Multiple arrhythmia signal classification methods based on spectral analysis

Abstract

The present application relates to the technical field of arrhythmia classification, and discloses a plurality of arrhythmia signal classification methods based on spectrum analysis. The method comprises collecting target object electrocardio signal data and preprocessing, generating standardized electrocardio signal sequence; performing multi-scale time-frequency transformation on the sequence, extracting time-frequency domain feature matrix, and constructing spectrum energy distribution atlas and phase change trajectory atlas accordingly. Based on the former, the dominant frequency band interval is divided and the energy proportion weight is marked; according to the latter, the phase offset of adjacent cardiac cycles is calculated, and the phase stability index is generated. Fusion of the two generates the initial classification feature vector, and the adjacent signal segment space correlation is dynamically compensated to generate the optimized feature vector. Matching the preset arrhythmia type feature library determines the candidate set, and then the feature similarity level is screened to output the final classification result, which can realize accurate classification of a plurality of arrhythmia signals.

Description

Technical Field

[0001] This invention relates to the field of cardiac arrhythmia classification technology, specifically to a method for classifying various cardiac arrhythmia signals based on spectral analysis. Background Technology

[0002] In the field of cardiovascular disease diagnosis, electrocardiogram (ECG) signals contain crucial information about cardiac activity, and the accurate classification of arrhythmia signals plays a vital role in early disease screening, condition assessment, and treatment planning. With the advancement of medical technology, the clinical requirements for the timeliness and accuracy of arrhythmia signal classification are constantly increasing, but existing classification methods still have many limitations and cannot meet the needs of practical applications.

[0003] Existing methods for classifying cardiac arrhythmia signals mostly rely on single-scale time-frequency analysis techniques, which can only capture the characteristics of electrocardiogram signals in a specific time or frequency dimension, and cannot comprehensively reflect the complex time-frequency variation patterns of different types of cardiac arrhythmia signals. For example, some methods use short-time Fourier transform for time-frequency analysis. The time resolution and frequency resolution of this transform are mutually constrained. For cardiac arrhythmia signals with sudden heart rate changes or irregular rhythms, it is difficult to accurately extract their instantaneous frequency and energy change characteristics, which can easily lead to feature omissions in subsequent classification processes.

[0004] Existing methods often suffer from limitations in feature extraction, with most focusing only on a single dimension of the electrocardiogram (ECG) signal's energy or phase characteristics. Classification based solely on energy features ignores the cardiac cycle rhythm variations inherent in phase information, while the differences between different types of arrhythmias (such as atrial fibrillation and premature ventricular contractions) often lie in their varying phase stability. Conversely, relying solely on phase features fails to reflect the differences in signal energy distribution across different frequency bands, and variations in spectral energy distribution are crucial for distinguishing between sinus rhythm and ectopic rhythms. This single-feature extraction approach easily leads to incomplete classification features, affecting the reliability of the classification results.

[0005] Existing methods lack utilization of the spatial correlation between adjacent signal segments and fail to perform dynamic compensation processing during feature vector construction. ECG signals are continuous, and the features of adjacent signal segments are inherently correlated. However, the initial feature vector is generated only based on the current signal segment, without considering this correlation. This makes it susceptible to interference factors such as signal noise and baseline drift, resulting in insufficient stability of the feature vector. In the subsequent matching and filtering stage, existing methods mostly use a single similarity threshold for judgment without hierarchical filtering. When multiple similar arrhythmia types exist, it is difficult to accurately distinguish them, leading to misclassification. Especially in complex scenarios where multiple arrhythmias coexist, the classification accuracy drops significantly.

[0006] Existing methods rely heavily on small sample data for feature library construction, resulting in poor adaptability to ECG signals from target subjects of different ages and physical conditions. When faced with diverse arrhythmia signals in clinical settings, the matching degree between the pre-set feature library and the actual signal features decreases, further affecting classification performance. These issues collectively limit the clinical application of existing arrhythmia signal classification methods, necessitating a classification method capable of comprehensively extracting features, dynamically optimizing vectors, and accurately filtering results to improve the overall performance of arrhythmia signal classification. Summary of the Invention

[0007] The purpose of this invention is to provide a variety of arrhythmia signal classification methods based on spectral analysis to solve the problems mentioned in the background art.

[0008] To achieve the above objectives, this invention provides a method for classifying various arrhythmia signals based on spectral analysis, the method comprising:

[0009] Collect and preprocess the electrocardiogram (ECG) signal data of the target object to generate a standardized ECG signal sequence;

[0010] Multi-scale time-frequency transformation is performed on standardized electrocardiogram signal sequences to extract time-frequency domain feature matrices;

[0011] Construct a spectral energy distribution map and a phase change trajectory map based on the time-frequency domain feature matrix;

[0012] The dominant frequency bands are divided based on the spectral energy distribution map, and the energy proportion weight of each frequency band is marked.

[0013] The phase shift between adjacent cardiac cycles is calculated based on the phase change trajectory map to generate a phase stability index.

[0014] By integrating energy proportion weights and phase stability indices, an initial arrhythmia classification feature vector is generated.

[0015] The initial arrhythmia classification feature vector is dynamically compensated by the spatial correlation of adjacent signal segments to generate an optimized classification feature vector.

[0016] Based on the matching degree between the optimized classification feature vector and the preset arrhythmia type feature library, a set of candidate arrhythmia types is determined;

[0017] Based on the feature similarity of each type in the candidate arrhythmia type set, hierarchical screening is performed to output the final arrhythmia classification result.

[0018] Preferably, the specific implementation process of performing multi-scale time-frequency transformation on the standardized electrocardiogram signal sequence includes:

[0019] A sliding window with variable window length is used to segment and extract standardized electrocardiogram signal sequences.

[0020] Perform windowed Fourier transform and wavelet packet decomposition on each signal segment;

[0021] The Fourier transform results and wavelet packet decomposition coefficients are concatenated in frequency band order to form a time-frequency domain feature matrix.

[0022] The length of the sliding window is dynamically adjusted according to the R-wave interval within the signal segment.

[0023] Preferably, the process of constructing the spectral energy distribution map includes:

[0024] Extract the energy amplitude of each frequency band from the time-frequency domain feature matrix;

[0025] An energy distribution curve is generated with frequency band as the horizontal axis and energy amplitude as the vertical axis.

