P. de Chazal

Automated processing of the single-lead electrocardiogram for the detection of obstructive sleep apnoea

A method for the automatic processing of the electrocardiogram (ECG) for the detection of obstructive apnoea is presented. The method screens nighttime single-lead ECG recordings for the presence of major sleep apnoea and provides a minute-by-minute analysis of disordered breathing. A large independently validated database of 70 ECG recordings acquired from normal subjects and subjects with obstructive and mixed sleep apnoea, each of approximately eight hours in duration, was used throughout the study. Thirty-five of these recordings were used for training and 35 retained for independent testing. A wide variety of features based on heartbeat intervals and an ECG-derived respiratory signal were considered. Classifiers based on linear and quadratic discriminants were compared. Feature selection and regularization of classifier parameters were used to optimize classifier performance. Results show that the normal recordings could be separated from the apnoea recordings with a 100% success rate and a minute-by-minute classification accuracy of over 90% is achievable.

A Patient-Adapting Heartbeat Classifier Using ECG Morphology and Heartbeat Interval Features

An adaptive system for the automatic processing of the electrocardiogram (ECG) for the classification of heartbeats into one of the five beat classes recommended by ANSI/AAMI EC57:1998 standard is presented. The heartbeat classification system processes an incoming recording with a global-classifier to produce the first set of beat annotations. An expert then validates and if necessary corrects a fraction of the beats of the recording. The system then adapts by first training a local-classifier using the newly annotated beats and combines this with the global-classifier to produce an adapted classification system. The adapted system is then used to update beat annotations. The results of this study show that the performance of a patient adaptable classifier increases with the amount of training of the system on the local record. Crucially, the performance of the system can be significantly boosted with a small amount of adaptation even when all beats used for adaptation are from a single class. This study illustrates the ability to provide highly beneficial automatic arrhythmia monitoring and is an improvement on previously reported results for automated heartbeat classification systems

A parametric feature extraction and classification strategy for brain-computer interfacing

Parametric modeling strategies are explored in conjunction with linear discriminant analysis for use in an electroencephalogram (EEG)-based brain-computer interface (BCI). A left/right self-paced typing exercise is analyzed by extending the usual autoregressive (AR) model for EEG feature extraction with an AR with exogenous input (ARX) model for combined filtering and feature extraction. The ensemble averaged Bereitschaftspotential (an event related potential preceding the onset of movement) forms the exogenous signal input to the ARX model. Based on trials with six subjects, the ARX case of modeling both the signal and noise was found to be considerably more effective than modeling the noise alone (common in BCI systems) with the AR method yielding a classification accuracy of 52.8 /spl plusmn/ 4.8% and the ARX method an accuracy of 79.1 /spl plusmn/ 3.9% across subjects. The results suggest a role for ARX-based feature extraction in BCIs based on evoked and event-related potentials.

Telephony-based voice pathology assessment using automated speech analysis

A system for remotely detecting vocal fold pathologies using telephone-quality speech is presented. The system uses a linear classifier, processing measurements of pitch perturbation, amplitude perturbation and harmonic-to-noise ratio derived from digitized speech recordings. Voice recordings from the Disordered Voice Database Model 4337 system were used to develop and validate the system. Results show that while a sustained phonation, recorded in a controlled environment, can be classified as normal or pathologic with accuracy of 89.1%, telephone-quality speech can be classified as normal or pathologic with an accuracy of 74.2%, using the same scheme. Amplitude perturbation features prove most robust for telephone-quality speech. The pathologic recordings were then subcategorized into four groups, comprising normal, neuromuscular pathologic, physical pathologic and mixed (neuromuscular with physical) pathologic. A separate classifier was developed for classifying the normal group from each pathologic subcategory. Results show that neuromuscular disorders could be detected remotely with an accuracy of 87%, physical abnormalities with an accuracy of 78% and mixed pathology voice with an accuracy of 61%. This study highlights the real possibility for remote detection and diagnosis of voice pathology.

