ECG Signal Analysis Method and Its Device

Abstract

The present application proposes a method for electrocardiogram signal analysis, which includes: preprocessing the electrocardiogram signal of a subject to obtain the preprocessed electrocardiogram signal; using 0.04 - 0.4 Hz as the full frequency band and 0.15 - 0.4 Hz as the high frequency band, performing bispectral analysis according to the preprocessed electrocardiogram signal, and calculating the high-low synchronization ratio index - synchHighLow; synchHighLow = #imgabs0#; where #imgabs1# is the coupling energy of the preprocessed electrocardiogram signal in the high frequency band; #imgabs2# is the coupling energy of the preprocessed electrocardiogram signal in the full frequency band. synchHighLow reflects the relative magnitude of the coupling energy between the high frequency band and the full frequency band in the form of the logarithm of the ratio, can intuitively display the migration of the coupling energy of the electrocardiogram signal between high and low frequencies, reveal the activity changes of high and low frequencies, and is helpful for evaluating the balance state of cardiac autonomic nerves.

Description

Technical Field

[0001] This application relates to the analysis technology of electrocardiogram (ECG) signals, and particularly to an ECG signal analysis method and its device. Background Art

[0002] In the field of ECG signal analysis, the estimation of power spectrum density (PSD) based on the RR interval sequence is very common. The LF / HF index extracted from the PSD belongs to the heart rate variability index and is an important index reflecting the balance of the cardiac autonomic nerves. When performing PSD analysis, the low-frequency band (LF) is usually set to 0.04 - 0.15 Hz, and the high-frequency band (HF) is set to 0.15 - 0.4 Hz to obtain the power of the LF and HF bands and the ratio LF / HF between them. Currently, it is generally believed that the LF power reflects the combined regulation of the sympathetic and vagus nerves on the heart rate and is related to the baroreceptor reflex of blood pressure; the HF power is related to the efferent activity of the vagus nerve and is mainly affected by respiratory activity; the LF / HF index can reflect the proportion of high-frequency and low-frequency energies, thereby inferring the balance state of the cardiac autonomic nerves.

[0003] The power spectrum is called the second-order spectrum, which is the Fourier transform of the autocorrelation function of the signal. It is the one-dimensional Fourier transform of the second-order cumulant, and the power spectrum can only reflect the amplitude information at each frequency. The following is a typical application example of PSD estimation in ECG signal analysis:

[0004] (1) Data preprocessing: Input the ECG signal for processing such as denoising and baseline drift removal;

[0005] (2) Perform QRS complex detection and extract the RR interval sequence;

[0006] (3) Perform ectopic beat detection, replacement, and detrending processing on the RR interval sequence;

[0007] (4) Select a PSD estimation method, such as the AR model method, to calculate the PSD;

[0008] (5) Divide the low-frequency band (LF) into 0.04 - 0.15 Hz and the high-frequency band (HF) into 0.15 - 0.4 Hz, and calculate the power of each band and LF / HF;

[0009] (6) For each ECG data, repeat the above five steps, and respectively statistically analyze whether there are significant differences in the average levels of high-frequency and low-frequency powers and LF / HF between the healthy group and the disease group, and draw a conclusion on the autonomic nerve function state of the disease group.

[0010] When processing and analyzing electrocardiogram (ECG) signals, people are accustomed to assuming that the signals or noises follow a Gaussian distribution. However, in reality, most real signals are non-Gaussian. Traditional power spectrum analysis methods lose some useful information of non-Gaussian processes. When actually processing signals, power spectrum analysis has the following disadvantages: 1) Loss of phase information. The LF / HF index only focuses on the ratio of low-frequency and high-frequency powers and only reflects the energy distribution of the signal in different frequency bands. For changes in the phase relationship of cardiac electrical activities, power spectrum indicators are difficult to detect; 2) Inability to obtain non-Gaussian characteristics. The ECG signal is essentially a non-Gaussian signal, and power spectrum indicators cannot extract the non-Gaussian information of the signal, which is not conducive to accurately judging the true electrical activity state of the heart; 3) Weak ability to resist Gaussian noise. During the actual acquisition of ECG signals, noise interference is often accompanied. Power spectrum analysis cannot suppress Gaussian noise, and the accuracy of its calculation results is easily significantly affected by Gaussian noise, leading to deviations in the assessment of the heart state; 4) Inability to obtain non-linear characteristics. Power spectrum analysis belongs to a linear analysis method and cannot extract the non-linear characteristics of the signal. For non-linear systems, the power spectrum cannot provide the non-linear dynamic characteristics of the signal. Summary of the Invention

[0011] In view of the above problems, the present application aims to propose an ECG signal processing method to capture the non-linear characteristics and phase coupling information of the signal. The present application also aims to propose an ECG signal processing method to reflect the balance state of the autonomic nervous system through the analysis results.

