Recurrent Neural Network Methods for Extracting Dynamic Balance Variables during Gait from a Single Inertial...

Area of Science:

• Biomechanics
• Wearable technology
• Machine learning

Background:

• Dynamic balance monitoring is crucial for preventing falls in the elderly population.
• Current methods may be limited in real-world applicability for continuous gait analysis.
• Inertial Measurement Units (IMUs) offer a potential solution for unobtrusive balance assessment.

Purpose of the Study:

• To develop and evaluate recurrent neural network (RNN) models for extracting gait balance variables from IMU data.
• To compare the performance of different RNN architectures (LSTM, GRU) in predicting balance parameters.
• To identify the most accurate and efficient model for real-life gait balance monitoring.

Main Methods:

• Thirteen healthy young and thirteen healthy older adults participated in the study.
• A single IMU was worn on the sacrum during walking to collect data.
• Inclination Angles (IA) and their Rates of Change (RCIA) were measured as ground truth.
• Four RNN models (uni-LSTM, bi-LSTM, uni-GRU, bi-GRU) were trained and evaluated using root-mean-squared errors (RMSEs) and percentage relative RMSEs (rRMSEs).

Main Results:

• The bidirectional Gated Recurrent Unit (bi-GRU) model with a weighted Mean Squared Error (MSE) demonstrated the highest prediction accuracy.
• The bi-GRU model exhibited superior computational efficiency compared to other models.
• The model effectively identified statistical differences in gait balance between young and older adults.

Conclusions:

• The bi-GRU model is the optimal choice for prolonged, real-life monitoring of gait balance.
• This technology can significantly contribute to fall risk management strategies in the elderly.
• Accurate dynamic balance assessment using IMUs and RNNs holds promise for proactive healthcare interventions.