Determining jumping performance from a single body-worn accelerometer using machine learning.

Area of Science:

• Sports Science
• Biomechanics
• Machine Learning

Background:

• External peak power in countermovement jumps is crucial for athlete training monitoring.
• Force platforms are the gold standard but impractical for field testing.
• Existing alternatives using flight time or inertial sensors lack sufficient accuracy.

Purpose of the Study:

• To develop and validate a machine learning model for accurate estimation of countermovement jump peak power using a single body-worn accelerometer.
• To investigate the impact of different features, sensor locations, and machine learning models on accuracy.

Main Methods:

• Collected data from 69 athletes performing 696 countermovement jumps (with/without arm swing).
• Extracted characteristic features (functional principal components) from a single body-worn accelerometer.
• Utilized a support vector machine model with a novel surrogate model optimization procedure.

Main Results:

• The optimized support vector machine model using the resultant signal from a lower back sensor achieved a peak power RMSE of 2.3 W·kg⁻¹ (5.1% of mean).
• Nested cross-validation and an independent holdout test confirmed the model's accuracy (2.0 W·kg⁻¹).
• The developed model demonstrated lower error than previous studies but is not yet field-deployable.

Conclusions:

• Machine learning models utilizing functional data representations can effectively reduce complexity for time-aligned signals.
• The novel optimization procedure is robust and applicable to other low-cost, noisy objective functions.
• While promising, the current model requires further refinement for practical field-based athlete monitoring.