We have developed an adaptive matched filtering algorithm based upon an artificial neural network (ANN) for QRS detection. We use an ANN adaptive whitening filter to model the lower frequencies of the ECG which are inherently nonlinear and nonstationary. The residual signal which contains mostly higher frequency QRS complex energy is then passed through a linear matched filter to detect the location of the QRS complex. We developed an algorithm to adaptively update the matched filter template from the detected QRS complex in the ECG signal itself so that the template can be customized to an individual subject. This ANN whitening filter is very effective at removing the time-varying, nonlinear noise characteristic of ECG signals. Using this novel approach, the detection rate for a very noisy patient record in the MIT/BIH arrhythmia database is 99.5%, which compares favorably to the 97.5% obtained using a linear adaptive whitening filter and the 96.5% achieved with a bandpass filtering method.

No abstract

SummaryBackground: Atrial fibrillation is often first recognized after a complication such as embolic stroke has occurred. Limited data are available for the prospective identification of patients at risk for developing atrial fibrillation.Hypothesis: Demonstration of areas of slow conduction in the atrium by means of P-wave signal averaging may identify individuals at risk for atrial fibrillation.Methods: P-wave signal averaging from the surface electrocardiogram was performed in 199 normal controls and 81 patients with paroxysmal atrial fibrillation using an automated, P-triggered, high-resolution signal for analysis.Results: Of the variables measured, the filtered P-wave duration and P-wave integral were significantly different between controls and patients (filtered P-wave duration 120 ± 9 vs. 145 ± 21 and P-wave integral 666 ± 208 vs. 868 ± 352), whereas the terminal root-mean-square (RMS) voltages (RMS 20, RMS 30, RMS 40) showed no significant differences between the two groups. Regression analysis of the first and second measurement of the filtered P-wave duration obtained during consecutive tests showed excellent reproducibility (r and r 2 of 0.96 and 0.92). The duration of the filtered P wave showed no age dependence but was shorter in women.

No abstract

The application of multi-layer perceptron artificial neural network model to detect the QRS complex in the ECG signal processing is presented. The objective is to improve the heart beat detection rate under the presence of severe background noise. An adaptively tuned multi-layer perceptron structure is used to model the non-linear, time varying background noise. The noise is removed by subtracting the predicted noise from the original signal. Preliminary experiment results indicate that the ANN based approach consistently out-perform the conventional band-pass filtering approach and the linear adaptive filtering approach. Such performance enhancement is most critical toward the development of a practical automated on-line ECG Arrhythmia monitoring system. 0-7803-0557-4/92$03.00 0 1992

We present several methods to reduce excessive number of neurons and synaptic weights in a feedforward, multi-layer perceptron artificial neural network (ANN). To reduce the synaptic weights, we replace the original weight matrix by a product of two smaller matrices so that the number of multiplications required can be reduced. To reduce the hidden units, we exploit the correlation among the outputs of the hidden neurons in the same layer. We propose a method to identil5, and remove redundant hidden units and update the weights of the remaining neurons. This approach offers potentially good performance without retraining. But when retraining is applied to fine-tune the reduced network, the updated weights become very good initial conditions enabling much faster training compared with training with random initial conditions.

No abstract

In this paper, an artificial neural network (ANN) is applied to perform the task of acoustic-to-articulatory inversion. The objective is to model the highly non-linear mapping from LPC code to corresponding articulatory parameters with a multi-layer perceptron ANN structure. Such information will facilitate the study of the relationships between the acoustic signal and the physical vocal tract which produces it. Several novel approaches for devising the ANN structure have been evaluated. Specifically, in this paper the performance using two learning algorithms, a back-propagation (BP) algorithm, and a random optimization (RM) algorithm, is compared. To reduce excessive, redundant hidden units in the multi-layer perceptron model, a singular value decomposition is applied to either the weight matrix or the output covariance matrix of the hidden units to check their corresponding ranks. In both cases, their ranks are closely related to the number of essential decision regions in the input data.