E.J. Ciaccio

Biosignal pattern recognition and interpretation systems

A general framework is given to describe pattern recognition and interpretation. Pattern analysis stages are described, with consideration of difficulties in implementation and uncertainties present at each level. The main forms of pattern analysis-statistical, syntactic, and artificial intelligence (connectionist and symbolic) methods-have different strengths and weaknesses, depending on the stage of pattern analysis at which they are used. In general, statistical, syntactic, and connectionist techniques are used for pattern recognition, and statistical and symbolic techniques are used for pattern interpretation. Largely, pattern interpretation involves reasoning with uncertainty. Multichannel recordings increase the information available about specific physiologic events, at the expense of processing complexity.<<ETX>>

Biosignal pattern recognition and interpretation systems. 3. Methods of classification

The following topics are discussed: Bayes/minimum distance classifiers; maximum likelihood classification estimation; k-nearest neighbor classification; entropy criteria; syntactic techniques; string matching; the Cocke-Younger-Kasami parsing algorithm; syntactic learning; finite-state automata; neural network classification techniques; learning vector quantization; cluster swapping; hierarchial clustering procedures.<<ETX>>

Biosignal pattern recognition and interpretation systems. 2. Methods for feature extraction and selection

Some feature extraction methods used in biomedical signal pattern recognition are presented. Particular attention is given to nontransformed signal characteristics, transformed signal characteristics, structural descriptors, graph descriptors, and feature selection methods. It is noted that the wide variety of techniques used for feature extraction presents two problems: which techniques should be used and how to select from among the features that each extraction technique generates. Selected features are best only by some standard; therefore, techniques for generation of features tend not to be very portable from one pattern processing problem to another. Production of salient features is the connecting link between prototypical and symbolic representations of a class. Often, thresholds govern the selection of features. Many techniques do not generate independent features; therefore, there is redundancy in the data, which potentially affects both efficiency and accuracy in pattern recognition.<<ETX>>

Biosignal pattern recognition and interpretation systems. 4. Review of applications

Applications in the engineering literature of pattern analysis to biomedical signals (electrophysiological (muscle, nerve, and heart) and biomechanical) are presented. The signals include EEG, EMG, ECG and blood pressure. Feature selection techniques often use: 1) thresholding or 2) reasoning with uncertainty. Many pattern analysis systems are only partial systems: they perform stages of pattern recognition or pattern interpretation, but they are not complete systems. Several general trends can be observed: usually, but not always, statistical, syntactic, and connectionist techniques are used for pattern recognition, and statistical and symbolic techniques are used for pattern interpretation. Often, single time series are analyzed (rather than multichannel data). Descriptions of shape are important to pattern analysis at the feature, property, and recognizer (classification, discrimination, detection, and parameter estimation) stages. Typically, statistical techniques are used at various stages of pattern recognition, which assume (perhaps incorrectly) Gaussian distributions.<<ETX>>

Algorithm for reduction of mechanical noise in arterial pulse recording with tonometry

The quality of arterial pulse recording is diminished by mechanical noise, mainly in the form of movement by the subject, which is unavoidable in environments such as treadmill stress testing aid in aerospace settings. A method is presented for minimizing motion artifacts in such recordings. The pulse shape is obtained, through the principle of tonometry, by means of two piezoelectric (crystal) sensors positioned at the pulse site. One motion sensor output is subtracted from the other (pulse sensor=pulse+mostly common-mode motion) output. Observations show that the mechanical gain at each sensors interface differs. Subject movements can shift the transducer slightly and makes the gain time-varying. An algorithm is developed that estimates the best fit of pulse signal vs. noise signal in real time, using the method of least squares. An equation is presented to correct the pulse.<<ETX>>

Localized spatial discrimination of epicardial conduction paths after linear transformation of variant information

We present a method for the localized statistical discrimination of class populations based on the Karhunen-Loève and Fukunaga-Koontz transforms. These transforms provide features that model the variance of a sample distribution. The spatial series of a 196 channel epicardial electrogram recording from an arrhythmogenic postinfarct canine were analyzed. For each type of rhythm studied, Karhunen-Loève and Fukunaga-Koontz expansions were computed from five training sets of spatial data, corresponding to five locations across the surface of the heart. Nonparametric statistical tests were then used for discriminant analysis to compare properties representative of the distribution from each proposed class. In a comparison of properties from sinus rhythm to those of two ventricular tachycardias, several spatial regions exhibited statistically significantly different propagation characteristics. These areas were observed by visual inspection of electrogram activation maps to be characterized by conductive gradients, which differed in magnitude and direction from one rhythm to another. The regions in which the propagation characteristics are of greatest difference in each tachycardia were centered upon sites of conduction block, manifested by reentrant circuit rhythms. Therefore, the importance of the technique for the localization of specific electrophysiologic events is demonstrated. This study extends previous work of our group on biosignal pattern recognition to encompass localized spatial data.

Pattern recognition and interpretation of electromyogram data from cat jaw muscle

This study investigates the effect of emotional behavior on the masseteric muscle EMG response patterns. Two experimental protocols are utilized: (1) does not elicit emotional behavior (stick chewing) and (2) elicits emotional behavior (hypothalamic stimulation). The Karhunen-Loève transform is used to compute features which exactly represent the correlated patterns of mean-zero observations, with data compression and noise immunity. Using nonparametric tests, it is found that the populations of biting and hissing features are significantly different (p < 0.05), with increased statistical significance as the size of the training set is increased. No statistically significant difference is seen in a test of the two biting populations.

Array sensor for arterial pulse recording-reduction of motion artifact

A linear array of three high sensitivity piezoelectric crystals are spaced 0.75 cm apart and bonded to a plastic block. The crystals are used in the thickness compression mode to measure the pressure pulse. The signal from each crystal is amplified and processed for noise reduction. The array is aligned perpendicular to the artery, and information from the off-axis sensors is averaged and subtracted from the on-axis signal. In addition, the sensors are aligned on-axis for determination of blood velocity from the phase difference between sensors. Experimental results indicate that motion artifact is mostly common-mode and that the contact stress on the surface of the skin is therefore uniform. Preliminary cancellation techniques provided are simple yet effective in restoring disturbed pulse recordings.<<ETX>>

Detection of statistically significant differences in cat jaw muscle movement using the EMG linear envelope

A study of jaw movement in the cat using experimental protocols to elicit biting and hissing responses is discussed. The EMG was recorded and its morphology was characterized by employing preprocessing and pattern analysis techniques for feature extraction. Time series electromyographic (EMG) episodes were conditioned by reduction to 40 representative sample points (approximating the EMG linear envelope), aligning, and taking the Karhuenen-Loeve expansion. To test the hypothesis that biting and hissing responses were reproducible, and yet statistically different from each other, the F test was used to determine the significance of differences in standard deviation of the KLE eigenvectors. It was found that jaw movement was significantly different for biting versus hissing protocols.<<ETX>>

A least squares algorithm with one phase weight for adaptive cancelation of noise from biomedical signals

A least squares algorithm is presented which uses one weight for phase shift correction per reference input for adaptive noise canceling with one or multiple reference inputs. The method uses an iterative gradient search procedure which assumes that phase shifts between inputs are small compared to the frequency of meaningful, correlated, periodic noise components. In a simulation in which the primary signal contained additive, correlated sinusoidal and random noise, the weights converged stably and rapidly to the minimum of the performance surface. The algorithm is potentially useful in real-time medical applications, where minimizing the cost of implementation is essential.<<ETX>>