Deep Learning Techniques in Estimating Ankle Joint Power Using Wearable IMUs

Abstract

Estimating ankle joint power can be used to identify gait abnormalities, which is usually achieved by employing a complicated biomechanical model using heavy equipment settings. This paper demonstrates deep learning approaches to estimate ankle joint power from two Inertial Measurement Unit (IMU) sensors attached at foot and shank. The purpose of this study was to investigate deep learning models in estimating ankle joint power in practical scenarios, in terms of variance in walking speeds, reduced number of extracted features and inter-subject model adaption. IMU data was collected from nine healthy participants during five walking trials at different speeds on a force-plate-instrumented treadmill while an optical motion tracker was used as ground truth. Three state-of-the-art deep neural architectures, namely Long Short-Term Memory (LSTM), Convolutional Neural Network (CNN) and, fusion of CNN and LSTM (CNN-LSTM), were developed, trained, and evaluated in predicting ankle joint power by extracting few simple, meaningful features. The proposed architectures were found efficient and promising with higher estimation accuracies (correlation coefficient, R > 0.92 and adjusted R-squared value > 83%) and lower errors (mean squared error <; 0.06, and mean absolute error <; 0.13) in inter-participant evaluations. Performance evaluations among the three deep regressors showed that LSTM performed comparatively better. Lower standard deviations in mean squared error (0.029) and adjusted R-squared value (5.5%) proved the proposed model's efficiency for all participants.