A multi-lead electrocardio feature recognition method and system for arrhythmia attack

Abstract

This invention discloses a multi-lead electrocardiogram feature identification method and system for arrhythmia attacks. The method assesses signal reliability based on the respiratory modulation pattern of each lead and generates weighted parameters. Ventricular waveform boundary localization is performed on the weighted fused signal. Based on the localization results, the inter-beat variation of translead activation time difference is tracked, and conduction abnormalities are marked in matrix form. Within the sinus rhythm reference range defined by the conduction abnormality markings, alternating odd and even beat amplitude features are extracted from the multi-lead residual signals to generate a prodromal score. The warning triggering conditions are determined by combining the repolarization dispersion test. During the triggering period, the same-direction attenuation diffusion direction of the amplitude in each lead is analyzed segment by segment. Myocardial spatial heterogeneity is assessed through differential contraction at the attenuation segment boundaries, and the attack type is classified. The residuals of the termination frame waveform and historical templates are compared segment by segment. After verifying the consistency between the cardioversion type and the attack type, an attack identification report is output, providing a verifiable basis for arrhythmia type attribution in clinical practice.

Description

Technical Field

[0001] This invention relates to the field of medical signal processing technology, and in particular to a method and system for identifying multi-lead electrocardiogram features during arrhythmia episodes. Background Technology

[0002] Ventricular tachycardia, ventricular fibrillation, and other rapid arrhythmias are major causes of sudden cardiac death. Multi-lead surface electrocardiography is a core tool for long-term monitoring; however, patient activity and changes in skin-electrode contact cause the reliability of signals in each lead to fluctuate continuously during monitoring. Existing recognition systems generally process each lead equally, with leads of lower quality being directly included in the calculation, which can easily lead to false alarms or missed detections in highly interference-prone environments.

[0003] Before the onset of ventricular arrhythmia, the electrical heterogeneity during ventricular repolarization will produce quantifiable, subtle amplitude changes between adjacent heartbeats. Accurately extracting this precursor feature in the context of artifacts remains a weak link in existing protocols. The multi-lead waveform at the time of arrhythmia termination contains electrophysiological information of the cardioversion mechanism. Comparing it with various types of historical feature systems and verifying the consistency of the attribution of onset and termination is of reference value for subsequent clinical decision-making. However, there is no complete protocol in the current system for this aspect. Summary of the Invention

[0004] This invention provides a method and system for identifying multi-lead electrocardiogram features of arrhythmia attacks. By adaptively evaluating and weightedly fusing the reliability of multi-lead signals, a conduction abnormality labeling system based on translead excitation timing differences is established. Combined with beat-by-beat extraction based on the alternating pattern of odd and even beat amplitudes and sliding assessment of the prodromal score, as well as spatial tracking of the same-direction amplitude decay and class-by-class residual comparison of the historical template of the termination frame, a systematic identification of the entire process of arrhythmia attacks is achieved, and an attack identification report containing the consistency verification conclusion of termination type and attack type is output.

[0005] The first aspect of this invention proposes a method for identifying multi-lead electrocardiogram features during arrhythmia episodes, comprising the following steps: Collect multi-lead electrocardiogram (ECG) signals, and perform respiratory modulation depth recognition of each lead based on the multi-lead ECG signals to generate lead confidence weights; The ventricular wave group boundary is located step-by-step using the multi-lead ECG signal and the lead confidence weights to generate a wave group time sequence. The wave group time sequence is used to perform cross-lead time difference step-by-step irregular tracking to form a time difference distribution matrix. The time difference distribution matrix is ​​used to evaluate the time difference abnormal segments and conduction velocity to obtain conduction abnormality markers. The sinus rhythm residual is traced beat by beat to form a micropotential residual sequence by the multi-lead electrocardiogram signal and the conduction abnormality marker. Based on the micropotential residual sequence, the alternation amplitude screening of odd and even beats is carried out to obtain the pre-attack score. Based on the pre-attack score, the multi-lead repolarization dispersion test is performed beat by beat to determine the warning triggering condition. Based on the aforementioned warning triggering conditions, multi-lead amplitude isotropic beat-by-beat measurement is performed to obtain the amplitude coordination coefficient. The amplitude coordination coefficient is used to measure the contraction rate of the attenuation zone boundary to form a spatial heterogeneous distribution. The spatial heterogeneous distribution is then used to locate the heterogeneous region and establish an epidural type record. The seizure type and termination boundary are located from the seizure type record to determine the seizure conclusion parameters. The seizure conclusion parameters are used to perform termination frame wave group morphology tracing to generate termination cardioversion markers. Based on the termination cardioversion markers, the cardioversion process is tracked and the consistency with the type prediction is output to generate a seizure identification report.