[0026] The energy distribution curve is Gaussian smoothed and normalized.

[0027] Peak points in the energy distribution curve that exceed three standard deviations above the global mean are designated as the dominant frequency bands.

[0028] Preferably, the calculation process of the phase change trajectory map includes:

[0029] Extract the fundamental frequency phase angle sequence from the time-frequency domain feature matrix;

[0030] The phase angle difference between adjacent R-wave peak points is calculated as the instantaneous phase offset;

[0031] A phase change trajectory is generated with the cardiac cycle number as the horizontal axis and the instantaneous phase shift as the vertical axis;

[0032] The phase change trajectory is linearly fitted, and the root mean square of the fitting residual is used as the phase stability index.

[0033] Preferably, the process of generating the initial arrhythmia classification feature vector includes:

[0034] The energy proportion weights of each dominant frequency band are arranged from low to high according to the frequency band to form an energy feature sub-vector;

[0035] The phase stability index is combined with the average R-wave interval to form a phase characteristic sub-vector;

[0036] The energy feature vector and the temporal feature vector are z-score standardized and then concatenated.

[0037] The QT interval variation coefficient is added as an additional dimension to the concatenated vector.

[0038] Preferably, the specific implementation steps for dynamically compensating the initial arrhythmia classification feature vector based on the spatial correlation of adjacent signal segments include:

[0039] The classification feature vectors of two cardiac cycles before and after the current signal segment are selected as the reference vector set;

[0040] Calculate the cosine similarity between the feature vector of the current signal segment and each reference vector;

[0041] Compensation weight coefficients are assigned based on the similarity level;

[0042] Perform a convex combination operation between the weighted average of the reference vector set and the current feature vector.

[0043] Preferably, the process of determining the candidate arrhythmia type set includes:

[0044] Calculate the Mahalanobis distance between the optimized classification feature vector and the template vector of each class in the feature library;

[0045] Template types with a Mahalanobis distance less than a preset threshold are selected as the primary candidate set;

[0046] Outliers that are more than twice the median distance from other types in the primary candidate set are removed.

[0047] The remaining types are sorted in ascending order of distance to generate a set of candidate arrhythmia types.

[0048] Preferably, the implementation process of hierarchical screening based on the feature similarity of each type in the candidate arrhythmia type set includes:

[0049] Extract the top three main candidate types from the set of candidate arrhythmia types;

[0050] Calculate the spectral energy distribution similarity and phase trajectory morphology similarity of the main candidate types respectively;

[0051] When both similarity metrics exceed their respective thresholds, the secondary feature verification process is triggered.

[0052] In the secondary feature verification, the slope of the ST segment and the symmetry index of the T wave are compared;

[0053] The final arrhythmia classification result was revised based on the verification results.

[0054] Preferably, the secondary feature verification process includes:

[0055] Perform ST-T segment separation processing on the current signal segment to obtain independent ST segment and T wave signals;

[0056] Calculate the absolute value of the average slope of segment ST within 80ms after point J;

[0057] Extract the symmetry index on both sides of the T-wave apex;

[0058] The absolute value of the slope is combined with the symmetry index to form a verification feature pair;

[0059] Compare and verify the matching degree of feature pairs with the typical value range of each major candidate type.

[0060] Preferably, the method further includes:

[0061] After outputting the final classification result, the optimized classification feature vector of the current signal segment is added to the feature library;

[0062] When the number of similar samples in the feature library exceeds a preset limit, the template vector update process is triggered.

[0063] The feature vectors of similar samples are recalculated using a clustering algorithm;

[0064] The feature library is updated by replacing the original template vector with the new cluster centers.

[0065] Compared with the prior art, the beneficial effects of the present invention are:

[0066] In the signal preprocessing stage, this method preprocesses the acquired raw ECG signal data and generates a standardized ECG signal sequence, which can eliminate irrelevant information such as noise, baseline drift, and electromyographic interference contained in the raw signal. At the same time, it unifies the format and amplitude range of ECG signals from different sources and under different acquisition conditions, so that subsequent time-frequency analysis and feature extraction can be carried out on a consistent signal basis, avoiding feature extraction deviations caused by signal differences, and making the analysis in each subsequent stage more targeted and effective.

[0067] Multi-scale time-frequency transform is used to extract time-frequency domain feature matrices from standardized electrocardiogram (ECG) signal sequences. Compared to existing single-scale time-frequency analysis techniques, this method can capture the changing characteristics of ECG signals at different time and frequency scales. Different types of arrhythmia signals exhibit different behaviors in the time-frequency domain, and some arrhythmia signal features only appear at specific scales. Multi-scale time-frequency transform can comprehensively cover these features, fully preserving the time-frequency domain information of ECG signals and providing rich data support for subsequent atlas construction and feature calculation.

[0068] By constructing a spectral energy distribution map and a phase change trajectory map based on the time-frequency domain feature matrix, the energy and phase characteristics of electrocardiogram (ECG) signals can be captured simultaneously. The spectral energy distribution map can intuitively present the energy distribution of the signal in different frequency bands, while the phase change trajectory map can reflect the phase evolution of the cardiac cycle. The two maps complement each other, avoiding the one-sidedness of existing methods that only focus on a single feature dimension, making the extracted features more comprehensive and better reflecting the essential differences between different arrhythmia signals.

[0069] By dividing the dominant frequency bands based on the spectral energy distribution map and assigning weights to their energy proportions, the frequency band information that plays a crucial role in classifying arrhythmias can be highlighted. The energy of different types of arrhythmia signals is often concentrated in specific frequency bands. By dividing the dominant frequency bands and assigning weights, the feature contribution of key frequency bands can be strengthened, while the interference of irrelevant frequency bands can be weakened. This makes the focus more prominent in the subsequent feature fusion process and improves the ability of feature vectors to identify arrhythmia types.

[0070] Calculating the phase shift between adjacent cardiac cycles based on phase change trajectory maps and generating a phase stability index can accurately reflect the stability of cardiac rhythm. One of the core manifestations of arrhythmia is the abnormality of cardiac rhythm. The phase shift directly reflects the change in the time interval and phase relationship between adjacent cardiac cycles, while the phase stability index quantifies this change, effectively distinguishing between regular and irregular arrhythmia types and providing an important basis for differentiating different types of arrhythmias.

