Author(s): TAO, K., YAN, X., Institution: BEIJING SPORT UNIVERSITY, Country: CHINA, Abstract-ID: 1037

INTRODUCTION:
Badminton is a dynamic and highly technical sport, which requires rapid responses and transitions for serve, stroke, and racquet- and ball-handling skills. The potentiation of training requires athletes to adapt to the kinetic and physiological variations. Such adaptations are determined by training loads, which include a combination of training volume, intensity and frequency. Moderate to extreme training loads increase the performance yet risk the fatigue and sports injury. A knowledge of relationship between internal and external loads were thereby investigated by a spectrum of studies. Recent application of using inertial measurement units (IMU) along with machine learning algorithms offer a promising solution for on-field capture of relations between internal and external loads.
METHODS:
Thirty female first-tier badminton athletes, were recruited and randomly assigned to a standard indoor badminton court for singles and doubles matches. Each game adopts a best of three format, using the official 21-point or 15-point rally scoring system. Before warm-up, the participants put on vests without wearing IMU devices. After completing warm-up session, a Catapult Vector S7 is equipped with vests of players to collect external training load data. Fifteen minutes after match, athletes were required to evaluate their internal training load levels based on subjective perceptions of muscle and cardiovascular situations via a printed Foster's 0-to-10 point SRPE scale questionnaire. Statistical analysis was performed to select kinematic features. We further applied Linear Regression (LR), Bayesian Ridge Regression (BRR), Random Forest (RF) and Gradient Boosting (GB) pipelines to validate the accuracy and robustness of models.
RESULTS:
The findings demonstrated that RF outperformed other models achieving the goodness-of-fit to 0.88 and reaching the least errors of minimal features. LR acquired the most accountability of the direct response to internal and external loads, though the accuracy merely exceeded 0.55. Also, elite badminton players displayed a stable, coordinated movement patterns, with Matthew’s Correlation Coefficient (MCC) being 90%, showcasing a strong predictive power. Expanding sample size (N) significantly enhanced model stability and generalizability (e.g., goodness-of-fit increased from 0.49 when N = 10 to 0.76 when N = 20).
CONCLUSION:
This research specifically targeted at female athletes and highlighted the predictive ability through IMU movement data and machine learning pipelines in badminton. We further discussed the impact of the change of competition rules on the relationship between external and internal load. These findings provide a quantitative foundation for monitoring, injury prevention, and personalized training programs for coaching.