[0006] A second aspect of this invention provides a multi-lead electrocardiogram feature recognition system for arrhythmia episodes, comprising: The signal acquisition module is used to acquire multi-lead electrocardiogram (ECG) signals and generate lead confidence weights based on the multi-lead ECG signals to perform respiratory modulation depth recognition of each lead. The conduction analysis module is used to perform beat-by-beat localization of ventricular wave group boundaries on the multi-lead ECG signals and the lead confidence weights to generate wave group time series sequences. The wave group time series sequences are used to perform beat-by-beat irregular tracking of cross-lead time difference to form a time difference distribution matrix. The time difference distribution matrix is ​​used to evaluate the time difference abnormal segments and conduction velocity to obtain conduction abnormality markers. The early warning module is used to perform sinus rhythm residual tracking on a beat-by-beat basis on the multi-lead electrocardiogram signal and the conduction abnormality marker to form a micropotential residual sequence, perform odd-even beat alternation amplitude screening based on the micropotential residual sequence to obtain an episodic prodrome score, and perform multi-lead repolarization dispersion test on a beat-by-beat basis according to the episodic prodrome score to determine the early warning triggering conditions. The spatial positioning module is used to measure the amplitude coherence of multi-lead amplitudes step by step according to the warning triggering conditions to obtain the amplitude coordination coefficient, use the amplitude coordination coefficient to measure the contraction rate of the attenuation segment boundary to form a spatial heterogeneous distribution, and perform heterogeneous region positioning of the spatial heterogeneous distribution to establish an attack type record. The conclusion output module is used to locate the seizure type and termination boundary from the seizure type record, determine the seizure conclusion parameters, perform termination frame wave group morphology tracing on the seizure conclusion parameters to generate a termination cardioversion mark, and output a seizure identification report based on the consistency between the cardioversion process and the type prediction based on the termination cardioversion mark.

[0007] The beneficial effects of this invention are reflected in the following points: First, the method of measuring lead reliability in this invention differs from conventional schemes that use signal-to-noise ratio or baseline drift amplitude as indicators. The following pattern of ECG signals in each lead to respiratory modulation is constrained by the electrode-skin contact state. When the electrode is loose or the impedance is unstable, this following pattern is disordered. Therefore, the modulation deviation rate becomes an endogenous indicator reflecting the electrode state, without the need for external sensor intervention. The lead reliability weights generated accordingly continuously adapt to changes in electrode state. The irregularity of translead activation time difference is quantified by the degree of differential directional unidirectional accumulation rather than the absolute value of the time difference, thus distinguishing between fixed bundle branch block and the progressive conduction delay caused by myocardial ischemia in terms of quantitative characteristics. Second, the extraction of precursor electrophysiological changes in this invention is based on sinus rhythm residual signals rather than the original ECG signals. After constructing a reference template with normal conduction beats, the signals are subtracted beat by beat. The micropotential component becomes the dominant signal in the residual domain, and the beat-by-beat alternation pattern of the odd-even beat residual amplitude difference thus has quantitative significance, avoiding the masking of weak precursor features by high-amplitude QRS components. The prodromal score is used here as a time-domain pre-screening condition, limiting the warning trigger to the period when the prodromal signal and repolarization index are simultaneously established, reducing false alarms from a single indicator. Finally, the analysis of amplitude decay during the attack is extended from "whether it decreases" to "in which direction it spreads": the diffusion direction of the same downward event in each lead is vectorized using frontal plane coordinates, and spatial heterogeneity is assessed by the difference in contraction rates between the anterior and posterior ends of the decay segment, allowing focal and diffuse attacks to be distinguished based on the geometric contraction morphology of the decay boundary. The morphological comparison of the termination frame uses the median template of historical deviation for each type as the benchmark rather than a single normal sinus rhythm template, and the consistency verification between the type classification and the attack stage label provides a self-consistent basis for the report output. Attached Figure Description

[0008] The accompanying drawings illustrate specific examples of the technical solutions described in this invention and, together with the detailed embodiments, form part of the specification, serving to explain the technical solutions, principles, and effects of this invention.

[0009] Figure 1 This is a flowchart illustrating a method for identifying multi-lead electrocardiogram features during arrhythmia according to the present invention.

[0010] Figure 2 This is a schematic diagram of the construction of the time difference distribution matrix and the spatial aggregation of transmission anomalies in this invention.

[0011] Figure 3 This is a schematic diagram of the alternating amplitude extraction of odd and even beat micropotential residuals in this invention.

[0012] Figure 4 This is a structural block diagram of a multi-lead electrocardiogram feature recognition system for arrhythmia episodes according to the present invention.

[0013] Wherein: 1-lead number identifier; 2-diagonal zero-value element; 3-high irregularity element; 4-low irregularity element; 5-focal aggregation area; 6-global diffuse morphology; 7-missing occupant element; 8-conduction abnormality threshold marker; 9-ECG waveform beat sequence label; 10-sinus rhythm reference template; 11-odd beat micropotential residual waveform; 12-even beat micropotential residual waveform; 13-aligned effective overlapping segment; 14-inter-beat amplitude difference; 15-dynamic noise benchmark; 16-alternating segment marker. Detailed Implementation

[0014] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0015] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.

