Federated learning for predicting clinical outcomes in patients with COVID-19.

Federated learning (FL) is a method used for training artificial intelligence models with data from multiple sources while maintaining data anonymity, thus removing many barriers to data sharing. Here we used data from 20 institutes across the globe to train a FL model, called EXAM (electronic medical record (EMR) chest X-ray AI model), that predicts the future oxygen requirements of symptomatic patients with COVID-19 using inputs of vital signs, laboratory data and chest X-rays. EXAM achieved an average area under the curve (AUC) >0.92 for predicting outcomes at 24 and 72 h from the time of initial presentation to the emergency room, and it provided 16% improvement in average AUC measured across all participating sites and an average increase in generalizability of 38% when compared with models trained at a single site using that site's data. For prediction of mechanical ventilation treatment or death at 24 h at the largest independent test site, EXAM achieved a sensitivity of 0.950 and specificity of 0.882. In this study, FL facilitated rapid data science collaboration without data exchange and generated a model that generalized across heterogeneous, unharmonized datasets for prediction of clinical outcomes in patients with COVID-19, setting the stage for the broader use of FL in healthcare.

Federated Learning used for predicting outcomes in SARS-COV-2 patients.

'Federated Learning' (FL) is a method to train Artificial Intelligence (AI) models with data from multiple sources while maintaining anonymity of the data thus removing many barriers to data sharing. During the SARS-COV-2 pandemic, 20 institutes collaborated on a healthcare FL study to predict future oxygen requirements of infected patients using inputs of vital signs, laboratory data, and chest x-rays, constituting the "EXAM" (EMR CXR AI Model) model. EXAM achieved an average Area Under the Curve (AUC) of over 0.92, an average improvement of 16%, and a 38% increase in generalisability over local models. The FL paradigm was successfully applied to facilitate a rapid data science collaboration without data exchange, resulting in a model that generalised across heterogeneous, unharmonized datasets. This provided the broader healthcare community with a validated model to respond to COVID-19 challenges, as well as set the stage for broader use of FL in healthcare.

Motor Unit Potential Jitter: A New Measure of Neuromuscular Transmission Instability.

A new measure of neuromuscular transmission instability, motor unit potential (MUP) jitter, is introduced. MUP jitter can be estimated quickly using MUP trains (MUPTs) extracted from electromyographic (EMG) signals acquired using conventional clinical equipment and needle EMG electrodiagnostic protocols. The primary motivation for developing MUP jitter is to avoid the technical demands associated with estimating jitter using conventional single fiber EMG techniques. At the core of the MUP jitter measure is a classifier capable of labeling a set of aligned MUP segments as single fiber MUP segments, i.e., parts of MUPs generated predominantly by a single fiber and not significantly contaminated by contributions from other fibers. For a set of MUPs generated by the same MU, these segments will have varying occurrence times within the MUPs, but will have consistent morphology across the MUPs. Pairs of sets of single fiber MUP segments generated by different fibers of the same MU and tracked across a MUPT can be used to estimate neuromuscular transmission instability. Aligning MUP segments is achieved using dynamic time warping. Results based on 680 simulated MUPTs show that MUP jitter can be estimated with an average error rate as low as 8.9%. Also, one or more sets of single fiber MUP segments can be detected in 85.3% of the studied trains. The analysis for a single MUPT can be completed in 3.6 s on average using a conventional personal computer.

Extraction and Classification of Multichannel Electromyographic Activation Trajectories for Hand Movement Recognition.

This paper proposes a system for hand movement recognition using multichannel electromyographic (EMG) signals obtained from the forearm surface. This system can be used to control prostheses or to provide inputs for a wide range of human computer interface systems. In this work, the hand movement recognition problem is formulated as a multi-class distance based classification of multi-dimensional sequences. More specifically, the extraction of multi-channel EMG activation trajectories underlying hand movements, and classifying the extracted trajectories using a metric based on multi-dimensional dynamic time warping are investigated. The developed methods are evaluated using the publicly available NINAPro database comprised of 40 different hand movements performed by 40 subjects. The average movement error rate obtained across the 40 subjects is 0.09±0.047. The low error rate demonstrates the efficacy of the proposed trajectory extraction method and the discriminability of the utilized distance metric.

