A comparison of neural feature extraction methods for brain-machine interfaces.

Brain-machine interferes (BMIs) have shown promise in augmenting people's control of their surroundings, especially for those suffering from paralysis due to neurological disorders. This paper describes an experiment using the rodent model to explore information available in neural signals recorded from chronically implanted intracortical microelectrode arrays. In offline experiments, a number of neural feature extraction methods were utilized to obtain neural activity vectors (NAVs) describing the activity of the underlying neural population while rats performed a discrimination task. The methods evaluated included standard techniques such as binned spike rates and local field potential spectra as well as more novel approaches including matched-filter energy, raw signal spectra, and an autocorrelation energy measure (AEM) approach. Support vector machines (SVMs) were trained offline to classify left from right going movements by utilizing features contained in the NAVs obtained by the different methods. Each method was evaluated for accuracy and robustness. Results show that most algorithms worked well for decoding neural signals both during and prior to movement, with spectral methods providing the best stability.

Estimation of neural energy in microelectrode signals.

We considered the problem of determining the neural contribution to the signal recorded by an intracortical electrode. We developed a linear least-squares approach to determine the energy fraction of a signal attributable to an arbitrary number of autocorrelation-defined signals buried in noise. Application of the method requires estimation of autocorrelation functions R(ap)(tau) characterizing the action potential (AP) waveforms and R(n)(tau) characterizing background noise. This method was applied to the analysis of chronically implanted microelectrode signals from motor cortex of rat. We found that neural (AP) energy consisted of a large-signal component which grows linearly with the number of threshold-detected neural events and a small-signal component unrelated to the count of threshold-detected AP signals. The addition of pseudorandom noise to electrode signals demonstrated the algorithm's effectiveness for a wide range of noise-to-signal energy ratios (0.08 to 39). We suggest, therefore, that the method could be of use in providing a measure of neural response in situations where clearly identified spike waveforms cannot be isolated, or in providing an additional 'background' measure of microelectrode neural activity to supplement the traditional AP spike count.

Action potential classifiers: a functional comparison of template matching, principal components analysis and an artificial neural network.

Multiunit neural activity occurs often in electrophysiological studies when utilizing extracellular electrodes. In order to estimate the activity of the individual neurons each action potential in the recording must be classified to its neuron of origin. This paper compares the accuracy of two traditional methods of action potential classification--template matching and principal components--against the performance of an artificial neural network (ANN). Both traditional methods use averages of action potential shapes to form their corresponding classifiers while the artificial neural network 'learns' a nonlinear relationship between a set of prototype action potentials and assigned classes. The set of prototypic action potentials and the assigned classes is termed the training set. The training set contained action potentials from each class which exhibited the full range of amplitude variability. The ANN provided better classification results and was more robust in analysis of across-animal data sets than either of the traditional action potential classification methods.

An in-vitro investigation of prosthetic heart valve cavitation in blood.

Vapor cavities produced by low pressure fluid flow conditions have been observed in the vicinity of mechanical heart valves for many years. As cavities collapse during pressure recovery, they can produce stresses large enough to cause pitting of the valve occluders and lysing or activation of blood cells. To date, no method has been presented for the quantification of mechanical heart valve cavitation in blood because it has only been detected optically in transparent blood analog fluids. This paper describes a novel method for quantifying cavitation intensities in opaque fluids such as blood. It is based on the detection of high frequency pressure oscillations (35-350 kHz) at a location 4.5 cm proximal to a Björk-Shiley monostrut mitral valve in a mock circulatory loop driven by a Penn State Electric Ventricular Assist Device. The pressure oscillations which result from cavity collapse are used to quantify cavitation intensities in blood. One time domain and three frequency domain parameters have been developed to quantify cavitation intensity during a single valve closure event and over an ensemble of closure events. The time domain parameter is the Root Mean Squared (RMS) value of the pressure signal after it has been high-pass filtered at 35 kHz. The other three parameters are derived from the power spectrum of the pressure signal. One is the maximum value of the power spectrum between 100 and 200 kHz, another is the area under the power spectrum between 35 and 400 kHz, and the last is the volume under a 3-dimensional time vs. frequency vs. power spectrum plot. The parameters are averaged over a random sample of pressure traces to determine an average cavitation intensity for each operating condition studied. In addition, cavitation pressure fluctuations and hemolysis rates were determined simultaneously at several different mock flow loop operating conditions using porcine blood, and the relationships between various measures of cavitation intensity and the associated index of hemolysis have been established. Hemolysis was shown to increase with cavitation intensity.

Ratio of variance to mean of action potential counts for an auditory nerve fiber model with second-order refractory behavior.

Sequences of interval times were generated with the first- and second-order interval probability dependence proposed by Gaumond et al. [J. Neurophysiol. 48, 856-873 (1982)] to characterize cochlear nerve fiber spontaneous action potential (AP) sequences. The ratio F(T) of the variance to the mean of the count of APs in bins of T s was computed for the simulated sequences, and compared to the corresponding measure for cochlear nerve fiber data presented by Lowen and Teich [J. Acoust. Soc. Am. 92, 803-806 (1992)]. The simulations were generally consistent with the data in the range 15 < T < 100 ms.

The effects of metals on a transcutaneous energy transmission system.