[0016] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0017] The technical solutions of the embodiments of this application will be described below.

[0018] like Figure 1 As shown, this embodiment of the invention provides a method for identifying multi-lead electrocardiogram features during arrhythmia, including the following steps S101-S105: Step S101: Collect multi-lead ECG signals, and generate lead confidence weights based on the respiratory modulation depth recognition of each lead based on the multi-lead ECG signals.

[0019] Specifically, multi-lead ECG signals are acquired. Following a standard 12-lead electrode layout, skin abrasion, application of conductive gel, and electrode adhesion are performed. The differential analog signal at the skin-electrode interface is picked up by the acquisition end of each lead electrode. This signal is then amplified by an instrumentation amplifier with a gain of 1000, and quantized using analog-to-digital conversion at a sampling rate of 500Hz and a resolution of 16 bits. For example, a sampling rate of 500Hz and a resolution of 16 bits are used; the sampling rate can also be 250Hz or 1000Hz, with the duration of each time window and the number of sampling points involved in inter-frame statistics proportionally adjusted for the sampling rate. The quantization sequences of each lead are packaged frame by frame according to the lead number and aligned to the same time axis to form a multi-lead ECG signal with a two-dimensional frame structure. The row dimension corresponds to the lead number, and the column dimension corresponds to the sampling points within the frame. The sampling value is in μV, and the number of channels is consistent with the actual number of solidified leads. The quantization sequence is saved as a 16-bit signed integer along with two acquisition parameters: 1000x gain and 500Hz sampling rate. Each frame header includes a sampling timestamp and one bit each of saturation and motion artifact flags. The flags are set one by one in subsequent processing: first, a high-pass filter with a cutoff of 0.5Hz is used to filter out low-frequency baseline drift frame by frame, and then second-order notch filters at 50Hz and 60Hz are used to eliminate power frequency and harmonic interference in sequence. The results of the two-stage filtering are written back to the corresponding channels of the multi-lead ECG signal. The normal ECG amplitude falls within ±5mV. If the absolute value of the amplitude of a single frame still exceeds 5mV, a saturation flag is set at the beginning of the channel of that frame. During an attack, the patient's limbs frequently tremble, and motion artifacts are superimposed as broadband noise on the multi-lead ECG signal. The mean square error (MSE) between each channel is calculated frame by frame. In some embodiments, the median or moving average of the MSE of the previous few frames in a channel is used as the baseline MSE of the current channel. Frame segments with an MSE exceeding 8 times the baseline for 5 consecutive frames are marked with motion artifacts, and these marked frames are not included in subsequent modulation depth identification. If the amplitude of a lead electrode is below the detection limit of 5 μV for 10 consecutive frames, the lead is considered detached, and the entire multi-lead ECG signal of that channel is marked as invalid, and the quantized sequence is not included in the valid data range.

[0020] In some embodiments, the step of generating lead confidence weights by identifying respiratory modulation depth in each lead based on the multi-lead ECG signal includes: measuring the amplitude envelope of each lead on a cycle-by-cycle basis to obtain an amplitude envelope sequence; performing respiratory frequency band and residual fluctuation layer-by-layer identification on the amplitude envelope sequence to determine the respiratory modulation cycle; performing modulation deviation identification and deviation rate assessment on a beat-by-beat basis according to the respiratory modulation cycle to generate a modulation depth index; and performing lead confidence layer-by-layer sorting and weight allocation based on the modulation depth index to obtain lead confidence weights.

[0021] Amplitude envelope sequences were obtained by measuring the amplitude envelope of each lead in a multi-lead ECG signal on a cycle-by-cycle basis. A sliding detection window was set for each channel of the multi-lead ECG signal, with the nominal duration of the R wave as the window length. This window length corresponds to approximately 80ms to 120ms of sampling points. The sampling point with the largest absolute value within the window was taken as the R wave peak point. The peak amplitudes were arranged sequentially according to the beat number to form a peak point sequence. During an attack, when the heart rate suddenly rises or the interval between two adjacent beats is extremely compressed, the sliding window may misjudge the P wave peak as a peak point. Therefore, a false peak identification was performed first. Peaks with an interval between two adjacent peak points less than 55% of the average of the previous 15 beats in that lead were identified as false peaks and removed from the peak point sequence. The peak points were then replaced by linear interpolation of the two true peak points before and after the false peak. After cleaning, the peak points of the entire segment of the same lead are fitted with a cubic spline curve. Adjacent peak points are used with the beat interval as spline node segments, ensuring the continuity of the first and second derivatives at the nodes. The spline is uniformly resampled according to the original sampling rate of the multi-lead ECG signal, and a continuous time series with the same duration as the multi-lead ECG signal is output as the amplitude envelope sequence of that lead. After each lead is completed in sequence, the amplitude envelope sequences of each lead are merged into a floating-point matrix and stored uniformly using the lead number as the row index. The number of rows in the matrix is ​​equal to the number of effective leads, and the number of columns is equal to the total number of time-domain samples of the multi-lead ECG signal. Each row also includes two counts: the number of effective peak points participating in the spline fitting and the number of spurious peaks removed. These are retained records of the envelope fitting quality of that lead for verification. The slowly varying components of the amplitude envelope sequence quantify the R-wave projection modulation of the corresponding leads at each respiratory phase: during deep inspiration, the diaphragm moves downward and the heart axis rotates in the frontal plane, increasing the angle between the precordial leads and the ECG vector, resulting in a local maximum in the amplitude envelope sequence at the peak of inspiration; at the end of expiration, the heart position returns to normal, and the amplitude envelope sequence falls accordingly, with the fluctuation period and amplitude driven by the respiratory rhythm.