Automated processing of shoeprint images based on the Fourier transform for use in forensic science

The development of a system for automatically sorting a database of shoeprint images based on the outsole pattern in response to a reference shoeprint image is presented. The database images are sorted so that those from the same pattern group as the reference shoeprint are likely to be at the start of the list. A database of 476 complete shoeprint images belonging to 140 pattern groups was established with each group containing two or more examples. A panel of human observers performed the grouping of the images into pattern categories. Tests of the system using the database showed that the first-ranked database image belongs to the same pattern category as the reference image 65 percent of the time and that a correct match appears within the first 5 percent of the sorted images 87 percent of the time. The system has translational and rotational invariance so that the spatial positioning of the reference shoeprint images does not have to correspond with the spatial positioning of the shoeprint images of the database. The performance of the system for matching partial-prints was also determined.

Automatic classification of ECG beats using waveform shape and heart beat interval features

The paper presents the classification performance of an automatic classifier of the electrocardiogram (ECG) for the detection of normal, premature ventricular contraction and fusion beat types. Both linear discriminants and feedforward neural networks were considered for the classifier model. Features based on the ECG waveform shape and heart beat intervals were used as inputs to the classifiers. Data was obtained from the MIT-BIH arrhythmia database. Cross-validation was used to measure the classifier performance. A classification accuracy of 89% was achieved which is a significant improvement on previously published results.

Using wavelet coefficients for the classification of the electrocardiogram

This study investigates the automatic classification of the Frank lead electrocardiogram (ECG) into different pathophysiological disease categories. Coefficients from the discrete wavelet transform are used to represent the ECG diagnostic information and a comparison of the performance of classifiers processing feature sets generated using different mother wavelets is made. Fifteen feature sets are calculated from three Daubechies wavelets, with the decomposition level varied between 3 and 7. The classification performance of each feature set was optimised using automatic feature selection and by combining classifications of multi-beat ECG information. Throughout the study a database of 500 ECG records with examples from 7 disease categories was used. The classification of each record is known with 100% confidence and is based on ECG independent information. Using multiple runs of 10-fold cross-validation to obtain all results, it was shown that the overall classification performance of the different feature sets was 71.6-74.2%. In addition, the wavelet order and level had little influence on the overall performance. Analysis of the automatically chosen features reveal that time-frequency bands in the vicinity of the QRS onset and the T-wave are consistently selected.

Automatic classification of sleep apnea epochs using the electrocardiogram

This study investigated the automatic prediction of epochs of sleep apnea from the electrocardiogram. A large independently validated database of 70 single lead ECGs, each of approximately 8 hours in duration, was used throughout the study. Thirty five of these records were used for training and 35 retained for independent testing. After considering a wide variety of features the authors found that features based on the power spectral density estimates of the R-wave maxima and R-R intervals to be the most discriminating. Results show that a classification rate of approximately 89% is achievable.

Parametric models and spectral analysis for classification in brain-computer interfaces

Parametric modelling strategies and spectral analysis are explored in conjunction with linear discriminant analysis to facilitate an EEG based direct-brain interface for use by disabled people. A self-paced typing exercise is analysed by employing for feature extraction, respectively, an autoregressive model, an autoregressive with exogenous input model, and a time-frequency decomposition of the data. Modelling both the signal and noise is found to be more, effective than modelling the noise alone with the former yielding an accuracy of 70.7% and the latter an accuracy of 57.4%. Experiments, using the raw samples of a short-time power spectral density estimate of each trial as features, yielded an accuracy of 62.5%.

Low computational cost classifiers for ECG diagnosis using neural networks

We investigate and compare a number of computationally efficient classifiers for categorising the Frank lead ECG as normal or one of six disease conditions using a neural network expert system. These include a power spectral density estimate, and two discrete wavelets, the Daubechies wavelet of order 10 (db 10) and the Symlet wavelet of order 8 (sym8) applied to a single beat of the X, Y and Z Frank leads. Simple statistical parameters derived from these transforms and from reconstructed filtered signals were used as inputs to a neural network with no hidden units and a softmax output stage. We used multiple runs of 10 fold cross validation to obtain estimates of classifier performance. Best results were obtained for the db 10 parameters when age and sex were also added. Overall accuracy was 68.8/spl plusmn/0.6%. These results are comparable to those derived from neural nets trained with over 229 scalar parameters (70.9/spl plusmn/0.6%) and were derived at much lower computational cost. The methods derived can be easily implemented in real time using a DSP processor.