[0012] The ECG signal analysis method of the present application includes:

[0013] Preprocess the ECG signal of the subject to obtain the preprocessed ECG signal; take 0.15 - 0.4 Hz as the high-frequency band and 0.04 - 0.4 Hz as the full-frequency band, and perform bispectrum analysis according to the preprocessed ECG signal to calculate the high-low synchronization ratio - synchHighLow;

[0014] synchHighLow = ;

[0015] Wherein, is the coupling energy of the preprocessed ECG signal in the high-frequency band; is the coupling energy of the preprocessed ECG signal in the full-frequency band.

[0016] Preferably, synchHighLow is used as an evaluation index for the balance state of the autonomic nervous system.

[0017] Preferably, the preprocessing includes: denoising, QRS wave detection, ectopic beat detection, ectopic beat replacement, and detrending.

[0018] Preferably, the denoising includes wavelet or adaptive denoising.

[0019] Preferably, the QRS wave detection includes wavelet or differential threshold QRS wave detection.

[0020] Preferably, the ectopic beat detection includes percentage filtering, median filtering, or standard deviation filtering.

[0021] Preferably, the ectopic beat replacement includes mean replacement, median replacement, or difference interpolation replacement.

[0022] Preferably, the detrending includes wavelet detrending or smoothing detrending.

[0023] The electrocardiogram (ECG) signal analysis device of the present application includes:

[0024] A preprocessing unit that preprocesses the input ECG signal to obtain a preprocessed ECG signal;

[0025] A synchHighLow calculation unit that performs bispectral analysis on the preprocessed ECG signal with 0.15 - 0.4 Hz as the high-frequency band and 0.04 - 0.4 Hz as the full-frequency band according to synchHighLow = , and calculates synchHighLow;

[0026] Wherein, is the coupling energy of the preprocessed ECG signal in the high-frequency band; is the coupling energy of the preprocessed ECG signal in the full-frequency band.

[0027] The present application modifies the frequency band of the bispectral synchFastSlow index to a frequency band suitable for ECG analysis, and innovatively obtains the bispectral high-low synchronization ratio index synchHighLow. synchHighLow reflects the relative magnitude of the coupling energy between the high frequency and the full frequency band through the logarithmic form of the ratio, can intuitively show the migration of the coupling energy of the ECG signal between the high and low frequencies, and reveal the activity changes of the high and low frequencies. This index has the advantages of bispectral analysis, can extract the non-linear coupling information of the high and low frequencies of the ECG signal, and is also helpful for evaluating the autonomic nerve balance state, providing a key reference for disease detection and prognosis evaluation. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 is a schematic diagram of the bispectral scope of action.

[0029] Figure 2 is a schematic diagram of the difference between the power spectrum and the bispectrum during phase detection.

[0030] Figure 3 is the application flowchart of the bispectral synchHighLow in the present application.

[0031] Figure 4 The two-dimensional bispectrum graph (left) and three-dimensional bispectrum graph (right) of healthy individuals.

[0032] Figure 5 The two-dimensional bispectrum graph (left) and three-dimensional bispectrum graph (right) of patients with end-stage renal disease.

[0033] Figure 6 Scatter plot of synchHighLow versus LF / HF based on 2h of daytime and nighttime electrocardiogram data.

[0034] Figure 7 Scatter plot of synchHighLow versus LF / HF based on 5min of electrocardiogram data.

[0035] Figure 8 Schematic diagram of the statistical results of synchHighLow in healthy individuals, patients with chest pain due to acute coronary syndrome, and patients with acute myocardial infarction.

[0036] Figure 9a ROC curve graph for synchHighLow to distinguish between healthy individuals and patients with chest pain due to acute coronary syndrome.

[0037] Figure 9b ROC curve graph for synchHighLow to distinguish between healthy individuals and patients with acute myocardial infarction.

[0038] Figure 10a Schematic diagram of the statistical results of synchHighLow in healthy individuals and patients with end-stage renal disease.

[0039] Figure 10b ROC curve graph for synchHighLow to distinguish between healthy individuals and patients with end-stage renal disease. Detailed implementation manner

[0040] Next, the present application will be described in detail with reference to the accompanying drawings.