[0022] The respiratory modulation period is determined by hierarchical identification of respiratory frequency bands and residual fluctuations in the amplitude envelope sequence. A sliding window is used to segment the amplitude envelope sequence of each lead. In some embodiments, the window length is 30s and the step size is 5s; in other embodiments, the window length can be calculated according to the required frequency resolution and sampling rate, and the step size is 1 / 6 to 1 / 4 of the window length, to balance temporal resolution and spectral estimation stability. Within the window, a Discrete Fourier Transform is performed on the amplitude envelope sequence after subtracting the window mean. The peak frequency of the power spectral density, f_resp, is located in the 0.1Hz to 0.5Hz respiratory frequency band. Candidate values ​​for the respiratory modulation period are calculated using T_resp = 1 / f_resp, where T_resp is the candidate value (in seconds) and f_resp is the peak frequency of the power spectral density in that band. Each lead yields one candidate value, and the median of all candidates is taken as the comprehensive respiratory modulation period for that window. Leads with deviations exceeding 20% ​​are then eliminated one by one, and the median is taken again, until the deviation of each candidate is less than 20% of the comprehensive value, or the number of eliminated leads reaches half of the total number of leads. The setting of the lower and upper bounds of the frequency band is crucial to accuracy: slowly varying peaks below 0.1Hz originate from electrode impedance drift or changes in body position, while high-frequency peaks above 0.5Hz originate from residuals of short-term changes in the RR interval. If these two types of components are not included in the search range without limitation, they will pull the estimated respiratory modulation period towards the long-period or short-period ends, respectively. During an attack, the amplitude envelope sequence exhibits irregular fluctuations, with multiple peaks of similar amplitude appearing in the power spectrum. When the ratio of the highest peak power to the second highest peak power within the respiratory frequency band is less than 1.5, it is judged as a competing spurious peak. The respiratory modulation period is then calculated by taking the frequency corresponding to the single highest peak in the narrow frequency band from 0.15Hz to 0.40Hz. If the highest power peak in the narrow frequency band is still less than 3 times the average power of the frequency band, the respiratory modulation period of that window is marked as missing. Each analysis window records the comprehensive period value, the number of participating leads, the competing spurious peak flag, and two types of settings: narrow frequency band recalculation and interpolation substitution, for subsequent modulation deviation calculations to be retrieved by window.

[0023] Based on the respiratory modulation cycle, modulation deviation is identified and the deviation rate is assessed for each lead to generate a modulation depth index. The amplitude envelope sequence of each lead is divided into continuous segments according to the number of sampling points N_resp corresponding to the respiratory modulation cycle. N_resp is obtained by multiplying the duration of the respiratory modulation cycle by the sampling rate and rounding down. Within each segment, the sampling point number is linearly mapped to a phase from 0° to 360°. A quadratic polynomial least squares fit is performed with the phase angle as the independent variable and the amplitude envelope sequence sampling value as the dependent variable. Optionally, the fitting order can be second-order or third-order to accommodate asymmetrical modulation waveforms between the inspiratory and expiratory branches. The predicted value at each phase is taken as the expected modulation amplitude. The deviation rate per sampling point is e = |ah| / r, where e is the modulation deviation rate at that point, a is the measured value of the amplitude envelope sequence, h is the expected modulation amplitude, and r is the peak-to-valley difference of the amplitude envelope sequence in this segment. The root mean square of the deviation rates of all sampling points within the current respiratory modulation cycle segment is taken, which is the modulation depth index for this segment of the lead. The entire path relies solely on continuous samples of the amplitude envelope sequence and the duration and sampling rate of the respiratory modulation cycle. Phase fitting replaces the timing of each R-wave beat, distinguishing it from methods that directly extract each R-wave beat. When the peak-to-trough difference of a certain segment is less than 20% of the average peak-to-trough difference of the previous 5 segments in that lead, the average of the previous 5 segments is used as the denominator for this segment, and a replacement denominator label is added to the segment record. The root mean square deviation of the modulation depth index is stored using a composite index of lead number and respiratory cycle segment number for retrieval during the stratified sorting stage. The deviation rate sequence within a segment is stored concurrently with the segment using the phase angle [0°, 360°) as the independent variable for verification. The lower the modulation depth index, the closer the amplitude envelope sequence follows the respiratory rhythm. Leads with persistently high deviation rates are often caused by electrode impedance instability or local electromyographic interference, and their index rises isolatedly in individual leads. However, when the episodic phase truly begins and the amplitude envelope sequence is completely decoupled from respiration, the modulation depth index rises patchily in multiple leads during the episodic phase.

