S. Hamid Nawab

High-yield decomposition of surface EMG signals

OBJECTIVE Automatic decomposition of surface electromyographic (sEMG) signals into their constituent motor unit action potential trains (MUAPTs). METHODS A small five-pin sensor provides four channels of sEMG signals that are in turn processed by an enhanced artificial intelligence algorithm evolved from a previous proof-of-principle. We tested the technology on sEMG signals from five muscles contracting isometrically at force levels ranging up to 100% of their maximal level, including those that were covered with more than 1.5cm of adipose tissue. Decomposition accuracy was measured by a new method wherein a signal is first decomposed and then reconstructed and the accuracy is measured by comparison. Results were confirmed by the more established two-source method. RESULTS The number of MUAPTs decomposed varied among muscles and force levels and mostly ranged from 20 to 30, and occasionally up to 40. The accuracy of all the firings of the MUAPTs was on average 92.5%, at times reaching 97%. CONCLUSIONS Reported technology can reliably perform high-yield decomposition of sEMG signals for isometric contractions up to maximal force levels. SIGNIFICANCE The small sensor size and the high yield and accuracy of the decomposition should render this technology useful for motor control studies and clinical investigations.

Approximate Signal Processing

It is increasingly important to structure signal processing algorithms and systems to allow for trading off between the accuracy of results and the utilization of resources in their implementation. In any particular context, there are typically a variety of heuristic approaches to managing these tradeoffs. One of the objectives of this paper is to suggest that there is the potential for developing a more formal approach, including utilizing current research in Computer Science on Approximate Processing and one of its central concepts, Incremental Refinement. Toward this end, we first summarize a number of ideas and approaches to approximate processing as currently being formulated in the computer science community. We then present four examples of signal processing algorithms/systems that are structured with these goals in mind. These examples may be viewed as partial inroads toward the ultimate objective of developing, within the context of signal processing design and implementation, a more general and rigorous framework for utilizing and expanding upon approximate processing concepts and methodologies.

Decomposition of indwelling EMG signals

Decomposition of indwelling electromyographic (EMG) signals is challenging in view of the complex and often unpredictable behaviors and interactions of the action potential trains of different motor units that constitute the indwelling EMG signal. These phenomena create a myriad of problem situations that a decomposition technique needs to address to attain completeness and accuracy levels required for various scientific and clinical applications. Starting with the maximum a posteriori probability classifier adapted from the original precision decomposition system (PD I) of LeFever and De Luca (25, 26), an artificial intelligence approach has been used to develop a multiclassifier system (PD II) for addressing some of the experimentally identified problem situations. On a database of indwelling EMG signals reflecting such conditions, the fully automatic PD II system is found to achieve a decomposition accuracy of 86.0% despite the fact that its results include low-amplitude action potential trains that are not decomposable at all via systems such as PD I. Accuracy was established by comparing the decompositions of indwelling EMG signals obtained from two sensors. At the end of the automatic PD II decomposition procedure, the accuracy may be enhanced to nearly 100% via an interactive editor, a particularly significant fact for the previously indecomposable trains.

IPUS An Architecture for the Integrated Processing and Understanding of Signals

The Integrated Processing and Understanding of Signals (IPUS) architecture is presented as a framework that exploits formal signal processing models to structure the bi-directional interaction between front-end signal processing and signal understanding processes. This architecture is appropriate for complex environments, which are characterized by variable signal to noise ratios, unpredictable source behaviors, and co-occuring objects whose signal signatures can distort each other. A key aspect of this architecture is that front-end signal processing is dynamically modifiable in response to scenario changes and to the need to re-analyze ambiguous or distorted data. This architecture tightly integrates the search for the appropriate front-end signal processing configuration with the search for plausible interpretations. In our opinion, this dual search, informed by formal signal processing theory, is a necessary component of perceptual systems for intelligent agents that must interact with complex environments. In order to explain this architecture in detail, an example of its use in an implemented system for acoustic signal interpretation is presented. ****************************************************************** This work was supported by the Rome Air Development Center of the Air Force Systems Command under contract F30602-91-C-0038, and by the Office of Naval Research under contract N00014-92-J-1450. The content does not necessarily reflect the position or the policy of the Government, and no official endorsement should be inferred. This report extends work on non-control aspects of IPUS reported in Technical Report 91-34. Please refer to that report for more information on control in IPUS.

Detecting freezing-of-gait during unscripted and unconstrained activity

We present a dynamic neural network (DNN) solution for detecting instances of freezing-of-gait (FoG) in Parkinsons disease (PD) patients while they perform unconstrained and unscripted activities. The input features to the DNN are derived from the outputs of three triaxial accelerometer (ACC) sensors and one surface electromyographic (EMG) sensor worn by the PD patient. The ACC sensors are placed on the shin and thigh of one leg and on one of the forearms while the EMG sensor is placed on the shin. Our FoG solution is architecturally distinct from the DNN solutions we have previously designed for detecting dyskinesia or tremor. However, all our DNN solutions utilize the same set of input features from each EMG or ACC sensor worn by the patient. When tested on experimental data from PD patients performing unconstrained and unscripted activities, our FoG detector exhibited 83% sensitivity and 97% specificity on a per-second basis.

