Elsevier

Expert Systems with Applications

Volume 40, Issue 11, 1 September 2013, Pages 4490-4495
Expert Systems with Applications

Adaptive neuro fuzzy selection of heart rate variability parameters affected by autonomic nervous system

https://doi.org/10.1016/j.eswa.2013.01.055Get rights and content

Abstract

Heart rate variability (HRV) parameters can be used as specific indicator of autonomic nervous system (ANS) behavior. ANS, with its main two branches, sympathetic and parasympathetic, may be considered as a coordinated neuronal network which controls heart rate continually. Many parameters define heart rate variability in different domains such as time, frequency or nonlinear. An excessively high computational complexity can occur when developing models for medical applications when the best set of inputs to use is not known. To build a model that can predict a specific process output, it is desirable to select a subset of variables that are truly relevant or the most influential to this output. This procedure is typically called variable selection, and it corresponds to finding a subset of the full set of recorded variables that exhibits good predictive abilities. In this study an architecture for modeling complex systems in function approximation and regression was used, based on using adaptive neuro-fuzzy inference system (ANFIS). Variable searching using the ANFIS network was performed to determine how the ANS branches affect the most relevant HRV parameters. The method utilized may work as a basis for examination of ANS influence on HRV activity.

Highlights

► Architecture for modeling complex systems in function approximation and regression. ► Heart rate variability parameters used for regression and prediction analysis of autonomic nervous system activity. ► Finding which of the autonomic nervous system parameters have the largest influence on heart rate variability parameters. ► Variable selection method.

Introduction

Heart rate variability (HRV) estimation is a useful non-invasive tool for the assessment of autonomic nervous system (ANS) behavior. Sympathetic and vagal tracts of ANS are interacting at each moment. Reduced HRV is a marker of impaired functioning of the ANS and can be considered a risk factor in chronic diseases. Power spectral analysis of HRV has been widely used for evaluating the ANS function. Three vital oscillatory components were identified. The physiological meaning of the very-low-frequency (VLF) component is still discussed, the low-frequency (LF) component (0.04–0.15 Hz) reflects baroreflex sympathetic control of blood pressure, and the high-frequency (HF) component (0.15–0.4 Hz) reflects respiratory rhythm and is related to parasympathetic control of heart rate. In previous article we used two measures, the cardiac vagal index (CVI) and the cardiac sympathetic index (CSI), which indicate vagal and sympathetic functions separately (Petković and Ćojbašić, 2011). Toichi et al. (1997) have recommended these two measures. These two indices have found to be more reliable than those obtained by the other methods. Lin et al. (2010) have developed methodology for mining physiological condition from HRV analysis using CVI and CSI indexes.

The standard measurements intervening in the analysis of HRV include time domain indices, geometric methods and components of the frequency domain (Berntson et al., 1997). Time domain analysis addresses how much variability there is in heart rate. Time domain values result from simple statistical calculations managed on the set of adjacent intervals. Frequency domain analysis is used to divide the total variance of the heart rate into the variance accounted for by underlying groups of frequencies. Measurements of HRV are generally performed on the basis of 24 h Holter recordings (long-term recordings) or on shorter periods ranging from 0.5 to 5 min (short-term recordings). The use of long or short-term recordings depends on the type of study that has to be realized.

Although the analysis of HRV has gained popularity as a simple and non-invasive tool for assessing autonomic function in both normal subjects and in patients in a variety of clinical settings, the potential of this tool in evaluating the impermanent changes in cardiac autonomic modulation in response to physiological or pathological stimuli has not been clarified adequately. Previous studies suggest the possibility of autonomic dysfunction in patients with some diseases and HRV may be useful to detect the diseases progression (Aksoyek et al., 1999, Goto et al., 2001, Lin et al., 2006, Pancera et al., 1999, Xinbao et al., 2006).

Some nonlinear techniques are able to describe the HRV activity (Chunhua and Xinbao, 2004, Xinbao et al., 2006). Heart rate signals have fractal characteristics or self-similarity and fractal dimensionality. The detrended fluctuation analysis (DFA) is used to quantify the fractal scaling properties (Ashkenazy et al., 2001, Chen et al., 2010, Chiu et al., 2003, Flynn et al., 2005). The HRV fluctuation is characterized by two scaling exponents, one for short range and one for long range.

ANS has different effect on each of the HRV parameters. It is often possible to measure the value of many physical signals (variables), but it is not necessarily known which of them are relevant and required to solve the problem. An excessively high computational complexity can occur when developing multivariate models for medical applications when the best set of inputs to use is not known. The main problems to face here are that when the input dimensionality increases, the computational complexity and memory requirements of the model increase (in some cases even exponentially); learning is more difficult with unnecessary inputs.

Neural networks can be defined as an architecture comprising massively parallel adaptive processing elements interconnected via structured networks. Thus, the neural network models generated from this data must therefore rely on how effectively the chosen sensor data represents the system. Therefore, in order to build a model that can predict a specific process output, it is desirable to select a subset of variables that are truly relevant to this output. This procedure is typically called variable selection, and it corresponds to finding a subset of the full set of recorded variables that exhibits good predictive abilities (Anderson et al., 2000, Castellano and Fanelli, 2000, Cibas et al., 1996, Dieterle et al., 2003). A solution to the variable selection problem could be the utilization of prior knowledge in order to screen out the irrelevant variables. However, in most cases, the large number of variables and the complexity of the process dictate the application of more sophisticated methods. A more advanced approach is to consider the variable selection problem as an optimization procedure via genetic algorithms (Donald, 2002), where the objective is to minimize the error between the true values and the model predictions of the explained (output) variables, by selecting the proper explanatory (input) variables. One of the most powerful types of neural network system is adaptive neuro fuzzy inference system (ANFIS). ANFIS shows very good learning and prediction capabilities, which makes is and efficient tool to deal with encountered uncertainties in any system. Fuzzy Inference System (FIS) is the main core of ANFIS. FIS is based on expertise expressed in terms of ‘IF-THEN’ rules and can thus be employed to predict the behavior of many uncertain systems. FIS advantage is that it does not require knowledge of the underlying physical process as a precondition for its application. Thus the ANFIS integrates the fuzzy inference system with a back-propagation learning algorithm of neural network. There are several studies of the ANFIS for extraction of the fuzzy rules from the some models (Chan et al., 2011, Kwong et al., 2009).