[0024] Lead confidence weights are obtained by hierarchically sorting and assigning weights based on modulation depth indices. Leads are arranged in ascending order of modulation depth index. The quartiles Q1 and Q3 of the modulation depth index for all leads are calculated: leads below Q1 are assigned to the high-confidence layer, leads between Q1 and Q3 to the medium-confidence layer, and leads above Q3 to the low-confidence layer. The base weight for the high-confidence layer is set to 1.0. The medium-confidence layer is assigned a weight of 0.4 to 1.0 based on linear interpolation of the modulation depth index between Q1 and Q3. The low-confidence layer is uniformly set to 0.2. In this example, the base weights for the high, medium, and low layers are 1.0, 0.4 to 1.0 linearly, and 0.2, respectively. In different embodiments, the upper and lower bounds of each layer can be fine-tuned according to the proportion of the three layers and the target fusion signal-to-noise ratio. In a standard 12-lead system, adjacent limb leads share electrodes. Loosening of a single electrode can cause the modulation depth indices of leads II, III, and aVF to rise simultaneously. Therefore, a common-source discrimination is first performed. When the modulation depth indices of these three leads simultaneously exceed twice the mean of the other leads, it is considered common-source interference, and the baseline weight is forcibly reduced to 0.05 and labeled as a common-source lead group. The baseline weight of each lead is normalized by dividing by the sum of all leads to obtain the lead confidence weight vector. This vector is arranged by lead number, with an element dimension equal to the total number of leads, and the sum of all elements is always 1.0. During weighted fusion, this vector is directly scaled and accumulated lead by lead. The reason for using quartiles relative stratification instead of a fixed threshold is that different patients have significant differences in respiratory depth and electrode impedance, resulting in vastly different absolute ranges in the modulation depth indices of each lead. Relative stratification allows the lead confidence weights to reflect the relative quality of each lead signal in patients of varying body types, rather than uniformly suppressing the weights of the entire lead group in patients with shallow breathing.

[0025] Step S102: The ventricular wave group boundary is located frame by frame using the multi-lead ECG signal and lead confidence weights to generate a wave group time sequence. The wave group time sequence is used to perform cross-lead time difference frame by frame irregular tracking to form a time difference distribution matrix. The time difference distribution matrix is ​​used to evaluate the time difference abnormal segments and conduction velocity to obtain conduction abnormality markers.

[0026] In some embodiments, the step of generating a time series sequence by locating the ventricular group boundary on a beat-by-beat basis using the multi-lead ECG signal and the lead confidence weights includes: performing differential suppression of the last lead and high-weighted fusion on the multi-lead ECG signal according to the lead confidence weights to generate a weighted fused ECG; performing ventricular group boundary detection on a beat-by-beat basis on the weighted fused ECG to obtain group boundary markers; identifying the drift direction of residual offset on a beat-by-beat basis for the group boundary markers to form an offset direction sequence; and performing beat-by-beat heart rate interval and rhythm analysis based on the offset direction sequence to generate a time series sequence.

[0027] A weighted fused ECG is generated by differential suppression of the last lead and high-weight fusion of multi-lead ECG signals based on lead confidence weights. First, the lead confidence weight vectors are arranged in ascending order of weight, and the last lead with the lowest weight is located. Each sampling point of this last lead is subtracted by 0.5 times the sampling value of the adjacent high-weight lead at the same instant, resulting in a differential suppression sequence that replaces the original channel of the last lead. In one implementation, the differential coefficient is set to 0.5. In other implementations, the value can be adaptively determined based on the common-mode correlation coefficient between the last lead and the adjacent high-weight lead; when the correlation is high, the coefficient is taken to a value close to 1. The component of common-mode interference between the last lead and the adjacent high-weight lead in the broadband interference is cancelled out in the difference. Subsequently, for each sampling moment of the multi-lead ECG signal, the sampled value of each channel is multiplied by the corresponding element of the lead confidence weight vector and then summed. The sum of the confidence weights of all leads participating in the fusion is normalized. The weighted mean sequence at each moment is arranged in μV as a single-channel continuous amplitude sequence according to the original sampling rate, which is the weighted fused ECG. The total number of time-domain samples is strictly equal to that of each channel of the multi-lead ECG signal. The metadata fields record three items in parallel: the number of leads participating in the fusion, the number of the excluded last lead, and the normalized denominator value. During the onset of rapid arrhythmias such as ventricular tachycardia, the patient's limbs shake violently, and the amplitude of the motion artifact in the last lead can be several times that of the normal QRS complex. If it is directly included in the weighting, the broadband artifacts will still remain in large quantities in the onset segment of the weighted fused ECG after being scaled by the lead confidence weights. The differential operation is precisely to significantly suppress this part of the residue. When the minimum confidence weight of a lead is below 0.05, the differential calculation can no longer suppress the noise in the last lead. Therefore, the entire lead is excluded from the fusion, and its weight is subtracted from the normalized denominator. The exclusion number and reason code are written into the metadata.

