Comparison of machine learning classifiers for differentiating level and sport using movement data.

The purposes of this study were to determine if 1) recurrent neural networks designed for multivariate, time-series analyses outperform traditional linear and non-linear machine learning classifiers when classifying athletes based on competition level and sport played, and 2) athletes of different sports move differently during non-sport-specific movement screens. Optical-based kinematic data from 542 athletes were used as input data for nine different machine learning algorithms to classify athletes based on competition level and sport played. For the traditional machine learning classifiers, principal component analysis and feature selection were used to reduce the data dimensionality and to determine the best principal components to retain. Across tasks, recurrent neural networks and linear machine learning classifiers tended to outperform the non-linear machine learning classifiers. For all tasks, reservoir computing took the least amount of time to train. Across tasks, reservoir computing had one of the highest classification rates and took the least amount of time to train; however, interpreting the results is more difficult compared to linear classifiers. In addition, athletes were successfully classified based on sport suggesting that athletes competing in different sports move differently during non-sport specific movements. Therefore, movement assessment screens should incorporate sport-specific scoring criteria.

Machine Learning and Deep Neural Network Architectures for 3D Motion Capture Datasets.

Biomechanical movement data are highly correlated multivariate time-series for which a variety of machine learning and deep neural network classification techniques are possible. For image classification, convolutional neural networks have reshaped the field, but have been challenging to apply to 3D movement data with its intrinsic multidimensional nonlinear correlations. Deep neural networks afford the opportunity to reduce feature engineering effort, remove model-based approximations that can introduce systematic errors, and reduce the manual data processing burden which is often a bottleneck in biomechanical data acquisition. What classification techniques are most appropriate for biomechanical movement data? Baseline performance for 3D joint centre trajectory classification using a number of traditional machine learning techniques are presented. Our framework and dataset support a robust comparison between classifier architectures over 416 athletes (professional, college, and amateur) from five primary and six non-primary sports performing thirteen non-sport-specific movements. A variety of deep neural networks specifically intended for time-series data are currently being evaluated.

Beneficial techniques for spatio-temporal imaging in electrical impedance tomography.

OBJECTIVE: Electrical impedance tomography (EIT) typically reconstructs individual images from electrical voltage measurements at pairs of electrodes due to current driven through other electrode pairs on a body. EIT images have low spatial resolution, but excellent temporal resolution. There are four methods for integrating temporal data into an EIT reconstruction: filtering over measurements, filtering over images, combined spatial and temporal (spatio-temporal) regularization, and Kalman filtering. These spatio-temporal methods have not been directly compared, making it difficult to evaluate relative performance and choose an appropriate method for particular use cases. APPROACH: We (i) develop a common framework, (ii) develop comparison metrics, (iii) perform simulation and tank studies which directly compare algorithms, and (iv) report on relative advantages of the different algorithms. MAIN RESULTS: Temporal filtering is well understood, but often not considered as part of the imaging process despite a direct impact on image reconstruction quality. Spatio-temporal regularized techniques are not yet efficient but offer tantalizing advantages. Kalman filtering enables adaptive filtering for time-varying measurement/image noise at the cost of often over-regularized (sub-optimal) images which can now be understood in the same framework as the other techniques. Further research into efficient implementations of Gauss-Newton spatio-temporal regularization will allow temporal and spatial covariance to be explicitly defined for longer time series (n > 10 frames) where temporal regularization can be more effective. For the immediate analysis of temporally varying images, we recommend the use of adaptive (time-varying) temporal filtering of measurements followed by adaptive spatial regularization (hyperparameter selection) as the most computationally efficient and effective approach currently available. SIGNIFICANCE: The analysis of variation within regions of an EIT image to extract physiological measures (functional imaging), has become an important EIT technique where temporal and spatial aspects of analysis are tightly integrated. This work gives guidance on available methods and suggests directions for future research.

Comparison of bolus- and filtering-based EIT measures of lung perfusion in an animal model.

OBJECTIVE: Two main functional imaging approaches have been used to measure regional lung perfusion using electrical impedance tomography (EIT): venous injection of a hypertonic saline contrast agent and imaging of its passage through the heart and lungs, and digital filtering of heart-frequency impedance changes over sequences of EIT images. This paper systematically compares filtering-based perfusion estimates and bolus injection methods to determine to which degree they are related. APPROACH: EIT data was recorded on seven mechanically ventilated newborn lambs in which ventilation distribution was varied through changes in posture between prone, supine, left- and right-lateral positions. Perfusion images were calculated using frequency filtering and ensemble averaging during both ventilation and apnoea time segments for each posture to compare against contrast agent-based methods using Jaccard distance score. MAIN RESULTS: Using bolus-based EIT measures of lung perfusion as the reference frequency filtering techniques performed better than ensemble averaging and both techniques performed equally well across apnoea and ventilation data segments. SIGNIFICANCE: Our results indicate the potential for use of filtering-based EIT measures of heart-frequency activity as a non-invasive proxy for contrast agent injection-based measures of lung perfusion.

Electrical Impedance Tomography: Tissue Properties to Image Measures.