The objective of variable selection is three-fold: improving the prediction performance of the predictors, providing faster and more cost-effective predictors, and providing a better understanding of the underlying process that generated the data i.e., providing the most influential parameters on the predictor (Despagne and Massart, 1998, Guyon and Elisseeff, 2003, Papadokonstantakis et al., 2005).

In this study, nine time domain, two frequency-nonlinear HRV parameters and three frequencies domain HRV parameters were used for regression and prediction analysis of ANS activity (Berntson et al., 1997, Malik, 1996). As ANS indication parameters, we have used CVI and CSI indexes which are obtained from Poincare plot indices SD1 and SD2. These two parameters (CVI and CSI) indicate vagal and sympathetic function of ANS separately. The main goal was to find which of the ANS parameters have the largest influence on HRV parameters. To deal with this problem, we made an architecture for modeling complex systems in function approximation problems by using ANFIS. ANFIS is a soft computing methodology. Yardimci (2009) demonstrates the possibilities of applying soft computing (SC) to medicine-related problems. Some of these SC methodologies have already been applied for ECG and HRV signals processing for classification (Annuradha and Reddy, 2008, Anuradha and Reddy, 2009, Mohammadzadeh-Asl and Setarehdan 2006, Ceylan et al., 2009, Osowski et al., 2009, Ozbay and Tezel, 2009, Rajendra et al., 2003), analysis (Ubeyli, 2010), recognition (Maglaveres et al., 1998, Osowski et al., 2009), diagnosis (Hosseini et al., 2006), and prediction (Patil & Kumaraswamy, 2009). ANFIS has been used only for the classification of ECG signals (Ubeyli, 2009, Nazmy et al., 2009).

The proposed method in this article is similar to above mentioned variable selection methods except we increase number of inputs sequentially to determine on which combination of them ANS has the largest influence but to avoid overfitting between training and checking data. Limitations of the proposed method with the ANFIS network could be requirements for more training data to find optimal network parameters. This leads to more computational power and memory space usage.

Section snippets

Materials and method

HRV activity prediction is a very nonlinear regression problem, in which two ANS parameters could be used to predict behaviour of the HRV activity. These two ANS parameters should indicate two branches of the ANS activity, sympathetic and parasympathetic. The ANS indices have different influence on each of the HRV parameters. In this study, 14 parameters of HRV signal were extracted for the analysis. Two parameters characterize the ANS functions. Those are cardiac vagal index (CVI), and cardiac

Results

We performed a comprehensive search within the available inputs to select the set of inputs that most influence the output parameters (Fig. 2). Essentially, the functions build an ANFIS model for each combination and trains it for one epoch and reports the performance achieved. In the beginning, the one most influential input in predicting the output was determined.

The left-most input variable in Fig. 2 has the least error or the most relevance with respect to the output. The plot and results

Conclusion and discussion

Analysis of the heart rate variability provides a non-invasive and sensitive tool for the estimation of autonomic regulation of the heart. Autonomic nervous system (ANS), with its main two divisions; sympathetic and parasympathetic, may be regarded as a hierarchically coordinated neuronal network which controls heart rate continually. The main interest of measuring cardiac sympathetic and parasympathetic activity lies in its prognostic value in cardiovascular risk.

Heart rate variability (HRV)

References (41)

  • S. Papadokonstantakis et al.

    Variable selection and data pre-processing in NN modelling of complex chemical processes

    Computers and Chemical Engineering

    (2005)
  • A. Yardimci

    Soft computing in medicine

    Applied Soft Computing

    (2009)
  • R. Ceylan et al.

    A novel approach for classification of ECG arrhythmias: Type-2 fuzzy clustering neural network

    Expert Systems and Applications

    (2009)
  • H.G. Hosseini et al.

    The comparison of different feed forward neural network architectures for ECG signal diagnosis

    Medical Engineering and Physics

    (2006)
  • E.D. Ubeyli

    Adaptive neuro-fuzzy inference system for classification of ECG signals using Lyapunov exponents

    Computer Methods and Programs in Biomedicine

    (2009)
  • D. Petković et al.

    Adaptive neuro-fuzzy estimation of automatic nervous system parameters effect on heart rate variability

    Neural Computing & Application

    (2011)
  • C.W. Lin et al.

    Mining physiological conditions from heart rate variability analysis

    IEEE Computational Intelligence Magazine

    (2010)
  • G.G. Berntson et al.

    Heart rate variability: Origins, methods, and interpretative caveats

    Psychophysiology

    (1997)
  • M.T. Lin et al.

    Heart rate variability monitoring in the detection of central nervous system complications in children with enterovirus infection

    Journal of Critical Care

    (2006)
  • P. Pancera et al.

    Autonomic nervous system dysfunction in sclerodermic and primary Raynaud’s phenomenon

    Clinical Science

    (1999)
  • Cited by (0)

    View full text