[0071] By fusing energy proportion weights and phase stability indices to generate an initial arrhythmia classification feature vector, an organic combination of energy and phase features is achieved. These two features reflect the attributes of arrhythmia signals from different dimensions. The fused initial feature vector contains more comprehensive classification information and, compared to a single feature vector, can more accurately characterize the feature differences between different arrhythmia signals, laying a solid foundation for subsequent classification.

[0072] Dynamic compensation of the initial feature vector is achieved by leveraging the spatial correlation between adjacent signal segments, fully utilizing the continuous nature of ECG signals. The features of adjacent signal segments are inherently correlated; dynamic compensation can correct deviations in the initial feature vector caused by noise and interference based on this correlation, while simultaneously updating the dynamic characteristics of the signal over time. This results in an optimized feature vector that better reflects the actual changes in ECG signals, improving the stability and accuracy of the feature vector.

[0073] Matching the optimized feature vectors with a pre-defined feature library of arrhythmia types to determine the candidate set, and then performing hierarchical filtering based on feature similarity, can significantly improve the accuracy of classification results. The pre-defined feature library covers typical features of various arrhythmia types, and the matching process can quickly narrow down the classification range. Hierarchical filtering breaks through the limitations of existing single threshold judgments, gradually eliminating types with low similarity through multiple rounds of similarity comparison, accurately locating the most suitable arrhythmia type. Especially in scenarios where multiple arrhythmia features are similar, it can effectively reduce misjudgments and ensure the reliability of classification results. Attached Figure Description

[0074] Figure 1 This is a schematic diagram illustrating the working principle of the multi-arrhythmia signal classification method based on spectrum analysis described in this invention.

[0075] Figure 2 A flowchart for constructing a spectral energy distribution map;

[0076] Figure 3 The flowchart for calculating the phase change trajectory map. Detailed Implementation

[0077] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0078] Please see Figure 1 This invention provides a method for classifying various arrhythmia signals based on spectral analysis. The method includes: processing a standardized electrocardiogram (ECG) signal sequence through multi-scale time-frequency transformation to extract a time-frequency domain feature matrix containing rich time-frequency information. Based on the time-frequency domain feature matrix, two key maps are further constructed: a spectral energy distribution map and a phase change trajectory map. The spectral energy distribution map is used to divide the dominant frequency bands and mark the energy proportion weight of each band; the phase change trajectory map is used to calculate the phase shift between adjacent cardiac cycles, thereby generating a phase stability index characterizing rhythm stability. The energy proportion weight and the phase stability index are fused together to form an initial arrhythmia classification feature vector. The initial arrhythmia classification feature vector is dynamically compensated and optimized by analyzing the spatial correlation between adjacent signal segments to generate an optimized classification feature vector. The optimized classification feature vector is then matched with a preset arrhythmia type feature library to determine a candidate arrhythmia type set. The final arrhythmia classification result is output after hierarchical filtering of the feature similarity of each type within the candidate arrhythmia type set.

[0079] Example 1: See Figure 2The specific implementation process of multi-scale time-frequency transformation of standardized electrocardiogram signal sequences involves segmenting the signal sequence using a sliding window with a variable window length. The window length of the sliding window is dynamically adjusted according to the R-wave interval within the signal segment. This adjustment mechanism allows the window length to adapt to changes in heart rate. When the heart rate is faster and the R-wave interval is shorter, a shorter window length is used to capture more subtle time structures. When the heart rate is slower and the R-wave interval is longer, a longer window length is used to obtain richer low-frequency spectral information. This dynamic window length strategy overcomes the shortcomings of insufficient time resolution or frequency resolution that may be caused by fixed window length when analyzing variable heart rate signals. Each signal segment extracted by the sliding window needs to undergo two independent time-frequency analysis processes: windowed Fourier transform and wavelet packet decomposition. Windowed Fourier transform applies a window function to the signal segment and then performs a fast Fourier transform to obtain an overview of the signal's energy distribution in the classical frequency domain, providing the overall spectral characteristics of the signal. Wavelet packet decomposition is a more refined time-frequency analysis method that decomposes the signal into multiple levels. Each level of decomposition further divides the frequency band into two, generating a complete wavelet packet decomposition tree. This allows for the extraction of local time-varying features of the signal in a specific frequency band and provides excellent time positioning capabilities. The spectral amplitude sequence obtained from the windowed Fourier transform and the coefficient sequence of all terminal nodes obtained from wavelet packet decomposition are sorted and concatenated according to their corresponding center frequencies or frequency bands. The spectrum of the windowed Fourier transform covers the entire range from the DC component to the Nyquist frequency, while the coefficients of the wavelet packet decomposition provide detailed information on multiple sub-bands from low to high frequencies. The concatenation operation needs to ensure the correct alignment of the frequency axes, i.e., the continuity from low to high frequencies, ultimately forming a comprehensive time-frequency domain feature matrix. The rows of the time-frequency domain feature matrix typically correspond to time points or time windows, and the columns correspond to different frequency components or sub-bands. Each element value in the matrix represents the signal energy or amplitude at a specific time point and a specific frequency component. The construction of the time-frequency domain feature matrix provides the data foundation for subsequent spectrum generation.