Electrical impedance tomography (EIT) uses electrical stimulation and measurement at the body surface to image the electrical properties of internal tissues. It has the advantage of noninvasiveness and high temporal resolution but suffers from poor spatial resolution and sensitivity to electrode movement and contact quality. EIT can be useful to applications, where there are conductive contrasts between tissues, fluids, or gasses, such as imaging of cancerous or ischemic tissue or functional monitoring of breathing, blood flow, gastric motility, and neural activity. The past decade has seen clinical application and commercial activity using EIT for ventilation monitoring. Interpretation of EIT-based measures is complex, and this review paper focuses on describing the image interpretation "pathway." We review this pathway, from Tissue Electrical Properties, EIT Electrodes & Hardware, Sensitivity, Image Reconstruction, Image Processing to EIT Measures. The relationship is discussed between the clinically relevant parameters and the reconstructed properties. An overview is given of areas of EIT application and of our perspectives for research and development.Electrical impedance tomography (EIT) uses electrical stimulation and measurement at the body surface to image the electrical properties of internal tissues. It has the advantage of noninvasiveness and high temporal resolution but suffers from poor spatial resolution and sensitivity to electrode movement and contact quality. EIT can be useful to applications, where there are conductive contrasts between tissues, fluids, or gasses, such as imaging of cancerous or ischemic tissue or functional monitoring of breathing, blood flow, gastric motility, and neural activity. The past decade has seen clinical application and commercial activity using EIT for ventilation monitoring. Interpretation of EIT-based measures is complex, and this review paper focuses on describing the image interpretation "pathway." We review this pathway, from Tissue Electrical Properties, EIT Electrodes & Hardware, Sensitivity, Image Reconstruction, Image Processing to EIT Measures. The relationship is discussed between the clinically relevant parameters and the reconstructed properties. An overview is given of areas of EIT application and of our perspectives for research and development.

Shape deformation in two-dimensional electrical impedance tomography.

Electrical impedance tomography (EIT) uses measurements from surface electrodes to reconstruct an image of the conductivity of the contained medium. However, changes in measurements result from both changes in internal conductivity and changes in the shape of the medium relative to the electrode positions. Failure to account for shape changes results in a conductivity image with significant artifacts. Previous work to address shape changes in EIT has shown that in some cases boundary shape and electrode location can be uniquely determined for isotropic conductivities; however, for geometrically conformal changes, this is not possible. This prior work has shown that the shape change problem can be partially addressed. In this paper, we explore the limits of compensation for boundary movement in EIT using three approaches. First, a theoretical model was developed to separate a deformation vector field into conformal and nonconformal components, from which the reconstruction limits may be determined. Next, finite element models were used to simulate EIT measurements from a domain whose boundary has been deformed. Finally, an experimental phantom was constructed from which boundary deformation measurements were acquired. Results, both in simulation and with experimental data, suggest that some electrode movement and boundary distortions can be reconstructed based on conductivity changes alone while reducing image artifacts in the process.

Addressing the computational cost of large EIT solutions.

Electrical impedance tomography (EIT) is a soft field tomography modality based on the application of electric current to a body and measurement of voltages through electrodes at the boundary. The interior conductivity is reconstructed on a discrete representation of the domain using a finite-element method (FEM) mesh and a parametrization of that domain. The reconstruction requires a sequence of numerically intensive calculations. There is strong interest in reducing the cost of these calculations. An improvement in the compute time for current problems would encourage further exploration of computationally challenging problems such as the incorporation of time series data, wide-spread adoption of three-dimensional simulations and correlation of other modalities such as CT and ultrasound. Multicore processors offer an opportunity to reduce EIT computation times but may require some restructuring of the underlying algorithms to maximize the use of available resources. This work profiles two EIT software packages (EIDORS and NDRM) to experimentally determine where the computational costs arise in EIT as problems scale. Sparse matrix solvers, a key component for the FEM forward problem and sensitivity estimates in the inverse problem, are shown to take a considerable portion of the total compute time in these packages. A sparse matrix solver performance measurement tool, Meagre-Crowd, is developed to interface with a variety of solvers and compare their performance over a range of two- and three-dimensional problems of increasing node density. Results show that distributed sparse matrix solvers that operate on multiple cores are advantageous up to a limit that increases as the node density increases. We recommend a selection procedure to find a solver and hardware arrangement matched to the problem and provide guidance and tools to perform that selection.

The impact of electrode area, contact impedance and boundary shape on EIT images.

Electrical impedance tomography (EIT) measures the conductivity distribution within an object based on the current applied and voltage measured at surface electrodes. Thus, EIT images are sensitive to electrode properties (i.e. contact impedance, electrode area and boundary shape under the electrode). While some of these electrode properties have been investigated individually, this paper investigates these properties and their interaction using finite element method simulations and the complete electrode model (CEM). The effect of conformal deformations on image reconstruction when using the CEM was of specific interest. Observed artefacts were quantified using a measure that compared an ideal image to the reconstructed image, in this case a no-noise reconstruction that isolated the electrodes' effects. For electrode contact impedance and electrode area, uniform reductions to all electrodes resulted in ringing artefacts in the reconstructed images when the CEM was used, while parameter variations that were not correlated amongst electrodes resulted in artefacts distributed throughout the image. When the boundary shape changed under the electrode, as with non-symmetric conformal deformations, using the CEM resulted in structured distortions within the reconstructed image. Mean electrode contact impedance increases, independent of inter-electrode variation, did not result in artefacts in the reconstructed image.