Predicting Running Vertical Ground Reaction Forces Using Neural Network Models Based on an IMU Sensor

Area of Science:

• Biomechanics of running
• Sports injury prevention
• Wearable sensor technology

Background:

• Vertical ground reaction force (vGRF) is crucial for understanding running-related injuries (RRIs).
• Accurate vGRF measurement typically requires instrumented treadmills or force plates.
• Inertial measurement units (IMUs) offer a portable alternative but require synchronization with vGRF data.

Purpose of the Study:

• To develop and validate a synchronization method (STWS) for IMU and vGRF data during running.
• To create and assess artificial neural network (ANN) models (WNN and FFNN) for predicting vGRF from IMU acceleration data.
• To evaluate the accuracy of predicting peak vGRF and overall vGRF curves.

Main Methods:

• Fifteen runners (rearfoot and forefoot strikers) ran at 12, 14, and 16 km/h.
• A single IMU and an instrumented treadmill collected acceleration and vGRF data.
• The sliding time window synchronization (STWS) algorithm was used to align IMU and vGRF data. Wavelet neural network (WNN) and feed-forward neural network (FFNN) models predicted vGRF using acceleration data from the stance phase.

Main Results:

• The STWS algorithm achieved stride time synchronization with mean absolute errors < 11.2 ms.
• ANN models showed high correlation (r > 0.97) and agreement (NRMSE 4.6–9.2%, R² 0.93–0.99) between measured and predicted vGRF curves.
• Peak vGRF prediction had NRMSEs of 1.6–8.2%, with sagittal-axis acceleration data showing good accuracy and reduced input.

Conclusions:

• The STWS algorithm effectively synchronizes IMU and force plate data during running.
• Both WNN and FFNN models accurately predict vGRF, with potential for using sagittal-axis acceleration alone.
• This research provides a foundation for developing personalized ANN models for monitoring lower limb load in runners.