Transparent muscle characterization using quantitative electromyography: different binarization mappings.

Evaluation of patients with suspected neuromuscular disorders is typically based on qualitative visual and auditory assessment of needle detected eletromyographic (EMG) signals; the resulting muscle characterization is subjective and highly dependent on the skill and experience of the examiner. Quantitative electromyography (QEMG) techniques were developed to extract motor unit potential trains (MUPTs) from needle detected EMG signals, and estimate features capturing motor unit potential (MUP) morphology and quantifying morphological consistency across MUPs belonging to the same MUPT. The aim of this study is to improve available methods for obtaining transparent muscle characterizations from features obtained using QEMG techniques. More specifically, we investigate the following. 1) Can the use of binarization mappings improve muscle categorization accuracies of transparent methods? 2) What are the appropriate binarization mappings in terms of accuracy and transparency? Results from four different sets of examined limb muscles (342 muscles in total) demonstrate that four out of the 10 investigated binarization mappings based on transparent characterization methods outperformed the multi-class characterizers based on Gaussian mixture models (GMM) and the corresponding binarization mappings based on GMM. This suggests that the use of an appropriate binarization mapping can overcome the decrease in categorization accuracy associated with quantizing MUPT features, which is necessary to obtain transparent characterizations. This performance gain can be attributed to the use of more relevant features and tuned quantization to obtain more specific binary characterizations.

Feature selection for motor unit potential train characterization.

INTRODUCTION: Ten new features of motor unit potential (MUP) morphology and stability are proposed. These new features, along with 8 traditional features, are grouped into 5 aspects: size, shape, global complexity, local complexity, and stability. METHODS: We used sequential forward and backward search strategies to select subsets of these 18 features to discriminate accurately between muscles whose MUPs are predominantly neurogenic, myopathic, or normal. RESULTS AND CONCLUSIONS: Results based on 8102 motor unit potential trains (MUPTs) extracted from 4 different limb muscles (n = 336 total muscles) demonstrate the usefulness of these newly introduced features and support an aspect-based grouping of MUPT features.

Detecting neuropathy using measures of motor unit activation extracted from standard concentric needle electromyographic signals.

OBJECTIVE: Motor unit loss associated with neuropathic disorders affects motor unit activation. Quantitative electromyographic (EMG) features of motor unit activation estimated from the sequences of motor unit potentials (MUPs) created by concurrently active motor units can support the detection of neuropathic disorders. Interpretation of most motor unit activation feature values are, however, confounded by uncertainty regarding the level of muscle activation during EMG signal detection. A set of new features circumventing these limitations are proposed, and their utility in detecting neuropathy is investigated using simulated and clinical EMG signals. METHODS: The firing sequence of a motor neuron was simulated using a compartmentalized Hodgkin-Huxley based model. A pool of motor neurons was modelled such that each motor neuron was subjected to a common level of activation. The detection of the firing sequence of a motor neuron using a clinically detected EMG signal was simulated using a model of muscle anatomy combined with a model representing muscle fiber electrophysiology and the voltage detection properties of a concentric needle electrode. SIGNIFICANCE: Findings are based on simulated EMG data representing 30 normal and 30 neuropathic muscles as well as clinical EMG data collected from the tibialis anterior muscle of 48 control subjects and 30 subjects with neuropathic disorders. These results demonstrate the possibility of detecting neuropathy using motor unit recruitment and mean firing rate feature values estimated from standard concentric needle detected EMG signals.

Multifunction myoelectric control using multi-dimensional dynamic time warping.

Myoelectric control can be used for a variety of applications including powered protheses and different human computer interface systems. The aim of this study is to investigate the formulation of myoelectric control as a multi-class distance-based classification of multidimensional sequences. More specifically, we investigate (1) estimation of multi-muscle activation sequences from multi-channel electromyographic signals in an online manner, and (2) classification using a distance metric based on multi-dimensional dynamic time warping. Subject-specific results across 5 subjects executing 10 different hand movements showed an accuracy of 95% using offline extracted trajectories and an accuracy of 84% using online extracted trajectories.