[0028] The ventricular wave group boundary is detected beat-by-beat using weighted fused ECG data to obtain wave group boundary markers. A dynamic detection threshold of 1.6 times the root mean square amplitude of the weighted fused ECG historical samples over the past 200ms is used, and the instantaneous absolute values ​​are compared point-by-point. For example, the threshold is set to 1.6 and the historical window to 200ms; in other implementations, the threshold can be set to 1.4 to 2.0 and the historical window to 150ms to 300ms to adapt to different baseline heart rates. The point where the threshold is first exceeded and the exceeded state lasts for at least 20ms is marked as the candidate QRS front. The point with the largest absolute value of the weighted fused ECG in the subsequent 200ms search window is taken as the R-wave peak sampling point. The first-order difference absolute value of the weighted fused ECG is read point-by-point backward from the R-wave peak. The point where the first difference value is less than 10% of the peak value is marked as the wave group start sample number, and the wave group end sample number is obtained using the same method. Using the beat number as the primary key, {beat number, start sample number of the wave group, and end sample number of the wave group} are appended to the wave group boundary marker beat by beat. Each time domain range does not overlap, and the sample number is counted based on the weighted fused ECG sampling rate. The beat number monotonically increases from 1. If the interval between the start sample numbers of two adjacent wave groups is less than 40% of the average of the previous 10 beats, the latter is judged as a short-interval pseudo-beat and deleted. The original beat number position is filled with a missing beat marker, and the numbering is not rearranged. The benefits of dynamic thresholding are most evident in the rapid ventricular arrhythmia segment: the interval between adjacent QRS complexes is extremely compressed, and the leading edge of the subsequent beat falls on the descending slope of the preceding beat. A fixed threshold would mistakenly record the second threshold crossing of the descending segment of the preceding beat as a new beat, creating a duplicate record in the wave group boundary marker. However, the dynamic threshold is updated in real time with the power of the weighted fused ECG in the first 200ms. The threshold remains high in the descending segment after the R wave peak. When the slope crosses the threshold again, the threshold is higher than the current amplitude, and the leading edge judgment is no longer triggered. The wave group boundary marker maintains a single record per beat in the high ventricular rate segment.

[0029] For wavegroup boundary markers, residual offset is aligned and drift direction is identified frame-by-frame to form an offset direction sequence. The starting sample number of each wavegroup boundary marker is taken. The difference between the current frame and the previous frame is divided by the sampling rate to calculate the measured RR interval. The mean of the measured RR of the previous 5 frames is used to estimate the expected starting number of the current frame. The difference between the measured starting number and the expected number is then divided by the mean of the intervals between the starting numbers of the previous 5 frames to obtain the normalized relative offset. If the relative offset exceeds a threshold k, it is incremented by 1 (advanced); if it is below -k, it is decremented by 1; if the absolute value does not exceed k, it is 0. The frame number, drift direction label, measured RR interval, and missing frame flag are recorded frame-by-frame to form the offset direction sequence. For missing frame positions of wavegroup boundary markers, the drift direction is forcibly set to 0, and the missing frame flag is set to 1. The mean of the measured RR of the previous and next two frames is interpolated and substituted. The missing frame flags of all other frames are set to 0. The threshold k is taken as the 75th percentile of the absolute value of the relative offset of the most recent 30 beats from the group boundary marker, and is updated every 30 beats and written to the sequence metadata. The 30-beat rolling window and 75th percentile mentioned above are the values ​​used in this example. The more irregular the rhythm, the shorter the window length can be to speed up threshold tracking. If there are less than 30 effective beats, the standard deviation of the relative offset of all effective beats is used instead of k. If there are less than 5 effective beats, k is fixed at 0.05 and a low confidence label is attached to the corresponding record. If the drift direction of the offset sequence is +1 or -1 for 5 or more consecutive beats, it means that there is a systematic unidirectional offset at the beginning of the group, which is often seen in electrode loosening or body position change. Atrial fibrillation and other rhythm disorders cause random alternation of the offset direction sequence signs. The continuous unidirectional segment is characterized by consecutive signs of the same sign, and the random alternation segment is characterized by alternating positive and negative signs. The boundary beat number corresponds to the group boundary marker.