[0080] The construction of the spectral energy distribution map begins with extracting the energy amplitude of each frequency band from the time-frequency domain feature matrix. For the windowed Fourier transform part, the energy amplitude can be the square of the modulus of the spectral coefficients; for the wavelet packet decomposition part, the energy amplitude can be the sum of squares or variance of the coefficients at each node. These energy values ​​reflect the intensity distribution of the ECG signal across different frequency components. The original discrete energy distribution map is plotted with the frequency band number or center frequency as the horizontal axis and the calculated energy amplitude as the vertical axis. To obtain a smooth curve for easier analysis, this discrete energy distribution map is Gaussian smoothed. Gaussian smoothing is achieved by convolving with a Gaussian kernel function, which effectively suppresses spikes caused by random noise while preserving the main peaks and contours of the energy distribution. The smoothed energy distribution curve needs to be normalized so that the total area under the curve equals one. Normalization eliminates the influence of the absolute signal amplitude, making the spectra of ECG signals from different individuals and under different recording conditions comparable. On the normalized spectral energy distribution curve, it is necessary to identify the dominant frequency bands. These are frequency bands with energy significantly higher than the background noise. The identification method involves setting a threshold, such as the global average energy plus three standard deviations, and marking the frequency bands corresponding to peak points with energy values ​​exceeding this threshold. These bands typically correspond to the basic rhythm, high-frequency components, or other physiologically significant oscillations in the electrocardiogram (ECG) signal. The energy proportion weight of each marked dominant frequency band is obtained by calculating the ratio of the sum of the energies of all frequencies within that band to the total energy of the entire spectrum. This energy proportion weight quantifies the relative importance of that frequency band in the ECG signal power spectrum.

[0081] The dynamic adjustment mechanism of the variable window length during multi-scale time-frequency transformation is closely related to the physiological characteristics of ECG signals. The R-wave interval directly reflects the instantaneous heart rate. Adjusting the window length based on the R-wave interval means matching the analysis window with the length of the cardiac cycle. Analyzing within a cardiac cycle window length can better capture electrophysiological events synchronized with the heartbeat and reduce spectral leakage. The combination of windowed Fourier transform and wavelet packet decomposition fully utilizes the advantages of both methods. Windowed Fourier transform is computationally efficient in estimating the overall spectrum of the signal, while wavelet packet decomposition provides excellent descriptive capabilities for the local features of non-stationary signals. In particular, for subtle phenomena such as high-frequency fragmented potentials or microvolt-level T-wave alternation that may exist in ECG signals, wavelet packet decomposition can provide better time-frequency localization. The construction of the spectral energy distribution map not only depends on the calculation of energy values ​​but also includes post-processing steps such as smoothing and normalization. The choice of kernel width for Gaussian smoothing requires a trade-off; a kernel that is too narrow may not effectively smooth noise, while a kernel that is too wide may lead to the loss of peak details. Normalization makes the analysis method insensitive to changes in signal gain, improving robustness. The labeling of the dominant frequency band is a key decision-making step. The threshold setting is based on global statistics, which makes the recognition process adaptive and able to cope with signals with different signal-to-noise ratios. The labeled dominant frequency band and its energy proportion weight constitute the core part of the subsequent classification features that describe the spectral characteristics.

[0082] Example 2: See Figure 3The calculation of the phase change trajectory map is based on the fundamental frequency phase angle sequence extracted from the time-frequency domain feature matrix. This sequence reflects the instantaneous phase information of the basic rhythm of the electrocardiogram (ECG) signal. Obtaining this sequence requires further processing of the time-frequency domain feature matrix, typically focusing on the frequency components corresponding to the basic heart rate of the ECG signal. The calculation process selects the R-wave peak points located by the QRS wave detection algorithm within a continuous cardiac cycle, and reads the fundamental frequency phase angle values ​​corresponding to these peak points, thus forming a phase angle sampling sequence synchronized with the heartbeat sequence. The fundamental frequency phase angle difference between adjacent R-wave peak points is defined as the instantaneous phase offset. This difference calculation considers the circular characteristics of the phase angle and requires modulo operation to ensure the result falls within a reasonable range (e.g., [-π, π] radians). Using the cardiac cycle number as the abscissa and the calculated instantaneous phase offset as the ordinate, points are plotted and connected on a two-dimensional plane to generate the phase change trajectory. This trajectory visually demonstrates the dynamic evolution of the phase relationship between heartbeats. Linear fitting is performed on the generated discrete phase change trajectory points. The goal of the fitting is to find a straight line that minimizes the sum of the squares of the perpendicular distances from all trajectory points to this line. This fitted line represents the overall trend of phase shift. The difference between each actual phase shift point and its corresponding point on the fitted line is calculated. These differences are called fitting residuals. The root mean square (RMS) value of the fitting residuals is used as a phase stability index to quantify the degree of phase jitter. A smaller RMS value means that the phase shift is closely distributed around the trend line, indicating that the heart rhythm is stable, while a larger value suggests the presence of significant phase noise or arrhythmia.

[0083] The generation of the initial arrhythmia classification feature vector is a systematic feature fusion and standardization process. The construction of the energy feature sub-vector begins with sorting the identified dominant frequency bands according to their center frequencies from low to high. The energy proportion weight values ​​corresponding to each dominant frequency band are extracted in this order, forming a one-dimensional numerical sequence. This sequence is the energy feature sub-vector, which characterizes the distribution of signal power across different frequency bands. The temporal feature sub-vector is composed of two different temporal features. The first feature is the previously calculated phase stability index, which reflects the temporal stability of the rhythm. The second feature is the average R-wave interval measured directly from the raw ECG signal. The average R-wave interval represents the average heart rate over a period of time. These two features describe the temporal characteristics of the ECG signal from different perspectives. Before concatenating the energy feature vector and the temporal feature vector into a complete initial arrhythmia classification feature vector, they need to be z-score standardized. z-score standardization, also known as standard deviation standardization, involves subtracting the mean of each element in the feature vector from the mean of all training samples and then dividing by the standard deviation. After this process, the element values ​​in the feature vector will follow a distribution with a mean of 0 and a standard deviation of 1. This effectively eliminates the imbalance caused by differences in physical units and numerical ranges among different features.

[0084] The standardized energy feature vector and temporal feature vector are concatenated end-to-end to form a higher-dimensional comprehensive feature vector. This vector simultaneously incorporates the frequency domain energy characteristics and temporal rhythm characteristics of the signal. To further enhance the discriminative power of the feature vector, an additional feature dimension is appended to the end of the concatenated vector: the QT interval variation coefficient of the electrocardiogram (ECG) signal. Calculating the QT interval variation coefficient requires first measuring the QT interval over multiple consecutive heartbeat cycles. The QT interval is the time interval from the start of the QRS complex to the end of the T wave. Then, the standard deviation of these QT interval values ​​is calculated, and this standard deviation is divided by the average QT interval to obtain the variation coefficient. The QT interval variation coefficient reflects the stability of ventricular repolarization time and is indicative of certain types of arrhythmias. The final generated initial arrhythmia classification feature vector is a digital representation integrating multiple information such as spectral energy proportion weights, phase stability, average heart rate, and repolarization time variability, providing a foundation for subsequent classification decisions.

