Dawn MacIsaac

The short-time Fourier transform and muscle fatigue assessment in dynamic contractions.

The mean frequency of the power spectrum of an electromyographic signal is an accepted index for monitoring fatigue in static contractions. There is however, indication that it may be a useful index even in dynamic contractions in which muscle length and/or force may vary. The objective of this investigation was to explore this possibility. An examination of the effects of amplitude modulation on modeled electromyographic signals revealed that changes in variance created in this way do not sufficiently affect characteristic frequency data to obscure a trend with fatigue. This validated the contention that not all non-stationarities in signals necessarily manifest in power spectral parameters. While an investigation of the nature and effects of non-stationarities in real electromyographic signals produced from dynamic contractions indicated that a more complex model is warranted, the results also indicated that averaging associated with estimating spectral parameters with the short-time Fourier transform can control the effects of the more complex non-stationarities. Finally, a fatigue test involving dynamic contractions at a force level under 30% of peak voluntary dynamic range, validated that it was possible to track fatigue in dynamic contractions using a traditional short-time Fourier transform methodology.

Fatigue estimation with a multivariable myoelectric mapping function

A novel approach to muscle fatigue assessment is proposed. A function is used to map multiple myoelectric parameters representing segments of myoelectric data to a fatigue estimate for that segment. An artificial neural network is used to tune the mapping function and time-domain features are used as inputs. Two fatigue tests were conducted on five participants in each of static, cyclic and random conditions. The function was tuned with one data set and tested on the other. Performance was evaluated based on a signal to noise metric which compared variability due to fatigue factors with variability due to nonfatiguing factors. Signal to noise ratios for the mapping function ranged from 7.89 under random conditions to 9.69 under static conditions compared to 3.34-6.74 for mean frequency and 2.12-2.63 for instantaneous mean frequency indicating that the mapping function tracks the myoelectric manifestations of fatigue better than either mean frequency or instantaneous mean frequency under all three contraction conditions.

Automated Biosignal Quality Analysis for Electromyography Using a One-Class Support Vector Machine

This paper introduces the importance of biosignal quality assessment and presents a pattern classification approach to differentiate clean from contaminated electromyography (EMG) signals. Alternatively to traditional bottom-up approaches, which examine specific contaminants only, we present a top-down approach using a one-class support vector machine (SVM) trained on clean EMG and tested on artificially contaminated EMG. Both simulated and real EMG are used. Results are evaluated for each contaminant: 1) power line interference; 2) motion artifact; 3) ECG interference; 4) quantization noise; 5) analog-to-digital converter clipping; and 6) amplifier saturation, as a function of the level of signal contamination. Results show that different ranges of contamination can be detected in the EMG depending on the type of contaminant. At high levels of contamination, the SVM classifies all EMG signals as contaminated, whereas at low levels of contamination, it classifies the majority of EMG signals as contaminant free. A transition point for each contaminant is identified, where the classification accuracy drops and variance in classification increases. In some cases, contamination can be detected with the SVM when it is not visually discernible. This method is shown to be successful in detecting problems due to single contaminants but is generic to all forms of contamination in EMG.

EMG-based muscle fatigue assessment during dynamic contractions using principal component analysis

A novel approach to fatigue assessment during dynamic contractions was proposed which projected multiple surface myoelectric parameters onto the vector connecting the temporal start and end points in feature-space in order to extract the long-term trend information. The proposed end to end (ETE) projection was compared to traditional principal component analysis (PCA) as well as neural-network implementations of linear (LPCA) and non-linear PCA (NLPCA). Nine healthy participants completed two repetitions of fatigue tests during isometric, cyclic and random fatiguing contractions of the biceps brachii. The fatigue assessments were evaluated in terms of a modified sensitivity to variability ratio (SVR) and each method used a set of time-domain and frequency-domain features which maximized the SVR. It was shown that there was no statistical difference among ETE, PCA and LPCA (p>0.99) and that all three outperformed NLPCA (p<0.0022). Future work will include a broader comparison of these methods to other new and established fatigue indices.

