching knowledge with data-driven biomechanical diagnostics. Twenty-eight elite athletes (18-22 years, 3+ years training) performed standardized Tour à la Seconde while synchronized 14-lead EMG (2000Hz) and triaxial force plate data were captured. Biomechanical feature engineering encompassed: Neuromuscular dynamics: RMS amplitudes, inter-muscular coherence. Kinetic signatures: Vertical force phasing (Fzₚeak), medio-lateral impulse (Fyᵢmpulse), dynamic stability index. A hierarchical modeling approach employing PyCaret's automated ML optimized 12 classifiers via leave-athlete-out cross-validation. SHAP value decomposition elucidated biomechanical determinants of expert-rated performance (binary classification threshold: coach score ≥7/10). The XGBoost model achieved superior generalizability (86% cross-validation accuracy vs. LightGBM's 82%, logistic regression's 76%), with three biomechanical drivers identified: Rotational stability control: Fzₚeak timing (SHAP=0. 41) correlating with centrifugal force management. Ankle proprioception: Gastrocnemius activation intensity (0. 33) reflecting plantarflexion precision. Neuromuscular synergy: Inter-muscular coordination (0. 28), indicating kinetic chain efficiency. Notably, while decision trees showed complete training accuracy (100%), 21% test accuracy drop revealed critical overfitting risks. The model's strong congruence with expert evaluations (κ=0. 72, 95%CI: 0. 65-0. 79) validates its potential as a coach-assistive tool, particularly through real-time visualization of weight transfer dynamics and muscle activation sequencing. This work establishes a novel computational paradigm for aesthetic sports biomechanics, decoding cheerleading rotational skills through explainable multimodal learning. The framework's technical viability for real-time deployment—evidenced by latency-optimized feature engineering (< 50ms processing time) — AI-enhanced athletic training systems. Future research directions include inertial sensor fusion for 3D kinematic validation and cross-domain adaptation to gymnastics and dance.
Yutao Chen (Thu,) studied this question.