Abstract
This study investigated differences in energy-expenditure (EE) modeling between badminton players of varying competitive levels during aerobic training. It evaluated the impact of sensor quantity and sample size on prediction model accuracy and generalizability, providing evidence for personalized training-load monitoring. Fifty badminton players (25 elite, 25 enthusiasts) performed treadmill running, cycling, rope skipping, and stair walking. Data were collected using accelerometers (waist, wrists, ankles), a heart rate monitor, and indirect calorimetry (criterion EE). Multiple machine learning models (Linear Regression, Bayesian Ridge Regression, Random Forest, Gradient Boosting) were employed to develop EE prediction models. Performance was assessed using R(2), mean absolute percentage error (MAPE), and root mean square error (RMSE), with further evaluation via the Triple-E framework (Effectiveness, Efficiency, Extension). Elite athletes demonstrated stable, coordinated movement patterns, achieving the best values for R(2) and the smallest errors using minimal core sensors (typically dominant side). Enthusiasts required multi-site sensors to compensate for greater execution variability. Increasing sensors beyond three yielded no performance gains; optimal configurations involved 2-3 core accelerometers combined with heart rate data. Expanding sample size significantly enhanced model stability and generalizability (e.g., running task R(2) increased from 0.49 (N = 20) to 0.95 (N = 40)). Triple-E evaluation indicated that strategic sensor minimization coupled with sufficient sample size maximized predictive performance while reducing computational cost and deployment burden. Competitive level significantly influences EE modeling requirements. Elite athletes are suited to a "low-sensor, small-sample" scenario, whereas enthusiasts necessitate a "multi-sensor, large-sample" strategy.