The performance of a machine learning model in predicting accelerometer-derived walking speed

Area of Science:

• Biomechanics
• Machine Learning
• Wearable Technology

Background:

• Accurate long-term measurement of walking speed in large studies is difficult.
• Walking speed is a key indicator of health and mobility.
• Current methods for measuring walking speed are often impractical for large-scale, long-term studies.

Purpose of the Study:

• To develop and evaluate a machine learning classifier for predicting walking speeds.
• To assess the performance of the classifier using dual and single accelerometer setups.
• To determine the accuracy of classifying slow, moderate, and brisk walking speeds.

Main Methods:

• Trained an eXtreme Gradient Boosting (XGBoost) machine learning classifier.
• Used data from 24 adults with tri-axial accelerometers on the thigh and low back.
• Validated the classifier using leave-one-out cross-validation with 1, 3, and 5-second windows.

Main Results:

• The machine learning classifier achieved high accuracy in predicting walking speeds (slow, moderate, brisk) and jogging.
• Performance was comparable between dual and single accelerometer setups and across different window lengths.
• Highest accuracy reached 91% with a dual accelerometer setup and a 5-second window.

Conclusions:

• A machine learning classifier can accurately predict walking speeds using accelerometer data.
• Both dual and single accelerometer setups are effective for this prediction.
• This approach offers a feasible method for assessing walking speed in large-scale studies.