High-Resolution Tracking of Motor Disorders in Parkinson's Disease During Unconstrained Activity

Parkinsons disease (PD) can present with a variety of motor disorders that fluctuate throughout the day, making assessment a challenging task. Paper‐based measurement tools can be burdensome to the patient and clinician and lack the temporal resolution needed to accurately and objectively track changes in motor symptom severity throughout the day. Wearable sensor‐based systems that continuously monitor PD motor disorders may help to solve this problem, although critical shortcomings persist in identifying multiple disorders at high temporal resolution during unconstrained activity. The purpose of this study was to advance the current state of the art by (1) introducing hybrid sensor technology to concurrently acquire surface electromyographic (sEMG) and accelerometer data during unconstrained activity and (2) analyzing the data using dynamic neural network algorithms to capture the evolving temporal characteristics of the sensor data and improve motor disorder recognition of tremor and dyskinesia. Algorithms were trained (n = 11 patients) and tested (n = 8 patients; n = 4 controls) to recognize tremor and dyskinesia at 1‐second resolution based on sensor data features and expert annotation of video recording during 4‐hour monitoring periods of unconstrained daily activity. The algorithms were able to make accurate distinctions between tremor, dyskinesia, and normal movement despite the presence of diverse voluntary activity. Motor disorder severity classifications averaged 94.9% sensitivity and 97.1% specificity based on 1 sensor per symptomatic limb. These initial findings indicate that new sensor technology and software algorithms can be effective in enhancing wearable sensor‐based system performance for monitoring PD motor disorders during unconstrained activities.

A C++ software environment for the development of embedded signal processing systems

A new environment for the rapid development of embedded signal processing software is described. The environment encourages incremental design via modular and hierarchical structuring of applications, and additional features are included which support the prototyping, testing, implementation, and integration stages of the system design cycle. Written in C++, the environment is comprised of a scripting language for the definition of system components and a class library which includes a basic application framework. Support is provided for incorporating both numeric and symbolic signal representations, as well as integrating multiple signal processing techniques within a single application. A sophisticated control mechanism allows dynamic scheduling of signal processing operations according to algorithmically defined schema. Signal processing applications developed in this environment are themselves objects, and are suitable for embedding within a larger overall system.

Improved resolution of pulse superpositions in a knowledge-based system EMG decomposition

We have improved the accuracy (sensitivity x specificity) of a knowledge-based system from 90% to well above 95% in decomposing complex EMG 3-channel data into its constituent motor unit action potential (MUAP) trains. The key to achieving this improvement is our use of a probabilistic framework for resolving pulse superpositions through the application of utility maximization at the suprasegmental level.

Dynamic neural network detection of tremor and dyskinesia from wearable sensor data

We present a dynamic neural network (DNN) solution for detecting time-varying occurrences of tremor and dyskinesia at 1 s resolution from time series data acquired from surface electromyographic (sEMG) sensors and tri-axial accelerometers worn by patients with Parkinsons disease (PD). The networks were trained and tested on separate datasets, each containing approximately equal proportions of tremor, dyskinesia, and disorder-free data from 8 PD and 4 control subjects performing unscripted and unconstrained activities in an apartment-like environment. During DNN testing, tremor was detected with a sensitivity of 93% and a specificity of 95%, while dyskinesia was detected with a sensitivity of 91% and a specificity of 93%. Similar sensitivity and specificity levels were obtained when DNN testing was carried out on subjects who were not included in DNN training.

Dynamical Learning and Tracking of Tremor and Dyskinesia From Wearable Sensors

We have developed and evaluated several dynamical machine-learning algorithms that were designed to track the presence and severity of tremor and dyskinesia with 1-s resolution by analyzing signals collected from Parkinsons disease (PD) patients wearing small numbers of hybrid sensors with both 3-D accelerometeric and surface-electromyographic modalities. We tested the algorithms on a 44-h signal database built from hybrid sensors worn by eight PD patients and four healthy subjects who carried out unscripted and unconstrained activities of daily living in an apartment-like environment. Comparison of the performance of our machine-learning algorithms against independent clinical annotations of disorder presence and severity demonstrates that, despite their differing approaches to dynamic pattern classification, dynamic neural networks, dynamic support vector machines, and hidden Markov models were equally effective in keeping error rates of the dynamic tracking well below 10%. A common set of experimentally derived signal features were used to train the algorithm without the need for subject-specific learning. We also found that error rates below 10% are achievable even when our algorithms are tested on data from a sensor location that is different from those used in algorithm training.