Estimating Biomechanical Time-Series with Wearable Sensors: A Systematic Review of Machine Learning Techniques

Abstract

Wearable sensors have the potential to enable comprehensive patient characterization and optimized clinical intervention. Critical to realizing this vision is accurate estimation of biomechanical time-series in daily-life, including joint, segment, and muscle kinetics and kinematics, from wearable sensor data. The use of physical models for estimation of these quantities often requires many wearable devices making practical implementation more difficult. However, regression techniques may provide a viable alternative by allowing the use of a reduced number of sensors for estimating biomechanical time-series. Herein, we review 46 articles that used regression algorithms to estimate joint, segment, and muscle kinematics and kinetics. We present a high-level comparison of the many different techniques identified and discuss the implications of our findings concerning practical implementation and further improving estimation accuracy. In particular, we found that several studies report the incorporation of domain knowledge often yielded superior performance. Further, most models were trained on small datasets in which case nonparametric regression often performed best. No models were open-sourced, and most were subject-specific and not validated on impaired populations. Future research should focus on developing open-source algorithms using complementary physics-based and machine learning techniques that are validated in clinically impaired populations. This approach may further improve estimation performance and reduce barriers to clinical adoption.

A Comparison of Three Neural Network Approaches for Estimating Joint Angles and Moments from Inertial Measurement Units

“…Inertial measurement unit (IMU) sensors are gaining traction as a clinical gait analysis tool due to their improved accuracy, feasibility, ease-of-use, and importantly, applicability outside of the laboratory environment [ 1 , 2 ]. A meta-analysis of inertial sensor based gait analysis research concluded limited evidence in their application for determining joint kinematics, especially in non-sagittal (flexion-extension) motion [ 3 ].…”

“…A variety of methods have been proposed for this purpose, and reviewed by Ancillao et al [ 9 ]. To overcome accuracy limitations and the restricted subsets of parameters that can be determined, researchers have focused on applying machine learning methods to improve the prediction of GRFs, joint angles and joint moments [ 2 , 10 , 11 , 12 , 13 , 14 , 15 ], with initial efforts focused on predicting smaller subsets of data, such as single GRF and joint moment components [ 10 , 11 , 12 ], or in the case of Stetter et al [ 13 ] by predicting sagittal and frontal plane moments in isolation. Very recently gait researchers have trained machine learning models to predict all component joint angles [ 14 , 15 ] and moments across all lower limb joints [ 14 ].…”

“…Hence, they are also convenient for tasks involving the prediction of motion sequences. Unfortunately, LSTM networks are intensive to train and require large datasets, something rarely available in biomechanics [ 2 ]. Using inertial sensor data as inputs, LSTM networks have also been previously used to predict joint angles and moments [ 15 , 17 , 20 ].…”

Artificial Neural Networks in Motion Analysis—Applications of Unsupervised and Heuristic Feature Selection Techniques

“…This approach is for now limited to the determination of the 2D joint angles [ 13 , 14 ]. The use of machine learning, in particular, artificial neural networks, is gaining more and more relevance in biomechanical time-series estimation and has recently been reviewed by Gurchiek et al [ 15 ]. They found hybrid approaches incorporating domain knowledge by feature selection into the model to be helpful for accurate predictions.…”

Wearables-Only Analysis of Muscle and Joint Mechanics: An EMG-Driven Approach

“…While physics-based techniques exist for estimating these clinically relevant variables from wearables [20], [21], they require complex sensor arrays that discourage use outside of research contexts [22]. Regression algorithms have been proposed to reduce the number of required sensors [23], but at…”