[0030] Based on the offset direction sequence, beat-by-beat heart rate interval and rhythm analysis is performed to generate a wave group time series. The measured RR interval of each beat in the offset direction sequence is offset and corrected according to the drift direction: when marked as +1 (when the beat is advanced and the measured RR is shorter), r_c = r_m + 0.3 × s; when marked as -1, r_c = r_m - 0.3 × s; and when marked as 0, r_c = r_m, where r_c is the corrected RR interval, r_m is the measured RR, and s is the standard deviation of the measured RR in the first 10 beats. Optionally, the correction coefficient is 0.3, and the standard deviation window is the first 10 beats, or it can be 0.2 to 0.4 with an adaptive window length according to the heart rate. The rhythm type is determined by the ratio of the difference between two adjacent corrected RR intervals (dRR) to the mean of the corrected RR intervals of the previous 10 beats: a ratio below 0.1 is recorded as "regular," 0.1 to 0.3 as "slightly irregular," and above 0.3 as "significantly irregular." The beat number, corrected RR interval, and rhythm type are recorded for each beat, forming a wave group time sequence. The beat number corresponds one-to-one with the drift direction sequence. In the regular segment where the drift direction of the drift direction sequence is 0, the correction value is almost identical to the measured value, and the wave group time sequence directly reflects the true ventricular activation interval. In the unidirectional segment, the correction amplitude increases with the standard deviation of the previous 10 beats, and the rhythm label of the wave group time sequence in this segment is closer to the actual rhythm after excluding drift. In the early stages of ventricular fibrillation, ventricular excitation is extremely irregular. In the deviation direction sequence, the measured RR contraction is sudden and the drift direction frequently alternates. Once the rhythm of the wave group time sequence is marked as "significantly irregular" for 8 or more consecutive beats, a high-risk rhythm marker is triggered and written into the corresponding segment metadata field. Among them, the wave group time sequence entries corresponding to the entries with the deviation direction sequence missing beat marker of 1 are set separately and are not included in the high-risk statistics of these 8 consecutive beats, so as to separate them from the actual excitation beats.

[0031] In some embodiments, the step of using the wave group time series to perform inter-lead time difference frame-by-frame irregularity tracking to form a time difference distribution matrix includes: using the wave group time series to perform frame-by-frame calibration of the start time of each lead to obtain the start time of each lead; performing frame-by-frame calculation of the inter-lead time difference for the start time of each lead to form a lead-pair time difference sequence; performing frame-by-frame analysis of the progressively increasing trend of time difference deviation on the lead-pair time difference sequence to obtain the time difference irregularity; and performing a comprehensive evaluation of the spatial distribution of the irregularity based on the time difference irregularity to form a time difference distribution matrix.

[0032] The start time of each lead is obtained by calibrating the start time of each lead using the wave group time series. The RR intervals of the corrected wave group time series are sequentially accumulated according to the beat number to calculate the reference sampling number of each beat in the multi-lead ECG signal. An asymmetric search window of 30ms forward and 50ms backward is set around this reference number. In some embodiments, the forward window is 30ms and the backward window is 50ms; in other embodiments, the backward window can be extended according to the T wave duration and the forward window can be shortened according to the possibility of pre-excitation. The shorter forward window and longer backward window are to avoid interference from the T wave tail on differential detection. The absolute value of the first-order difference is calculated for each effective lead within the window. The point that first exceeds twice the mean within the window is the QRS start sampling number of that lead for that beat. Using the beat number and lead number as a composite primary key and the QRS start sampling number as the value field, the start time of each lead is entered one by one, with the total number of records not exceeding the product of the number of effective leads and the number of effective beats. In cases of frequent premature ventricular contractions (PVCs), the tail of the T wave in the preceding beat and the leading edge of the QRS complex in the following beat highly overlap. The differential signal may exceed the threshold at the tail of the T wave. Therefore, when the starting sampling number of a QRS complex deviates from the reference sampling number of the same beat in the time sequence of the ventricular ventricular excitation wave by more than 50 ms, it is judged as a false T wave trigger. The search starting point is moved back 10 ms and the sample is retrieved again. If the limit is exceeded twice, the label is invalidated and the beat number is retained as a continuous number. The range of the starting sampling numbers of QRS complexes in each lead within the same beat quantifies the dispersion of the arrival time of the ventricular excitation wave in all directions. The larger the range, the more uneven the conduction in the intraventricular zone in that beat. This range is recorded along with the record for verification of excitation dispersion. Beat segments marked as "significantly irregular" in the time sequence of the ventricular ventricular contraction show QRS morphological distortion and decreased start point stability. The corresponding lead-by-lead start time is recorded with morphological distortion labels for verification. In cases of bypass conduction in ventricular pre-excitation syndrome, the QRS initiation in the bypass side leads is significantly advanced. This time shift directly marks the location of the bypass on the lead dimension of the lead-by-lead start time.

