Author(s): ANDERSON-CONWAY, L., KOKOSZKA, A., ZHOU, C., SOKOLOFF, M., BROOKE, M., NASHIKKAR, P., PRIDDLE, B., STONE, B., VICKERS, D., PANTELOPOULOS, A., Institution: GOOGLE LLC, Country: UNITED STATES, Abstract-ID: 2553

INTRODUCTION:
Resistance training (RT) is a cornerstone of metabolic health and neuromuscular longevity [Westcott, W. L. (2012)]. However, unlike aerobic conditioning, which benefits from standardized metrics (heart rate, VO2max etc.), RT lacks accessible, objective quantification for load and fatigue. While Velocity-Based Training (VBT) serves as the current non-invasive gold standard for monitoring neuromuscular fatigue [Sánchez-Medina & González-Badillo, 2011], its reliance on specialized linear position transducers limits its utility to elite settings. Conversely, Rating of Perceived Exertion (RPE) remains the most common field method but is frequently compromised by subjective bias and poor intra-individual reliability [Zourdos et al., 2016].
METHODS:
This study proposes a novel framework for detecting muscle fatigue using consumer-grade, wrist-worn devices. We integrated tri-axial accelerometer, gyroscope, and barometric altimeter data to capture high-resolution kinematic signatures across common compound lifts (squats, deadlifts, bench press etc) and auxiliary lifts (biceps curl, tricep extension, shoulder press etc). In total, 13 exercises were used that were labelled by the proctors during the data collection. Data were collected from 334 participants who performed a total of 2104 sets of various RPEs, simultaneously monitored via VBT-gold standard equipment (Speed4lifts S.L., Madrid, Spain and Gymaware, Braddon, Australia). Furthermore, human labelers were employed to provide labels on sets, reps and rep phases from video data. We utilized a comparative machine learning approach, evaluating Random Forest, Gradient Boosting, and Neural Network architectures to classify "severely fatigued" (defined as >35% velocity loss) vs “low to mid fatigue” sets.
RESULTS:
Our analysis found a weak correlation (0.21) between self-reported RPE and objective VBT-defined fatigue, indicating RPE is a poor metric for muscle fatigue. Conversely, by leveraging multi-modal sensor data from consumer wearables, our trained model achieved moderate classification accuracy of 72% for severely fatigued sets.
CONCLUSION:
These results suggest that integrated kinematic sensors in consumer wearables can provide an objective, "lab-free" proxy for VBT. By providing near-to-real-time feedback on fatigue and consistency, this technology offers a scalable solution to optimize training volume, minimize injury risk, ensure high-quality repetitions and enhance metabolic outcomes in the general population.