[0041] The bispectrum has symmetry and periodicity

[0042] Regarding Symmetry:

[0043]

[0044] Regarding Or Symmetry:

[0045]

[0046] This is due to the periodicity of the Fourier transform, from which the periodicity of the bispectrum can be known:

[0047]

[0048] The original scope of the bispectrum is , and from the symmetry and periodicity of the above-mentioned bispectrum, it can be obtained that this scope is symmetric about , , ， symmetric, and periodically repeated about . As shown in Figure 1 , the original scope is the entire large square, which is divided into sixteen parts by the above-mentioned symmetry axes. That is, as long as the bispectrum information in the original scope is one-sixteenth of the original, it can be completely included. In this way, the computational complexity of the algorithm is reduced to one-sixteenth of the original.

[0049] The bispectrum can detect phase coupling information

[0050] Consider two signals:

[0051]

[0052]

[0053] It can be found that for and , their power spectra are exactly the same. In the bispectrum, the bispectrum of is 0, and the bispectrum of Figure 2 shows a pulse at (1, 2), that is, there is quadratic phase coupling (see

[0054] The electrocardiogram (ECG) signal analysis method of this application innovatively changes the synchFastSlow of bispectrum analysis to synchHighLow suitable for ECG signal analysis, referring to the remarkable success of the electroencephalogram (EEG) fast-slow synchronization ratio (synchFastSlow) index in anesthesia depth monitoring.

[0055] The bispectrum is also called the third-order spectrum. The third-order spectrum is a two-dimensional Fourier transform of the third-order autocorrelation function of the signal. As the simplest and most practical high-order spectrum, the bispectrum is simple to calculate and has all the advantages of high-order spectrum analysis. The bispectrum is a two-dimensional complex spectrum, and its peak represents the phase coupling strength of these two frequency components of the signal, characterizing the non-Gaussian characteristics of the signal.

[0056] The definition of the third-order spectrum can be further deduced from the definition of the second-order statistic and the second-order spectrum. For a zero-mean stationary process X(n), first obtain the third-order statistic (third-order autocorrelation function):

[0057]

[0058] The third-order spectrum is the two-dimensional Fourier transform of the third-order statistics:

[0059]

[0060] where: ,

[0061] When the signal is a finite-length sequence, the bispectrum estimate can be obtained through derivation:

[0062]

[0063] where is a finite-length discrete random signal of length N is the Fourier transform of:

[0064]

[0065] One of the most significant applications of bispectrum analysis is the bispectral index (BIS) for analyzing the depth of anesthesia. BIS is currently the most widely used index for evaluating the depth of anesthesia and sedation in clinical practice. synchFastSlow is a BIS subparameter applied to detect the depth of anesthesia, which can reflect the nonlinear synchronization (non-Gaussian coupling) of fast and slow frequency band neural oscillations, thus providing information on the depth of anesthesia. In the BIS index, the definition of synchFastSlow is:

[0066]

[0067] where,

[0068]

[0069]

[0070]

[0071]

[0072] In electroencephalogram (EEG) analysis, usually set representing the entire frequency band (full frequency band) to be 0.5 - 47 Hz, and representing the high-frequency band to be 40 - 47 Hz.

[0073] In the present invention, the frequency bands in the bispectrum synchFastSlow are changed to the full frequency band suitable for electrocardiogram (ECG) analysis (LF1 - HF2 set as 0.04 - 0.4 Hz) and the high - frequency band (HF1 - HF2 set as 0.15 - 0.4 Hz), thereby obtaining a new bispectrum high - low synchronization ratio index - synchHighLow.

[0074] The definition of synchHighLow is:

[0075] synchHighLow =

[0076] The meaning of synchHighLow is the logarithm of the ratio of the coupling energy in the high - frequency band to the coupling energy in the full frequency band. As can be seen from the definition, when synchHighLow increases, it indicates that the coupling energy of the ECG signal migrates from low frequency to high frequency, that is, the high - frequency activity increases; when synchHighLow decreases, it indicates that the coupling energy of the ECG signal migrates from high frequency to low frequency, that is, the low - frequency activity increases.

[0077] synchHighLow has all the advantages of the bispectrum index, can extract the high - and low - frequency non - linear coupling information of the ECG signal, and helps to better evaluate the autonomic nerve state. Introducing bispectrum - related indexes in ECG signal analysis is beneficial to extracting more abundant and reliable information, providing important references for the detection and prognosis evaluation of diseases.

[0078] To measure the effect of the present invention, healthy people and three groups of diseased people are used to conduct the correlation analysis between synchHighLow and LF / HF, the difference analysis of synchHighLow between healthy people and patient groups, and the effectiveness analysis of synchHighLow in detecting autonomic neuropathy respectively.

