An approach to cardiac arrhythmia analysis using hidden Markov models.

This paper describes a new approach to ECG arrhythmia analysis based on "hidden Markov modeling" (HMM), a technique successfully used since the mid-1970's to model speech waveforms for automatic speech recognition. Many ventricular arrhythmias can be classified by detecting and analyzing QRS complexes and determining R-R intervals. Classification of supraventricular arrhythmias, however, often requires detection of the P wave in addition to the QRS complex. The hidden Markov modeling approach combines structural and statistical knowledge of the ECG signal in a single parametric model. Model parameters are estimated from training data using an iterative, maximum likelihood reestimation algorithm. Initial results suggest that this approach may provide improved supraventricular arrhythmia analysis through accurate representation of the entire beat including the P wave.

Detection of late potentials by adaptive filtering.

Recent studies have shown small deflections in the terminal portion of the QRS complex in postmyocardial infarction (MI) patients with episodes of potentially dangerous ventricular arrhythmias. These very small deflections, termed late potentials, are difficult to observe in ECGs acquired with electrodes on the chest surface owing to myoelectric noise from underlying muscle and other environmental noise. Most research into enhancement of late potentials has focused on signal averaging. Our interest is in a mathematical signal-processing technique called time sequenced adaptive filtering. In ECG signals processed with adaptive filtering, late potentials are discernable in individual beats. This processing technique may also allow visualization of late potentials in signals from patients with intraventricular conduction delays. Preliminary studies used 12 normal subjects and 5 patients scheduled for electrophysiology studies (EPs), with histories of cardiopulmonary arrest, symptoms of unexplained syncope, or documented malignant ventricular arrhythmias. Three of the five patients studied by EPs had inducible sustained ventricular tachycardia. Subsequently, when ECG signals were obtained from surface electrodes and process by adaptive filtering, these same three patients had prolonged QRS durations (greater than 120 ms) with occurrence of additional low voltage activity (greater than or equal to 5-10 mV) in the terminal portion of the QRS. Two of these patients had bundle branch blocks as determined by a 12-lead ECG. The remaining patients and the control subjects had normal QRS durations with no additional activity. Two subjects in the control group had some activity in the terminal portion of the QRS without prolongation of QRS duration.(ABSTRACT TRUNCATED AT 250 WORDS)

Enhancement of low-level ECG components in noise with time-sequenced adaptive filtering.

Observation of low amplitude components in the ECG motivated our interest in time-sequenced adaptive filtering. This technique is applicable to signals that are cyclic in nature. Two simultaneously acquired signals are used in the technique. It is assumed that the underlying signal is correlated and that noise is uncorrelated between the two channels. Instead of one enhancer that continuously adjusts its characteristics over the time course of the signal, each cycle is subdivided into intervals; a given enhancer is used only on the same corresponding interval in successive cycles. This minimizes the signal range over which the enhancer must adjust its characteristics. In the time-sequenced approach, it is important that the enhancers adapt at the same rate. Thus, each has its own feedback coefficient derived from an average of 10 consecutive ECG cycles and an estimate of noise in intervals where no signal is present. Each feedback coefficient is augmented over the first 10 beats. To improve adaptation, a means to update each filter on the preceding beat and the current beat was developed. Lastly, a weight averaging scheme was developed to circumvent weight stalling. The procedure has enabled observation of both His activity and late potentials in individual beats from signals acquired from the chest surface.

Use of hidden Markov models for electrocardiographic signal analysis.

Hidden Markov modelling (HMM) is a powerful stochastic modelling technique that has been successfully applied to automatic speech recognition problems. We are currently investigating the application of HMM to electrocardiographic signal analysis with the goal of improving ambulatory ECG analysis. The HMM approach specifies a Markov chain to model a "hidden" sequence that in this case is the underlying state of the heart. Each state of the Markov chain has an associated output function that describes the statistical characteristics of measurement samples generated during that state. Given a measurement sequence and HMM parameter estimates, the most likely underlying state sequence can be determined and used to infer beat classification. Advantages of this approach include resistance to noise, ability to model low-amplitude waveforms such as the P wave, and availability of an algorithm for automatically estimating model parameters from training data. We have applied the HMM approach to QRS complex detection and to arrhythmia analysis with encouraging results.

A dipole plus quadrupole lead system for human electrocardiography.

A lead system was constructed to extract dipole and quadrupole components of cardiac sources from surface electrocardiograms (ECGs) recorded at 16 sites. The lead system was based on an analysis of a computerized model of a multipole equivalent cardiac generator in a homogeneous torso. The model was previously determined from extensive geometric and electrocardiographic data obtained from one subject. Dipole components estimated with the lead system were 89% accurate for the original subject. Evaluation of the lead system on this subject and in 59 other subjects included calculation of the effect of non-dipolar sources on the values of the estimated dipole components, comparison of the consistency of equivalent sources found independently at two origins in the heart region, and reconstruction of ECGs from lead system components. Dipole consistency at the origins was maintained over the wide range of age, weight, and body shape which characterized the subject population. Whereas quadrupole terms did not agree as well as the dipole terms, inclusion of the quadrupole reduced ECG reconstruction errors by a factor of about three compared to errors for the dipole alone. Together, the dipole and quadrupole accounted for almost 90% of the electrocardiographic information measured on the body surface with the D/Q lead system.