Comparative Analysis of Heart Rate Variability Parameters for Arrhythmia and Atrial Fibrillation using ANOVA

Heart Rate Variability (HRV) is an important criterion to check the cardiac health. Sudden HRV signifies the unhealthy condition of the heart, particularly when the person is suffering from a cardiac disease. HRV parameters on different patients of different ages, gender and health conditions are observed using time domain, geometrical domain and frequency domain. Statistical comparison is done on three different databases MIT/BIH Normal Sinus Rhythm (NSR), MIT/BIH Arrhythmia (AR) and MIT/BIH Atrial Fibrillation (AF) using Analysis of Variance (ANOVA) technique. Features of all the three domains show weak, moderate or strong significant changes as per the relation during comparison with respective databases. The best features were selected among the various HRV features which will be used for classification in future.

Electrocardiography (ECG) is a noninvasive method to extract electrical signals produced by Sino-Atrial (SA) node in the heart.By monitoring the ECG of a patient, prediction can be made which helps the doctor to take effective steps clinically 1,2 .A significant decrease in the mortality rate of sudden heart attack cases can be achieved by regularly monitoring the ECG.ECG signal consists of several features delineating from P, QRS and T wave.Among all these features Heart Rate (HR) is an important feature to diagnose the cardiac health 2 .HR is calculated using the RR interval (the difference between two consecutive R-peaks).RR interval is the most prominent feature of the HRV, while other features are directly or indirectly dependent on it.HRV components can be calculated in the time domain, geometrical domain and frequency domain 3 .
Various platforms are available nowadays for the collection of the database (offline and online) 4 .Firstly, pre-processing is done on the database to eliminate the undesired signal from the useful signal 5 .Generally, two types of unwanted signals are present which are Power Line Interference (PLI) and Baseline Wander (BLW).To remove these noises, Finite Impulse Response (FIR) or Infinite Impulse Response (IIR) digital filters are used.IIR Butterworth filter is used to diminish these noises 6 .The same filter of 4th order is used for pre-processing 7 .
Butterworth filter is an IIR filter and it is not hardware implementable due to its unstable nature.FIR filter to is used to remove these unexpected signals due to ease of its hardware implementable nature 8 .Secondly, feature extraction is doneto extract the parameter of HRV for different duration and domains 2,9 .There are various statistical tests which are used for the comparison in between and within groups like t-test, f-test, ANOVA etc.If the comparison is in between two groups then the t-test or f-test is used, while if more than two groups are there then ANOVA test is preferred.If we use the t-test or f-test on more than two groups, Type-I error will be introduced, which is unacceptable during statistical test performance.ANOVA test is a statistical test used for comparison or finding a relationship between or within features.Extraction of time and frequency domain features of HRV for the stress level evaluation is also used to executethe ANOVA test for the statistical comparison among different level of stress 10 .ANOVA test on HRV features for the evaluation between meditative state and non-meditative state of a person is used 11 .
In this paper, we have used three online databases of MIT/BIH Normal Sinus Rhythm (NSR), MIT/BIH Arrhythmia (AR) and MIT/BIH Atrial Fibrillation (AF).From these databases of ECGs, removal of artifact is done by using digital FIR Kaiser Window filter because it has better frequency response than other conventional windows.Later from the filtered ECG signal, we have extracted time domain, geometrical domain and frequency domain HRV features.In the end, comparison of different databases on the basis of best features is done by using ANOVA test.
The paper is structured as: Section 2 describes the different processing steps to be applied on different databases like MIT/BIH NSR, MIT/BIH AR and MIT/BIH, Section 3 discusses the results of the different processing steps and conclusion are presented in Section 4.

Methodology
The process for the statistical analysis of feature processing consists primarily of preprocessing, feature extraction, feature selection and classification phase 12,13 .In this paper, we are concentrating on the first three steps 14,15 .Figure 1 illustrates the block diagram of HRV feature processing.