[0079] Data

[0080] The experimental data were from THEW database (Telemetric and Holter ECG Warehouse, HYPERLINK"http: / / www.thew-project.org" http: / / www.thew-project.org), including 24-hour Holter electrocardiogram data of multiple populations. Four sub-databases were selected. One was the Normal sub-database (E-HOL-03-0202-003), including electrocardiogram data of 202 normal people; the second was the AMI sub-database (E-HOL-03-0160-001), including electrocardiogram data of 93 patients with acute myocardial infarction (AMI); the third was the CPP sub-database (E-HOL-12-1172-012), including electrocardiogram data of 802 patients with chest pain of acute coronary syndrome (CPP); the fourth was the ESRD sub-database (E-HOL-12-0051-016). The selected individuals were all end-stage renal disease (ESRD) patients undergoing hemodialysis treatment. They also had hypertension or diabetes and were high-risk populations with autonomic dysfunction abnormalities.

[0081] Correlation analysis of synchHighLow and LF / HF

[0082] The Normal sub-database representing the normal population was selected. All 24-hour Holter electrocardiogram records were manually analyzed through Kubios software. Two-hour relatively quiet data segments were manually selected from daytime (07:00–20:00) and nighttime (20:00-07:00) for analysis. In addition, 5-minute data segments without movement or napping were extracted from each electrocardiogram record (between 8 am and 5 pm). The correlation analysis of synchHighLow and LF / HF was carried out from two perspectives: short-term and long-term, daytime and nighttime.

[0083] Calculation of synchHighLow:

[0084] (1) For a certain electrocardiogram data, preprocessing operations were performed, including filtering, denoising, and baseline correction.

[0085] (2) QRS wave detection was performed, and the RR interval was obtained by differential calculation.

[0086] (3) Detection and replacement of RR interval ectopics, detrending.

[0087] (4) Divide the two main frequency bands in the bispectrum domain: the full frequency band LF1-HF2 is defined as 0.04 - 0.4 Hz, and the high frequency band HF1-HF2 is defined as 0.15 - 0.4 Hz, and calculate synchHighLow;

[0088] (5) Repeat the above four steps to obtain synchHighLow for all people.

[0089] (6) Conduct statistical analysis of the correlation, and the results are as Figure 6 shown. The left is the data result during the day, and the right is the data result at night.

[0090] In the 2h daytime and nighttime electrocardiogram data, the scatter plot of synchHighLow and LF / HF is as Figure 6 shown. It can be seen that synchHighLow and LF / HF generally show a linear correlation trend. The Spearman correlation analysis shows that the correlation coefficients of the two are -0.766 ( p = 0.00) and -0.797 ( p = 0.00), indicating that synchHighLow and LF / HF have a strong correlation. In the 5min electrocardiogram data, synchHighLow and LF / HF also show a linear correlation trend (see Figure 7 ), and the Spearman correlation coefficient is -0.883 ( p = 0.00), indicating that synchHighLow and LF / HF have a strong correlation. It can be seen that regardless of whether it is 2h or 5min data, regardless of whether it is daytime or nighttime data, synchHighLow and LF / HF show a significant strong correlation (especially for short-term data). Therefore, synchHighLow is very likely to be a supplementary indicator that can reflect the autonomic nerve balance state in addition to LF / HF.

[0091] Analysis of the differences in synchHighLow between healthy people and patient populations

[0092] Select Normal as the control group, and the CPP and AMI sub-databases represent the experimental groups. For each record, 5-minute time period data is selected for analysis.

[0093] The synchHighLow levels of each group are represented by the median and interquartile range. The statistical results show that compared with the healthy control group (-2.48, 0.95), the synchHighLow value of CPP patients (-2.14, 0.96) is significantly increased ( p = 0.00), and the synchHighLow value of AMI patients (-2.11, 0.86) is also significantly increased ( p= 0.00), there was no significant difference in synchHighLow between the CPP and AMI pathological groups (see Figure 8 ). This result proves that pathological factors can lead to enhanced signal synchronization and increased coupling strength in the high-frequency band, and thus show the phenomenon that the synchHighLow of patients is significantly larger than the normal value.

[0094] ROC analysis showed (see Figure 9a , 9b ) that the AUC of synchHighLow for differentiating CPP patients from healthy people was 0.667, while the AUC of LF / HF was 0.619; the AUC of synchHighLow for differentiating AMI patients from healthy people was 0.672, while the AUC of LF / HF was 0.648. It can be seen that the ability of synchHighLow to distinguish the disease group from normal people is significantly better than that of LF / HF, and synchHighLow shows potential in disease detection.