[0085] The calculation of phase change trajectory maps relies on accurate R-wave detection and fundamental frequency phase angle estimation. The accuracy of R-wave detection directly affects the calculation of instantaneous phase shift, thus requiring a QRS wave detection algorithm with strong anti-interference capabilities. Obtaining the fundamental frequency phase angle typically involves performing a Hilbert transform on the signal to calculate the analytic signal, then extracting the phase angle components of the analytic signal and filtering out the frequency components corresponding to the baseline heart rate. Instantaneous phase shift reveals minute changes in the duration of the heartbeat cycle, which may be difficult to detect in conventional RR interval analysis. Linear fitting is used to eliminate potential linear trends in the phase shift, such as slow changes caused by slight heart rate accelerations or decelerations, making the fitting residuals more representative of non-trend-based, potentially pathological, phase jitter. The phase stability index, as a scalar, effectively summarizes the smoothness and predictability of the phase change trajectory. The construction of the initial arrhythmia classification feature vector embodies the idea of ​​multi-feature fusion: the energy feature vector comes from the frequency domain, the temporal feature vector comes from the time domain and time-frequency domain, and the QT interval variation coefficient is a morphological and temporal combined index. z-score normalization is a common data preprocessing step in machine learning. It brings different features to the same order of magnitude, which helps improve the performance of distance-based classifiers. The elements in the energy feature sub-vector are arranged in ascending frequency order, which aligns with signal processing conventions and may help capture patterns related to frequency gradients. The mean R-wave interval, as a fundamental time-domain feature, provides macroscopic information about heart rate, complementing microscopic phase stability indices. The introduction of the QT interval coefficient of variation adds consideration for the stability of the repolarization process, enabling the feature vector to describe more physiological processes beyond depolarization and rhythm.

[0086] Example 3: Specific Implementation Steps for Dynamic Compensation of Initial Arrhythmia Classification Feature Vectors Based on Spatial Correlation of Adjacent Signal Segments. The aim is to utilize the temporal continuity of ECG signals to smooth classification features and reduce the impact of random fluctuations on classification results. Classification feature vectors corresponding to the two preceding and following cardiac cycles of the current signal segment to be analyzed are selected. These vectors are temporally adjacent to the initial arrhythmia classification feature vector of the current signal segment, forming a reference vector set. The reference vector set contains five classification feature vectors: the vectors of the two preceding cardiac cycles, the vector of the current cardiac cycle, and the vectors of the two following cardiac cycles. The cosine similarity between the initial arrhythmia classification feature vector of the current signal segment and each vector in the reference vector set is calculated. Cosine similarity measures the degree of similarity between two vectors in direction, with a value range of [-1, 1]. The closer the value is to 1, the more consistent the directions.

[0087] The calculation of cosine similarity follows the standard formula in the vector space model:

[0088]

[0089] in: Cosine similarity is a dimensionless scalar. The initial arrhythmia classification feature vector represents the current signal segment. This represents a specific reference classification feature vector within the reference vector set. (Symbol) This represents the dot product operation of vectors. (Symbol) This represents the Euclidean norm (or magnitude) of a vector. The numerator of the formula is the dot product of the two vectors, reflecting their similarity; the denominator is the product of the magnitudes of the two vectors, used to normalize the dot product. This formula ensures that the calculation result is independent of the absolute value of the vectors, depending only on their direction, satisfying the mathematical definition of cosine similarity, and both sides of the formula are scalars with consistent dimensions.

[0090] Based on the calculated cosine similarity, a compensation weight coefficient is assigned to each vector in the reference vector set. The principle for assigning compensation weight coefficients is that reference vectors with higher cosine similarity receive greater weights, indicating that they are closer to the feature patterns of the current signal segment and should play a greater role in the compensation process. A specific weight allocation method involves transforming the cosine similarity values ​​using a softmax function. The softmax function converts a set of real numbers into a probability distribution, where the ratio of the exponent of each number to the sum of the exponents of all numbers is used as its output, ensuring that the sum of all weight coefficients is 1. All five classification feature vectors in the reference vector set are then weighted and averaged according to their corresponding compensation weight coefficients. This weighted average produces a single, representative compensation reference vector that incorporates the feature information of the current signal segment's context. The final step involves performing a convex combination operation between the initial arrhythmia classification feature vector of the current signal segment and the weighted averaged compensation reference vector. A convex combination operation is a linear combination where all combination coefficients are non-negative and sum to 1.

[0091] Convex combination operations can be represented as:

[0092]

[0093] in: This represents the optimized classification feature vector. The initial arrhythmia classification feature vector represents the current signal segment. This represents the compensation reference vector obtained through weighted averaging. It is a mixing factor between 0 and 1, used to control the mixing ratio between the current vector and the reference vector. Mixing factor The value of can be dynamically adjusted based on the signal-to-noise ratio estimate of the current signal segment's feature vector or its average similarity to the reference vector. The optimized classification feature vector, resulting from the convex combination operation, retains the unique individual information of the current signal segment while incorporating the spatial context information provided by its preceding and following signal segments, making the generated feature vector more robust to transient noise and interference.

[0094] The process of determining the candidate arrhythmia type set is a stepwise optimization process based on distance metrics and statistical screening. The Mahalanobis distance between the optimized classification feature vector and the template vector of each arrhythmia category in the pre-defined arrhythmia type feature library is calculated. Mahalanobis distance, unlike Euclidean distance, considers the covariance relationship between feature dimensions, making it a more effective classification distance metric that can eliminate the influence of feature correlation. The calculation of Mahalanobis distance requires the inverse matrix of the covariance matrix of the corresponding category in the feature library. A pre-defined Mahalanobis distance threshold is set, and all template types with Mahalanobis distances less than this threshold are filtered out of the feature library, forming a preliminary candidate set. This threshold determines the breadth of the candidate set. Within the preliminary candidate set, potential outliers need to be further eliminated by calculating the average Mahalanobis distance between each type in the preliminary candidate set and all other types in the set. The median of these average Mahalanobis distances is calculated, and types with an average Mahalanobis distance exceeding twice the median are identified as outliers. Outliers mean that they differ significantly from most other types in the candidate set and may be a mismatch. The identified outlier types are removed from the initial candidate set to ensure that the remaining types are relatively concentrated in the feature space. Finally, the remaining arrhythmia types after outlier removal are sorted in ascending order according to their respective Mahalanobis distances, with the type with the smallest Mahalanobis distance listed first. This ordered list is the final set of candidate arrhythmia types.

