IMU-Based Deep Neural Networks: Prediction of Locomotor and Transition Intentions of an Osseointegrated Transfemoral Amputee

Abstract

This paper focuses on the design and comparison of different deep neural networks for the real-time prediction of locomotor and transition intentions of one osseointegrated transfemoral amputee using only data from inertial measurement units. The deep neural networks are based on convolutional neural networks, recurrent neural networks, and convolutional recurrent neural networks. The architectures' input are features in both the time domain and the time-frequency domain, which are derived from either one inertial measurement unit (placed above the prosthetic knee) or two inertial measurement units (placed above and below the prosthetic knee). The prediction of eight different locomotion modes (i.e., sitting, standing, level ground walking, stair ascent and descent, ramp ascent and descent, walking on uneven terrain) and the twenty-four transitions among them is investigated. The study shows that a recurrent neural network, realized with four layers of gated recurrent unit networks, achieves (with a 5-fold cross-validation) a mean F1 score of 84.78% and 86.50% using one inertial measurement unit, and 93.06% and 89.99% using two inertial measurement units, with or without sitting, respectively.

Learning-Based Motion-Intention Prediction for End-Point Control of Upper-Limb-Assistive Robots

“…Three different RNN models were designed in this study, including RNN, long short-term memory (LSTM) networks, and gated recurrent unit (GRU) networks [ 40 ]. All the RNNs consisted of three hidden layers, including two recurrent layers and one fully-connected linear layer.…”

Hybrid Learning Models for IMU-Based HAR with Feature Analysis and Data Correction

“…Deep learning networks have powerful non-linear representation learning capabilities and can automatically capture data features required for the classification of complex actions [ 35 ]. The authors [ 36 ] feed the characteristics of the time domain of inertial signals into nine different neural network models to complete the action classification and prediction of subjects. The model is based on CNN, RNN, and convolutional recurrent neural network (CRNN) architectures.…”

“…Comparing the existing HAR systems to an IMU node, the proposed system simulates the real-world environment as much as possible, which does not need to regulate the subject to complete the specified action within the specified time and limit of its action trajectory. Compared to existing AI-based HAR studies (e.g., SVM [ 19 , 26 ], DT [ 19 ], KNN [ 26 ], RF [ 26 ], HMM [ 27 ], AdaBoost [ 28 ], LSTM [ 36 , 39 ], and CNN [ 37 , 38 , 39 ]), we implement up to 19 classifiers and 6 generators and combine the two hyperparameter optimization methods of GS and RS to find the most suitable model. At the same time, we use feature engineering with two different concepts (i.e., selecting a subset of features with the most correlation with the output and the least correlation among these corresponding features) to reduce the feature dimension, find the ideal input features, further reduce the size of the classifier, and improve the time and accuracy of classification operations, providing a comprehensive study.…”

Biomechanics of Ascending and Descending Stairs in a Patient with Transfemoral Amputation and Neural Sensory Feedback: A Case Report

“…Gait impairment resulting from amputation has been objectively documented across various domains, including spatiotemporal and biomechanical parameters [10,11], as well as bioenergetics parameters. Individuals with transfemoral amputation, particularly those with dysvascular morbidity as the underlying cause, walk slower by 40% than normal, consume 2.5 times more energy [12], have increased oxygen consumption by about 20% compared to a healthy person [13], and face limitations in walking longer distances outdoors [14].…”