Sprint Assessment Using Machine Learning and a Wearable Accelerometer.

Field-based sprint performance assessments rely on metrics derived from a simple model of sprinting dynamics parameterized by 2 constants, v0 and τ, which indicate a sprinter's maximal theoretical velocity and the time it takes to approach v0, respectively. This study aims to automate sprint assessment by estimating v0 and τ using machine learning and accelerometer data. To this end, photocells recorded 10-m split times of 28 subjects for three 40-m sprints while wearing an accelerometer around the waist. Features extracted from the accelerometer data were used to train a classifier to identify the sprint start and regression models to estimate the sprint model parameters. Estimates of v0, τ, and 30-m sprint time (t30) were compared between the proposed method and a photocell method using root mean square error and Bland-Altman analysis. The root mean square error of the sprint start estimate was .22 seconds and ranged from .52 to .93 m/s for v0, .14 to .17 seconds for τ, and .23 to .34 seconds for t30. Model-derived sprint performance metrics from most regression models were significantly (P < .01) correlated with t30. Comparison of the proposed method and a physics-based method suggests pursuit of a combined approach because their strengths appear to complement each other.