[0095] The core idea of ​​the dynamic compensation mechanism is to utilize the autocorrelation of ECG signals over a short period of time, as the electrophysiological activities of adjacent cardiac cycles typically exhibit high similarity. Cosine similarity, as a measure of directional similarity, is insensitive to the absolute magnitude of feature vectors, which helps focus on comparing feature patterns rather than differences in energy magnitude. The allocation strategy of compensation weight coefficients gives higher weight to neighbors more similar to the current segment. The weighted average operation produces a "consensus" contextual feature representation. The mixing factor α in the convex combination operation provides flexibility, allowing adjustment of the level of trust in current information based on signal quality. The use of Mahalanobis distance improves the classifier's discriminative ability in feature-related situations. The initial candidate set selection, based on an absolute distance threshold, is a coarse-grained process. The outlier removal step, based on the relative distance distribution within the candidate set, is a fine-grained process, removing candidates inconsistent with the group and improving the cohesion of the candidate set.

[0096] Example 4: The implementation process of hierarchical screening based on the feature similarity of each type in the candidate arrhythmia type set constitutes the final stage of classification decision-making. The candidate arrhythmia type set is an ordered list containing several arrhythmia types with the smallest Mahalanobis distance. The top three types are extracted from the candidate arrhythmia type set as the main candidate types for in-depth analysis. These three main candidate types represent several possibilities that best match the characteristics of the ECG signal to be classified. The similarity of the signal to be classified with each main candidate type in spectral energy distribution is calculated. Spectral energy distribution similarity focuses on whether the distribution pattern of signal power in different frequency subbands is consistent. The calculation can use the histogram intersection method or the correlation coefficient method to compare the degree of agreement between the normalized spectral energy distribution curve of the signal to be classified and various typical spectral templates stored in the feature library. Simultaneously, the similarity of the signal to be classified with each major candidate type in terms of phase trajectory morphology is calculated. Phase trajectory morphology similarity assesses whether the shape and fluctuation pattern of the phase change trajectory are similar. Dynamic time warping algorithms or shape context-based descriptors can be used to quantify this morphological similarity. This index reflects the temporal dynamic characteristics of arrhythmia patterns.

[0097] Independent thresholds are set for spectral energy distribution similarity and phase trajectory morphology similarity, and these thresholds are determined based on statistical analysis of known training data. When both the calculated spectral energy distribution similarity and phase trajectory morphology similarity values ​​exceed their respective preset thresholds for a given primary candidate type, it indicates that the candidate type highly matches the signal to be classified in both frequency domain and temporal characteristics. At this point, the system automatically triggers a more refined and rigorous secondary feature verification process. The secondary feature verification process aims to utilize morphological features in ECG signals other than the spectrum and phase, namely the morphological features of the ST segment and T wave, to finally confirm or correct the primary classification results, thereby resolving certain arrhythmia types that may be confused in the primary feature space.

[0098] The secondary feature verification process includes ST-T segment separation of the current signal segment. This separation requires precise identification of the QRS complex endpoint and the T wave endpoint, thus isolating the ST segment and T wave from the complete cardiac cycle and obtaining independent ST segment and T wave signals for subsequent analysis. The average slope of the ST segment is calculated over a fixed time range of 80 milliseconds after point J, the connection point between the QRS complex and the ST segment. The average slope is calculated by linearly fitting the signal segment, taking the slope value of the fitted line, and calculating its absolute value as the ST segment slope index. This index reflects the upward or downward trend of the ST segment. The symmetry index is extracted from both sides of the T wave peak. Calculating the symmetry index requires locating the T wave peak point and then calculating the area or amplitude characteristics of the T wave signal within a certain time window to the left and right of the peak point. The symmetry of the T wave morphology is quantified by comparing the ratio or difference of the features on both sides. The calculated absolute value of the average slope of the ST segment is combined with the T wave symmetry index to form a verification feature pair containing two elements. This verification feature pair carries important morphological information about the ventricular repolarization period.

[0099] The verification feature pairs generated from the signal to be classified are compared with the typical value ranges of each major candidate type pre-stored in the feature library. The feature library stores typical value ranges of ST segment slope and T wave symmetry index for each arrhythmia type, obtained based on statistics from a large number of samples. The comparison process evaluates whether each element in the verification feature pair of the signal to be classified falls within the corresponding typical value range of the major candidate type. Based on the matching degree results between the verification feature pairs and the major candidate types, a final decision is made on the classification results based on primary feature similarity. A high matching degree confirms the type as the final result; a low matching degree may select a suboptimal candidate or output an "undetermined" result. Refer to Table 1, which shows the typical value ranges of several major candidate arrhythmia types that may be stored in the feature library, used for comparison operations in the secondary feature verification process.