[0095] Analysis of the effectiveness of synchHighLow in detecting autonomic neuropathy

[0096] Normal was selected as the control group, and the ESRD sub-database represented the pathological group of autonomic neuropathy. After removing the records with incomplete day-night data, 189 records from Normal and 43 records from ESRD were selected respectively. Manually select a 2-hour relatively quiet data segment from daytime (07:00–20:00) for analysis.

[0097] The ability of synchHighLow to distinguish ESRD patients from healthy people was evaluated by ROC curve analysis, and performance indicators such as the area under the curve (AUC) and sensitivity were obtained, and its performance was compared with that of LF / HF. In addition, in order to verify whether the combination of synchHighLow and LF / HF can improve the ability of a single indicator to detect autonomic neuropathy, binary logistic regression analysis was applied to obtain the predicted probabilities of the synchHighLow and LF / HF indicators, and then ROC analysis was performed on the predicted probabilities to observe whether the disease detection performance was improved.

[0098] The results showed (see Figure 10a , 10b ) that compared with the healthy control group (-2.44, 0.81), the synchHighLow value (-1.70, 0.74) of ESRD patients was significantly increased ( p(<0.05), indicating that autonomic neuropathy leads to enhanced signal synchronization and increased coupling strength in the high-frequency band. The results of ROC analysis showed that the AUC of synchHighLow for differentiating healthy individuals from ESRD was 0.828, much larger than the AUC of LF / HF (0.743), indicating that synchHighLow has a stronger ability to detect autonomic neuropathy than LF / HF. When -2.346 was selected as the threshold (with the maximum Youden index), the diagnostic sensitivity reached 0.977. The AUC of the predicted probabilities of synchHighLow and LF / HF was 0.839, greater than the AUC of either synchHighLow or LF / HF alone, indicating that the combined prediction model of synchHighLow and LF / HF can better detect autonomic neuropathy. synchHighLow provides useful information not contained in LF / HF, and their combined use is likely to show excellent effects in the detection and prognostic evaluation of autonomic diseases.

Claims

1. An electrocardiogram signal analysis method, comprising: Preprocessing the electrocardiogram signal of a subject to obtain a preprocessed electrocardiogram signal; Taking 0.15 - 0.4 Hz as the high-frequency band and 0.04 - 0.4 Hz as the full-frequency band, performing bispectrum analysis based on the preprocessed electrocardiogram signal, and calculating synchHighLow; synchHighLow = ; Among them, is the coupling energy of the preprocessed electrocardiogram signal in the high-frequency band; is the coupling energy of the preprocessed electrocardiogram signal in the full frequency band.

2. The electrocardiogram signal analysis method according to claim 1, wherein: synchHighLow is used as an evaluation index for the balance state of the cardiac autonomic nerves.

3. The electrocardiogram signal analysis method according to claim 1, wherein: The preprocessing includes: denoising, QRS wave detection, ectopic beat detection, ectopic beat replacement, and detrending.

4. The electrocardiogram signal analysis method according to claim 3, wherein: The denoising includes wavelet or adaptive denoising.

5. The electrocardiogram signal analysis method according to claim 3, wherein: The QRS wave detection includes wavelet or differential threshold QRS wave detection.

6. The electrocardiogram signal analysis method according to claim 3, wherein: The ectopic beat detection includes percentage filtering, median filtering, or standard deviation filtering.

7. The electrocardiogram signal analysis method according to claim 3, wherein: The ectopic beat replacement includes mean replacement, median replacement, or difference interpolation replacement.

8. The electrocardiogram signal analysis method according to claim 3, wherein: The detrending includes wavelet detrending or smoothing detrending.

9. An electrocardiogram signal analysis device, comprising: A preprocessing unit that preprocesses the input electrocardiogram signal to obtain a preprocessed electrocardiogram signal; The synchHighLow calculation unit, where synchHighLow = , performs bispectrum analysis on the preprocessed electrocardiogram signal with a high-frequency band of 0.15 - 0.4 Hz and a full-frequency band of 0.04 - 0.4 Hz, and calculates synchHighLow; Among them, is the coupling energy of the preprocessed electrocardiogram signal in the high-frequency band; is the coupling energy of the preprocessed electrocardiogram signal in the full frequency band.

10. The electrocardiogram signal analysis device according to claim 9, wherein: synchHighLow is used as an evaluation index for the balance state of the cardiac autonomic nerves.

Images