Smartwatch-Based Prediction of Single-Stride and Stride-to-Stride Gait Outcomes Using Regression-Based Machine...

Area of Science:

• Biomechanics
• Wearable Technology
• Machine Learning

Background:

• Gait spatiotemporal variability is crucial for fall risk assessment.
• Wrist-worn sensors are preferred by users but underutilized for gait analysis.
• Existing methods often require sensors at other body locations.

Purpose of the Study:

• To develop and evaluate a smartwatch application for monitoring gait spatiotemporal parameters.
• To assess the feasibility of using a consumer-grade smartwatch's inertial measurement unit (IMU) for gait analysis.
• To compare different machine learning models for predicting gait outcomes from IMU data.

Main Methods:

• Trained machine learning models (linear, ridge, SVM, random forest, xGB) using IMU data from an Apple Watch Series 5.
• Collected gait data from 41 young adults walking at three different speeds on a treadmill.
• Validated model performance against an optoelectronic system for single-stride and spatiotemporal variability metrics.

Main Results:

• Extreme gradient boosting (xGB) models excelled in predicting single-stride gait outcomes (7-11% error, 0.60-0.86 ICC).
• Support vector machine (SVM) models performed best for spatiotemporal variability (18-22% error, 0.47-0.64 ICC).
• Models effectively captured speed-related changes in spatiotemporal gait parameters.

Conclusions:

• Smartwatch IMUs combined with machine learning are feasible for monitoring gait spatiotemporal parameters.
• This approach supports remote and accessible gait analysis, potentially aiding fall risk monitoring.
• Further research can refine these models for clinical applications and diverse populations.