[0100] Table 1: Typical Value Ranges of ST-T Segment Characteristics for Major Cardiac Arrhythmias

[0101]

[0102] The hierarchical screening logic reflects a coarse-to-fine classification strategy. Primary screening rapidly narrows the candidate range based on global spectral and phase features, while secondary verification utilizes local morphological features for precise discrimination. Spectral energy distribution similarity distinguishes different types from the perspective of frequency components; for example, some arrhythmias may be accompanied by high-frequency or low-frequency energy variations. Phase trajectory morphological similarity provides information from the perspective of rhythm stability; atrial fibrillation typically exhibits a highly irregular phase trajectory. A dual-threshold triggering mechanism ensures that only cases highly suspected in both aspects proceed to the more resource-intensive secondary verification, improving system efficiency. The accuracy of ST-T segment separation processing is crucial and relies on a reliable waveform boundary detection algorithm. The ST segment slope is calculated using a window of 80 milliseconds after the J point, a clinically commonly used analysis interval capable of capturing meaningful ST segment changes. The T-wave symmetry index reflects the spatial coordination of the ventricular repolarization process; certain cardiac diseases can cause abnormal T-wave morphology. The use of verification feature pairs combines two key repolarization phase indicators—the ST segment and the T wave—providing a more comprehensive morphological view. The typical value range stored in the feature library needs to be obtained through statistical learning based on a large-scale, well-labeled clinical database. The range should be set to cover most of the variations in this type while still distinguishing it from other types. The final decision rule can be based on logical judgment, such as requiring both elements of the verification feature pair to fall within the range of a certain candidate type before confirmation, or it can be based on a weighted scoring mechanism. The entire hierarchical screening process integrates multi-dimensional information such as the frequency domain, time domain, time-frequency domain, and morphology of the signal, aiming to improve the accuracy and reliability of classifying complex arrhythmias.

[0103] Example 5: After outputting the final arrhythmia classification result, the system executes an important self-learning and optimization mechanism. This involves adding the optimized classification feature vector generated after the current signal segment completes the processing flow, along with its final classification label, as a new knowledge sample to the preset arrhythmia type feature library. The arrhythmia type feature library is essentially a database storing typical feature templates of known arrhythmia types. Each type contains multiple feature vector samples learned from historical data. Adding new samples allows the arrhythmia type feature library to continuously absorb new data, reflecting potential individual differences or new pattern variations, thereby maintaining its representativeness and timeliness. The addition operation must follow a certain data management strategy, setting a preset upper limit for the number of samples for each type of arrhythmia in the arrhythmia type feature library. This upper limit controls the size of each type of sample, preventing the arrhythmia type feature library from expanding indefinitely and affecting query speed, and ensuring the timeliness of each type of data.

[0104] When the number of samples for a certain type of arrhythmia in the arrhythmia type feature library reaches or exceeds its preset upper limit, the system automatically triggers a template vector update process for that type of arrhythmia. The goal of this template vector update process is to calculate a new, more representative feature center for that type of arrhythmia, replacing the old template vectors in the arrhythmia type feature library. After the update process is initiated, the system extracts the feature vectors of all samples belonging to that specific arrhythmia category from the arrhythmia type feature library; these vectors form the data basis for this update. Clustering algorithms are applied to this set of feature vectors to find new centers of data distribution. K-means clustering is a commonly used choice, aiming to divide all samples into K clusters such that the sum of the squared distances from each sample to the center of its cluster is minimized. In the template vector update scenario, the number of clusters K is usually set to 1, with the aim of finding a center point that can represent the entire sample set.

[0105] The K-means clustering algorithm (K=1) is applied to cluster all extracted feature vectors of the class. The calculation process iteratively optimizes the search for a point that minimizes the sum of distances to all other points in the set; this point is the cluster center. The calculated cluster center is a vector with the same dimension as the feature vectors, representing the most typical feature pattern of this type of arrhythmia under the current data distribution. This newly calculated cluster center vector will formally replace the original template vector stored in the arrhythmia type feature library. The original template vector may have been calculated from early training data during system initialization or generated from the previous update process. Replacing the original template vector with the new cluster center completes the feature library update. The updated template vector more accurately reflects the overall feature distribution of this type of arrhythmia after the addition of new samples. After the template vector is updated, all historical sample feature vectors under this class can be retained for subsequent analysis, or archived or cleaned up according to a strategy to make room for new samples.

[0106] To illustrate this process with a concrete example, suppose the system's preset arrhythmia type feature library has a maximum sample size of 1000 for the "ventricular premature beats" category. After a long period of operation, the number of samples in the "ventricular premature beats" category in the arrhythmia type feature library has gradually accumulated from the initial 800 to 1001. Reaching 1001 samples exceeds the preset maximum of 1000, a condition that triggers the template vector update process. The system automatically initiates the template vector update process for the "ventricular premature beats" category. The update process first retrieves and extracts all 1001 optimized classification feature vectors labeled "ventricular premature beats" from the arrhythmia type feature library. Each vector represents the characteristics of a historically successfully classified ventricular premature beat. K-means clustering is then applied to these 1001 feature vectors for cluster analysis. Since the goal is to find a central point, the number of clusters, K=1, is set. The K-means algorithm, after multiple iterations, converges to a final cluster center. This center vector is the mean of 1001 sample points in the feature space, best representing the central trend of this data set. The calculated new cluster center vector is used to replace the previously stored "ventricular premature beats" template vector in the arrhythmia type feature library. The old template vector might have been calculated based on the initial 800 samples, while the new template vector is calculated based on 1001 more abundant samples, theoretically making it more representative. After the template vector replacement operation is completed, the feature library update process for the "ventricular premature beats" category ends. The sample counter for the "ventricular premature beats" category in the arrhythmia type feature library can be reset to zero to begin a new round of sample accumulation, or some of the latest samples can be retained; the specific strategy depends on the system design. Subsequently, when the system classifies new ECG signals, the "ventricular premature beats" template vector used to calculate the Mahalanobis distance is the new vector generated after this update.

[0107] This dynamic update mechanism ensures that the arrhythmia type feature library is not a static, unchanging knowledge base, but a dynamic system that continuously evolves and optimizes with the influx of new data. Adding new samples to the arrhythmia type feature library is the foundation of knowledge accumulation, and the upper limit on the number of samples prevents unlimited growth of storage space and unnecessary increases in model complexity. The mechanism for triggering the template vector update process ensures that updates occur when the amount of data reaches a certain scale and is statistically significant. Using clustering algorithms to recalculate centroids is a data-driven approach; the new centroids reflect the overall distribution of all historical samples (including newly added ones), avoiding excessive influence of individual new samples on the template vector. Replacing the original template vector with new cluster centers allows the system to adapt to the slow drift that may exist in ECG features, such as changes caused by long-term changes in the patient's physiological state or minor changes in electrode position. The update process of "ventricular premature beats" in the example clearly demonstrates the complete closed loop from condition triggering, data extraction, centroid calculation to template replacement. Resetting the sample counter or formulating a sample rotation strategy ensures that the system always updates knowledge based on relatively new data, avoiding the impact of outdated data on model performance.