Removal of electrocardiogram artifacts in surface electromyography using a moving average method

This paper presents a moving average method for estimating and removing electrocardiogram (ECG) artifact in surface electromyography (sEMG) recordings. This method does not require an ECG-only recording (e.g., with muscles relaxed), which is often required by other methods. The moving average method is compared to a common template subtraction method using sEMG recordings that were contaminated by adding ECG recordings. The performance of the moving average method is comparable to the template subtraction method. It provides superior performance at low signal-to-noise ratios (SNR) and is less sensitive to SNR.

Training a multivariable myoelectric mapping function to estimate fatigue

The mapping index (MI) is a fatigue assessment index that uses multiple time-domain myoelectric features to train an artificial neural network (ANN) to track the progression of fatigue. This work showed that mapping functions trained using data from independent subjects and contraction conditions to yield a generalized mapping index (GMI) can assess fatigue as well as functions trained with subject and contraction-specific data to yield MI. Surface myoelectric signals were collected from nine healthy participants during isometric, cyclic and random fatiguing contractions. Two datasets were collected: one for tuning the functions and the other for testing. The performance of fatigue indices was evaluated using a newly proposed piece-wise linear signal to noise ratio. ANN based indices were compared to normalized spectral moments (NSM) and mean frequency (MF). GMI performed as well as MI and outperformed NSM and MF demonstrating that subject and contraction-specific baseline data is not needed in order to train a mapping function which can effectively assess fatigue.

Machine Learning for Classifying Learning Objects

Building an ontology for learning objects can be useful for translating such objects between learning contexts. Such translations are important because they afford learners and educators with the opportunity to a survey a wide selection of learning and teaching material. For instance, university instructors are sometimes required to assess curriculum from courses delivered from other programs or universities, even internationally. Often, the only learning object available to do so is the course outline made available in HTML format on a Web page. Generally there is an abundance of metadata available from such learning objects and this information can be used to generate useful components of the ontology. Other useful information can be derived from first establishing the domain of the object, electricity and computing for instance, or possibly history. Once extracted, the information representing learning objects can be stored as elements in an XML template. The purpose of this work was to develop and implement a machine learning strategy for classifying course outlines into pre-defined domains and sub-domains in order to provide this information to an ontology repository designed to aid in the translation of such objects. First some typical domains were identified. Then, 20-30 course outlines were chosen to represent each sub-domain. Next, frequency tables of words common to the course outlines for a given sub-domain were generated in order to compile an ordered list of synonyms used to represent the sub-domains. Finally, a new set of course outlines were randomly selected for classification based on an analysis of the synonym content of each. Establishing the frequency tables and completing the synonym analysis was automated completely thereby constituting the machine learning strategy

Detection of ADC clipping, quantization noise, and amplifier saturation in surface electromyography

This paper focuses on the detection and quantification of three types of noise, analog-to-digital converter (ADC) clipping, quantization noise, and amplifier saturation, in surface electromyography (sEMG) without prior information regarding the sEMG setup. ADC clipping can be detected by searching for consecutive minimum and maximum values in a signal. Quantization noise can be expressed as a signal-to-quantization noise ratio which is estimated from the smallest observable step size in the signal. Amplifier saturation is quantified using a normality test, as amplifier saturation will reduce the normality of the signal amplitude distribution. Experimental results, using simulated sEMG, demonstrate the successful application of these methods.

CleanEMG — Power line interference estimation in sEMG using an adaptive least squares algorithm

This paper presents an adaptive least squares algorithm for estimating the power line interference in surface electromyography (sEMG) signals. The algorithm estimates the power line interference, without the need for a reference input. Power line interference can be removed by subtracting the estimate from the original sEMG signal. The algorithm is evaluated with simulated sEMG based on its ability to accurately estimate power line interference at different frequencies and at various signal-to-noise ratios. Power line estimates produced by the algorithm are accurate for signal-to-noise ratios below 15 dB (SNR estimation error at 15 dB is 14.7995 dB + 1.6547 dB).

A novel approach to localized muscle fatigue assessment

A method for generating a function which maps a set of surface myoelectric parameters to a fatigue index is proposed in this work. This forms the basis of a novel approach to assessing localized muscle fatigue with the myoelectric signal. An artificial neural network with a multilayer perceptron architecture was utilized to tune the function to emphasize trends in input parameters which are due to fatigue. The concept was tested empirically under static, cyclic, and random conditions. Results indicate improved performance when compared to fatigue assessment performance of mean frequency estimates.