Knowledge-based and data-driven models in arrhythmia fuzzy classification.

OBJECTIVES: Fuzzy rules automatically derived from a set of training examples quite often produce better classification results than fuzzy rules translated from medical knowledge. This study aims to investigate the difference in domain representation between a knowledge-based and a data-driven fuzzy system applied to an electrocardiography classification problem. METHODS: For a three-class electrocardiographic arrhythmia classification task a set of fifteen fuzzy rules is derived from medical expertise on the basis of twelve electrocardiographic measures. A second set of fuzzy rules is automatically constructed on thirty-nine MIT-BIH database's records. The performances of the two classifiers on thirteen different records are comparable and up to a certain extent complementary. The two fuzzy models are then analyzed, by using the concept of information gain to estimate the impact of each ECG measure on each fuzzy decision process. RESULTS: Both systems rely on the beat prematurity degree and the QRS complex width and neglect the P wave existence and the ST segment features. The PR interval is not well characterized across the fuzzy medical rules while it plays an important role in the data-driven fuzzy system. The T wave area shows a higher information gain in the knowledge based decision process, and is not very much exploited by the data-driven system. CONCLUSIONS: The main difference between a human designed and a data driven ECG arrhythmia classifier is found about the PR interval and the T wave.

Input features' impact on fuzzy decision processes.

Many real-world applications have very high dimensionality and require very complex decision borders. In this case, the number of fuzzy rules can proliferate, and the easy interpretability of fuzzy models can progressively disappear. An important part of the model interpretation lies on the evaluation of the effectiveness of the input features on the decision process. In this paper, we present a method that quantifies the discriminative power of the input features in a fuzzy model. The separability among all the rules of the fuzzy model produces a measure of the information available in the system. Such measure of information is calculated to characterize the system before and after each input feature is used for classification. The resulting information gain quantifies the discriminative power of that input feature. The comparison among the information gains of the different input features can yield better insights into the selected fuzzy classification strategy, even for very high dimensional cases, and can lead to a possible reduction of the input space dimension. Several artificial and real-world data analysis scenarios are reported as examples in order to illustrate the characteristics and potentialities of the proposed method.

A characterization of HRV's nonlinear hidden dynamics by means of Markov models.

A study of the 24-h heart rate variability's (HRV) hidden dynamic is performed hour by hour, in order to investigate the evolution of the nonlinear structure of the underlying nervous system. A hierarchy of null hypotheses of nonlinear Markov models with increasing order n is tested against the hidden dynamic of the HRV time series. The minimum accepted Markov order supplies information about the nonlinearity of the HRV's hidden dynamic and consequently of the underlying nervous system. The Markov model with minimum order is detected for each hour of the RR time series extracted from seven 24-h electrocardiogram records of patients in different pathophysiological conditions, some including ventricular tachycardia episodes. Heart rate, pNN30, and LF/HF index plots are reported to serve as a reference for the description of the patient's cardiovascular frame during each examined hour. The minimum Markov order shows to be a promising index for quantifying the average nonlinearity of the autonomic nervous system's activity.

Dynamics extraction in multivariate biomedical time series.

A nonlinear analysis of the underlying dynamics of a biomedical time series is proposed by means of a multi-dimensional testing of nonlinear Markovian hypotheses in the observed time series. The observed dynamics of the original N-dimensional biomedical time series is tested against a hierarchy of null hypotheses corresponding to N-dimensional nonlinear Markov processes of increasing order, whose conditional probability densities are estimated using neural networks. For each of the N time series, a measure based on higher order cumulants quantifies the independence between the past of the N-dimensional time series, and its value r steps ahead. This cumulant-based measure is used as a discriminating statistic for testing the null hypotheses. Experiments performed on artificial and real world examples, including autoregressive models, noisy chaos, and nonchaotic nonlinear processes, show the effectiveness of the proposed approach in modeling multivariate systems, predicting multidimensional time series, and characterizing the structure of biological systems. Electroencephalogram (EEG) time series and heart rate variability trends are tested as biomedical signal examples.

Human heart rate variability and sleep stages.

With the aim of better understanding the dynamic changes in sympatho-vagal tone occurring during the night, human heart rate variability (HRV) during the various sleep stages was evaluated by means of autoregressive spectral analysis. Each recording consisted of an electroencephalogram, an electrooculogram, and electromyogram, and electrocardiogram, and a spirometry trace. All of the data were sampled and stored in digital form. Sleep was analysed visually, but HRV was analysed off-line by means of original software using Burg's algorithm to calculate the LF/HF ratio (LF: 0.04-0.12 Hz; HF: 0.15-0.35 Hz) for each sleep stage. Seven healthy subjects (four males; mean age 35 years) were enrolled in the study. Our findings show a progressive and significant reduction in the LF/HF ratio through sleep stages S1-S4, as a result of an increase in the HF component; this indicates the prevalence of parasympathetic activity during slow-wave sleep. During wakefulness, S1 and REM, the LF/HF values were similar and close to 1.

Classification of arrhythmic events in ambulatory electrocardiogram, using artificial neural networks.

We propose artificial neural networks (ANN) for ambulatory ECG arrhythmic event classification, and we compare them with some traditional classifiers (TC). Among them, the one based on the median method (heuristic algorithm) was chosen and taken as a quality reference in this study, while a back propagation based classifier, designed as an autoassociator for its peculiar capability of rejecting unknown patterns, was examined. Two tests were performed: the first to discriminate normal vs ventricular beats and the second to distinguish among three classes of arrhythmic events. The results show that the ANN approach is more reliable than the traditional classifiers in discriminating among many classes of arrhythmic events: 98% by ANN vs 99% by a TC for correctly classified normal beats, 98% by ANN vs 96% by TC for correctly classified ventricular ectopic beats, 96% by ANN vs 59% by TC for correctly classified supraventricular ectopic beats, and 83% by ANN vs 86% by median method for correctly classified aberrated atrial premature beats. This paper also tackles the problem of the management of classification uncertainty. Two concurrent uncertainty criteria have been introduced, to reduce the classification error of the unknown ventricular and supraventricular arrhythmic beats respectively. The error in ventricular beats case was kept close to 0% in average and for supraventricular beats was kept at 35% in average. So we can state that the ANN approach is powerful in classifying beats represented in the training set and that it manages the uncertainty in such a way as to reduce, in any case, the global error percentage.