Smartwatch-based prediction of single-stride and stride-to-stride gait outcomes using regression-based machine learning

Abstract

Spatiotemporal variability during gait is linked to fall risk and could be monitored using wearable sensors. Although many users prefer wrist-worn sensors, most applications position at other sites. We developed and evaluated an application using a consumer-grade smartwatch inertial measurement unit (IMU). Young adults (N = 41) completed seven-minute conditions of treadmill gait at three different speeds. Single-stride outcomes (stride time, length, width, and speed) and spatiotemporal variability (coefficient of variation of each single-stride outcome) were recorded using an optoelectronic system, while 232 single- and multi-stride IMU metrics were recorded using an Apple Watch Series 5. These metrics were input to train linear, ridge, support vector machine (SVM), random forest, and extreme gradient boosting (xGB) models of each spatiotemporal outcome. We conducted Model x Condition ANOVAs to explore model sensitivity to speed-related responses. xGB models were best for single-stride outcomes (relative mean absolute error [% error]: 7 - 11%; intraclass correlation coefficient [ICC2,1]: 0.60 - 0.86) and SVM models were best for spatiotemporal variability (% error: 18 - 22%; ICC2,1 = 0.47 - 0.64). Spatiotemporal changes with speed were captured by these models (Condition: p < 0.00625). Results support the feasibility of monitoring multi-stride spatiotemporal parameters using a smartwatch IMU and machine learning.