[0108] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0109] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for classifying various cardiac arrhythmia signals based on spectral analysis, characterized in that, Includes the following steps: Collect and preprocess the electrocardiogram (ECG) signal data of the target object to generate a standardized ECG signal sequence; Multi-scale time-frequency transformation is performed on standardized electrocardiogram signal sequences to extract time-frequency domain feature matrices; Construct a spectral energy distribution map and a phase change trajectory map based on the time-frequency domain feature matrix; The dominant frequency bands are divided based on the spectral energy distribution map, and the energy proportion weight of each frequency band is marked. The phase shift between adjacent cardiac cycles is calculated based on the phase change trajectory map to generate a phase stability index. By integrating energy proportion weights and phase stability indices, an initial arrhythmia classification feature vector is generated. The initial arrhythmia classification feature vector is dynamically compensated by the spatial correlation of adjacent signal segments to generate an optimized classification feature vector. Based on the matching degree between the optimized classification feature vector and the preset arrhythmia type feature library, a set of candidate arrhythmia types is determined; Based on the feature similarity of each type in the candidate arrhythmia type set, hierarchical screening is performed, and the final arrhythmia classification result is output. The energy percentage weight of each marked dominant frequency band is obtained by calculating the ratio of the sum of the energy of all frequency points in that band to the total energy of the entire spectrum. The energy percentage weight quantifies the relative importance of that frequency band in the power spectrum of the electrocardiogram signal. Linear fitting is performed on the phase change trajectory, and the difference between each actual phase offset point and the corresponding point on the fitted line is calculated. The difference is called the fitting residual, and the root mean square of the fitting residual is used as the phase stability index. The process of generating the initial arrhythmia classification feature vector includes: arranging the energy proportion weights of each dominant frequency band from low to high according to the frequency band to form an energy feature sub-vector; The phase stability index is combined with the average R-wave interval to form a phase feature vector; the energy feature vector and the phase feature vector are z-score standardized and then concatenated; the QT interval variation coefficient is added as an additional dimension to the concatenated vector.

2. The method for classifying various arrhythmia signals based on spectral analysis according to claim 1, characterized in that, The specific implementation process of performing multi-scale time-frequency transformation on the standardized electrocardiogram signal sequence includes: A sliding window with variable window length is used to segment and extract standardized electrocardiogram signal sequences. Perform windowed Fourier transform and wavelet packet decomposition on each signal segment; The Fourier transform results and wavelet packet decomposition coefficients are concatenated in frequency band order to form a time-frequency domain feature matrix. The length of the sliding window is dynamically adjusted according to the R-wave interval within the signal segment.

3. The method for classifying various arrhythmia signals based on spectral analysis according to claim 1, characterized in that, The process of constructing the spectral energy distribution map includes: Extract the energy amplitude of each frequency band from the time-frequency domain feature matrix; An energy distribution curve is generated with frequency band as the horizontal axis and energy amplitude as the vertical axis. The energy distribution curve is Gaussian smoothed and normalized. Peak points in the energy distribution curve that exceed three standard deviations above the global mean are designated as the dominant frequency bands.

4. The method for classifying various arrhythmia signals based on spectral analysis according to claim 1, characterized in that, The calculation process of the phase change trajectory map includes: Extract the fundamental frequency phase angle sequence from the time-frequency domain feature matrix; The phase angle difference between adjacent R-wave peak points is calculated as the instantaneous phase offset; A phase change trajectory is generated with the cardiac cycle number as the horizontal axis and the instantaneous phase shift as the vertical axis; The phase change trajectory is linearly fitted, and the root mean square of the fitting residual is used as the phase stability index.

5. The method for classifying various arrhythmia signals based on spectral analysis according to claim 1, characterized in that, The specific implementation steps for dynamically compensating the initial arrhythmia classification feature vector based on the spatial correlation of adjacent signal segments include: The classification feature vectors of two cardiac cycles before and after the current signal segment are selected as the reference vector set; Calculate the cosine similarity between the feature vector of the current signal segment and each reference vector; Compensation weight coefficients are assigned based on the similarity level; Perform a convex combination operation between the weighted average of the reference vector set and the current feature vector.

6. The method for classifying various arrhythmia signals based on spectral analysis according to claim 1, characterized in that, The process of determining the set of candidate arrhythmia types includes: Calculate the Mahalanobis distance between the optimized classification feature vector and the template vector of each class in the feature library; Template types with a Mahalanobis distance less than a preset threshold are selected as the primary candidate set; Outliers that are more than twice the median distance from other types in the primary candidate set are removed. The remaining types are sorted in ascending order of distance to generate a set of candidate arrhythmia types.

7. The method for classifying various arrhythmia signals based on spectral analysis according to claim 6, characterized in that, The implementation process of hierarchical screening based on the feature similarity of each type in the candidate arrhythmia type set includes: Extract the top three main candidate types from the set of candidate arrhythmia types; Calculate the spectral energy distribution similarity and phase trajectory morphology similarity of the main candidate types respectively; When both similarity metrics exceed their respective thresholds, the secondary feature verification process is triggered. In the secondary feature verification, the slope of the ST segment and the symmetry index of the T wave are compared; The final arrhythmia classification result was revised based on the verification results.

8. The method for classifying various arrhythmia signals based on spectral analysis according to claim 7, characterized in that, The secondary feature verification process includes: Perform ST-T segment separation processing on the current signal segment to obtain independent ST segment and T wave signals; Calculate the absolute value of the average slope of segment ST within 80ms after point J; Extract the symmetry index on both sides of the T-wave apex; The absolute value of the slope is combined with the symmetry index to form a verification feature pair; Compare and verify the matching degree of feature pairs with the typical value range of each major candidate type.

9. The method for classifying various arrhythmia signals based on spectral analysis according to claim 1, characterized in that, The method further includes: After outputting the final classification result, the optimized classification feature vector of the current signal segment is added to the feature library; When the number of similar samples in the feature library exceeds a preset limit, the template vector update process is triggered. The feature vectors of similar samples are recalculated using a clustering algorithm; The feature library is updated by replacing the original template vector with the new cluster centers.

Images