Author(s): TAO, K., MENG, K., GAO, B., QIU, J., Institution: BEIJING SPORT UNIVERSITY, Country: CHINA, Abstract-ID: 930

INTRODUCTION:
Dance, as a widely practiced physical activity worldwide, encompasses diverse styles and tempos, making it challenging to accurately assess energy expenditure. Traditional methods rely on empirical formulas embedded in ActiGraph accelerometers, calculating vector magnitude counts per minute to gauge dance intensity. However, these formulas often lead to significant bias in estimating energy expenditure. To address this issue, multiple wearable sensors and accelerometers have been utilized to reduce prediction bias. Nevertheless, this approach increases the complexity of energy expenditure models and impacts participants flexibility during dance. This study introduces the Triple-E principle: effectiveness, efficiency, and extension. By adhering to this criterion, we developed state-of-the-art (SOTA) machine learning models to assess energy consumption during dance with the highest accuracy, minimal complexity, and optimal suitability.
METHODS:
We enrolled 250 participants (mean age ±SD: 63.0 ±6.0 years) who engaged in one of three dance styles (ballroom, aerobic, square). Before the experiment, they completed a survey on anthropometric measures and exercise habits, underwent body composition exams, and wore CORTEX MetaMax 3B gas analyzers and ActiGraph wGT3X-BT accelerometers on five body sites. Ballroom dancers had two 10-minute dance sessions followed by a 10-minute rest. Aerobic dancers had one 20-minute session followed by a 10-minute rest. Square dancers had a 10-minute intense dance, a 10-minute soft dance, and a 10-minute rest. Participants rated perceived exertion using the RPE scale after 1-1.5 hours. After data cleaning, we obtained 311 physiological signal and 1555 acceleration count sequences.
RESULTS:
Empirical formulas inaccurately assessed dance energy expenditure, with MAPE exceeding 50% and RMSE surpassing 3.23. We developed a bidirectional stepwise regression model, achieving a 0.73 average goodness-of-fit, exploring optimal accelerometer sites for Effectiveness and Efficiency. Features extracted from raw data showed varying CCC: 0.37 (waist), 0.89 (left ankle), 0.94 (right ankle), 0.78 (left wrist), and 0.80 (right wrist). A random forest regression model minimized errors to 5% (MAPE) and 0.33 (RMSE) when incorporating all site data. Wrist accelerometers and heart rate sufficed for energy expenditure estimation, with RMSE values of 0.35 and 0.36, respectively, indicating a trade-off between effectiveness and efficiency. A neural network pipeline, based on extension, automatically assessed energy expenditure, yielding R-squared values of 0.842, F1-score of 0.709, and recall of 0.774.
CONCLUSION:
This study, to our knowledge, is the first to systematically assess energy expenditure in dance. Our contributions are twofold: introducing the Triple-E principle to regulate model aspects, and developing SOTA machine learning models for optimal sensor selection and accurate energy expenditure estimation.