database Collection
The online platform of Physionet is used for database collection.These databases are MIT/ BIH Normal Sinus Rhythm (NSR), MIT/BIH Arrhythmia (AR) and MIT/BIH Atrial Fibrillation (AF) with 18 records each.The sampling frequency of NSR, AR and AF database are 128 Hz, 360 Hz, and 250 Hz respectively.The database contains measurable data of the interval scale so, all the statistical parameters can be applied on it.eCg Pre-processing Several noises in ECG signals are introduced during the acquisition of the signal from the patient body.Among them, Power Line Interference (PLI) and Base Line Wander (BLW) are of main concern 16 .These two unwanted signals create hindrance while extracting the useful information from the ECG wave.Pre-processing is done using a digital filter to diminish these undesirable signals 17 .In this paper, FIR digital filters are used due to its simplicity in hardware implementation.Among various FIR filters, FIR Kaiser Window is used because it gives better frequency response as compared to other windowing technique.Useful bandwidth of ECG signal lies in between 0.5 Hz to 50 Hz.PLI is a highfrequency noise introduced due to the biomedical instrument used for the acquisition of ECG signal.PLI is eliminated using FIR Kaiser Low Pass Filter (LPF) having cut off frequency ('f c ') of 50 Hz.Respiration during signal acquisition is the main cause of BLW containing low-frequency noise.Removal of BLW can be done be using FIR Kaiser High Pass Filter (HPF), having 'f c ' of 0.5 Hz.Later, we have obtained a denoised/filtered signal which is used for HRV parameter extraction.

R-peak detection
ECG signal consists of mainly three waves P, QRS and T.Among these QRS wave is the prime focus, which can be further divided into three subwaves Q, R and S. R wave is the most prominent feature of QRS due to its high positive peak.In this paper, the Pan-Tompkins Algorithm for QRS and R peak detection is used 18 .For QRS detection, squaring and moving window integration is used.Since R wave is a positive peak, so squaringis done to eliminate all the negative peaks.Adaptive Thresholding is used for extraction of the R-peak from QRS wave after eliminating abnormal beats.

hRV Feature extraction
Feature extraction from the R-peak of a filtered signal is the essential part for HRV feature processing 17 .HRV features are divided into three domains :time domain, geometric domain and frequency domain.Spectrum for RR interval in time In Table 1

explanation of different domain features with their description and formulas for calculating them is explained. hRV Feature Selection
A number of HRV features are present in various domains as defined but only some of them shows a noteworthy effect.That's why correlation among different HRV features of various domains expressed at two significance levels (95% and 99%) is shown in this paper.HRV parameters which show high correlation factor get selected while others are discarded.For the comparison of three groups of ECG databases, ANOVA test is used in different domains.ANOVA test demonstrates the variation among the mean of HRV parameters for different databases.

ReSult and diSCuSSion
The input signal of ECG from different databases is used in this work which is demonstrated in figure 3. Fig. 3a represents the NSR ECG wave, Fig. 3b shows the AR ECG wave and Fig. 3c illustrates the AF ECG wave.
In Pre-processing stage FIR Kaiser Window is used to denoise the signal from unwanted noises like PLI and BLW using LPF and HPF respectively.Fig. 4a shows the removal of PLI using LPF and Fig. 4b represents the removal of BLW using HPF.
After getting filtered output, the Pan-Tompkins algorithm is used for real-time QRS detection.This algorithm is used for extracting the R-peak of QRS wave.R-peak is the most important The feature values can be affected by the sampling frequency and waveform duration of the database taken while processing.In this paper, we have generated the results at 1000 Hz sampling frequency for a long-term database.After observing the results of ANOVA test for these specifications, it is interpreted that some HRV features like SDSD, RMSSD, NN50, pNN50, aULF, aVLF, aHF, and aTotal gives a strong indication to reject H o .It represents that these features show a significant variation in their mean value.While, all other parameters show weak evidence against to reject H o , except mean HR which represents moderate significant change.Mostly HRV feature represents a strong significant change which denotes that there is a substantial difference occurring among different databases.After analyzing the p-value of ANOVA which signifies the noticeable variation among different databases, we have selected the best HRV features.Best selected features are SDSD, RMSSD, NN50, pNN50, mean HR aULF, aVLF, aHF and aTotal which we will be used for further classification in the future.

ConCluSion
This paper presents the statistical comparison among HRV features of different domain and different databases.FIR Kaiser digital filter is used to eliminate noise from NSR, AR and AF database.HRV feature extraction is implemented in time, geometrical and frequency domain, further for feature selection correlation is used.For the comparison between the three databases, ANOVA test has been used which yields the best features which will be used for classificationin future.The hardware implementation of these statistical tests will be done on FPGA using VIVADO tool as extension of this work in future.

Fig. 1 .
Fig. 1.Block diagram of methodology of HRV Feature processing

table 2 .
Comparison of HRV features using ANOVA