The effects of metal objects on the mutual inductance, self-inductance, and effective series resistance (ESR) of the coaxial coils of a transcutaneous energy transmission system (TETS) were investigated theoretically and experimentally. The theory considers a thin conducting sheet of infinite size aligned parallel to a current-carrying coil. Results of the theory indicate that coil parameters vary with the distance from the sheet to the coil. Changes in mutual and self inductance are independent of the conductivity and thickness of the sheet, with a larger percentage change for mutual inductance than for self inductance. Changes in ESR are proportional to the surface resistivity of the sheet. Experimental measurements using several aluminum sheets and a titanium alloy can in the presence of the TETS coils used for the Penn State artificial heart showed excellent agreement with the theory for inductance parameters and agreed within a factor of 2 for ESR measurements when skin effect was considered. It was generally observed that mutual inductance drops to 65% of its initial value as a metal sheet is moved to within 1-2 cm of the coil, while self-inductance drops to 80% of its initial value when the sheet is 2 cm from the coil. Measured changes in ESR tended to be higher than the calculated values with the discrepancy depending on distance to the metal object. Theory and measurements were extended to the case of lateral misalignment of the coils.

Models for the generation of the binaural difference response.

Two simple models are examined in order to explain the observation that a portion of the binaural-evoked response is less than the sum of monaural-evoked responses in human and animal subjects. The sum of monaural responses minus the binaural response is called the binaural difference (BD). Each model acts on binaural input signals and applies a single memoryless nonlinearity. One model (IE) applies a rectifying nonlinearity to the difference of input signals, while the other (EE) applies a compressive nonlinearity to the sum of input signals. These models are suggested by properties of inhibitory-excitatory (IE) and excitatory-excitatory (EE) neurons of the auditory brainstem. Parameters can be found that enable each model to produce a ratio of BD to summed monaural response which is invariant with input stimulus level. The IE model, but not the EE model, has a BD whose level is linearly related to input stimulus level.

Permanent circulatory support systems at the Pennsylvania State University.

Permanent circulatory support systems are required for patients in whom myocardial damage is irreversible and cardiac transplantation is not possible. Two systems are described which provide long term circulatory support: the left ventricular assist system and the total artificial heart. These systems are based on the design of a pusher plate actuated blood pump, driven by a small brushless dc electric motor and rollerscrew driver. An implantable motor controller maintains suitable physiologic flow rates for both systems and controls left-right balance in the total artificial heart. Other parts of the system include an intra-thoracic compliance chamber, transcutaneous energy and data transmission system, and internal and external batteries.

Analysis of cat multichannel acoustic brain-stem response data using dipole localization methods.

We have used dipole localization methods (DLMs) to examine multielectrode acoustic brain-stem response (ABR) data for 9 data files collected in 5 experiments on 2 cats. The analytical model for scalp surface voltage consisted of a single current dipole in a homogeneous spherical head model. DLM sources were compared to ABR generator sites determined in intracranial and lesion experiments described in the literature. DLM source along the mediolateral axis is consistent with a predominantly ipsilateral cochlear nucleus origin for P2, and a predominantly contralateral origin for P3. DLM source for P4 and P5 was near the midline plane, consistent with the suggestion of sources arrayed along both sides of the midline. Average DLM source position increased by 7.1 mm superiorly from P3 to P5, nearly what would be expected from movement from superior olive to inferior colliculus levels. A lack of DLM source movement in the postero-anterior direction is attributed to limitations of the spherical model. The overall progression of mean DLM equivalent source position at successive ABR peaks demonstrates the existence of substantive information about generator source position in scalp surface ABR field patterns. However, model constraints of available DLM techniques present severe restrictions on the physiological interpretation of results.

Multichannel data acquisition system for recording the acoustic brainstem response.

A system has been developed to sample and average multi-channel ABR data. All channels are sampled simultaneously. The simple design incorporates one sample-and-hold amplifier and one analog-to-digital converter in each channel. For sampling of 13 channels or less, sampling interval must exceed 36 microseconds for full 10-bit resolution. For sampling more than 13 channels, data transfer to the host computer limits minimum sampling interval to 20 + (1.2 X N) microsecond, where N is the number of channels sampled.

Response of cochlear nerve fibers to brief acoustic stimuli: role of discharge-history effects.

The probability that a cochlear nerve fiber spike discharge occurs during a time interval delta t depends on both the acoustic stimulus and on aftereffects from earlier spike discharges. We have examined the influence of discharge-history on post-stimulus time (PST) histogram responses to acoustic click and tone-burst stimuli. Discharge-history effects were found to include the modification of observed interpeak times of PST responses to clicks, a loss of distinct peaks in the click response of high characteristic frequency (CF greater than 5 kHz) fibers, and changes in the ratio of initial to steady-state response portions of tone-burst responses. The method used to separate discharge-history from stimulus-related factors is based on a model developed in Gaumond et al. [J. Neurophysiol. 48, 856-873 (1982)]. The results are in general agreement with those obtained by the method of Peter Gray [Biophys. J. 7, 759-777 (1967)], which discards from consideration those portions of the response record not preceded by a silent interval of 20 or 25 ms or more. Our method requires more assumptions about the spike train, but produces less variable results by utilizing more of the spike train data.