[0033] A lead pair time difference sequence is formed by calculating the time difference between lead pairs beat by beat at the start time of each lead. All pairs of valid leads at the start time of each lead are enumerated, and lead pair identifiers (ij) are defined with the smaller number as i and the larger number as j. The same lead pair is not counted twice in both directions. For each lead pair, the difference between the initial QRS sampling numbers of leads i and j is read beat by beat, multiplied by the sampling interval (the reciprocal of the sampling rate) to calculate the millisecond time difference value, and appended to the lead pair time difference sequence according to the lead pair group and beat number. When all 12 leads are valid, the lead pair time difference sequence has 66 subsequences. The length of each subsequence is equal to the number of beats in both leads of that pair that are valid. The number of valid beats in a subsequence is stored along with the metadata. When the number of leads takes other values, the number of subsequences is determined by the number of two possible combinations of the number of leads. A positive time difference value indicates that the QRS initiation in lead i is earlier than in lead j; the larger the absolute value, the greater the difference in the arrival of excitation between the two regions. In normal conduction, the time difference between ipsilateral lead pairs is stable across beats. In myocardial ischemia, the conduction delay on the blocked side fluctuates with perfusion changes. The time difference sequence of lead pairs fluctuates beat-by-beat in cross-ventricular septum lead pairs, which differs from the stable and high time difference pattern in fixed bundle branch block. When the label of a certain lead in a particular beat is invalid at the start of each lead pair, the time difference of all lead pairs containing that lead is missing for that beat. After marking the missing position in the time difference sequence of the lead pair, it is filled by linear extrapolation of the next 5 valid beats. The filled entry is accompanied by an estimated label and does not participate in subsequent trend intensity statistics. If a lead pair is missing for 20 consecutive beats, the entire subsequence is removed, and its valid beat count is recorded as 0, to avoid long-segment extrapolation values ​​contaminating subsequent trend intensity statistics.

[0034] The irregularity of time difference is obtained by performing a beat-by-beat progressive trend analysis on the lead-pair time difference sequence. For each lead-pair time difference sequence, the time difference value of each beat is taken, and the difference between adjacent beats is calculated sequentially as d_k = TD_k - TD_(k-1), where d_k is the time difference difference of the k-th beat, TD_k is the time difference value of the k-th beat, and TD_(k-1) is the time difference value of the previous beat. A sliding window of length 10 beats is used to traverse the d_k sequence. The larger of the positive and negative d_k beat counts within the window is divided by the window width of 10 to obtain the progressive trend strength of the current window. The average of the progressive trend strengths of all windows in the entire segment is taken as the scalar value of the lead-pair time difference irregularity. The scalar value and the number of participating windows are stored concurrently with the recording. The sliding window is 10 beats (in different embodiments, it can be between 8 and 16 beats). A longer window is less sensitive to short-term jitter and more sensitive to slow accumulation. When the estimated number of supplementary beats within the window exceeds 4, the window is not included in the mean because the supplementary segment d_k is obtained by linear extrapolation and the difference direction is artificially fixed; including it would artificially inflate the cumulative trend intensity. When the absolute value of the time difference sequence in a lead pair is high, but the difference direction of each beat alternates randomly, the cumulative trend intensity is close to 0.5, and the time difference irregularity output is low. For fixed bundle branch block, the time difference is consistently high, the difference is random, and the irregularity is low. For progressively worsening myocardial ischemia conduction, the time difference accumulates unidirectionally beat by beat, and the irregularity increases. The absolute values ​​of the two types of time differences are similar, but the irregularity values ​​differ. Within tens of minutes after the onset of acute myocardial infarction, the time difference value of the lead pair time difference sequence in the infarct-related lead pairs continuously increases unidirectionally in the delayed direction. The time difference irregularity first rises in these lead pairs, which can help locate the infarct-related conduction area.

[0035] A time difference distribution matrix is ​​formed by comprehensively evaluating the spatial distribution of time difference irregularities. For example... Figure 2As shown, all valid leads are sequentially labeled L1, L2, ..., Lm as lead number identifier 1. Using this as the row and column index, an m×m symmetric matrix is ​​constructed as the initial structure of the time difference distribution matrix, where m is the total number of valid leads. Each lead itself forms a diagonal zero-value element 2, which is preset to 0. The time difference irregularity records for each lead pair are traversed, and the corresponding scalar values ​​are symmetrically filled into row Li, column Lj, and row Lj, column Li. The 75th percentile of all valid elements is used as the conduction abnormality threshold marker 8. Lead pairs with positions exceeding the threshold are filled with the high irregularity element 3, those below the threshold are filled with the low irregularity element 4, and those with fewer than 15 valid beats are filled with the missing placeholder element 7 (value -1). Optionally, the abnormality threshold is taken as the 75th percentile of all valid elements, or an empirically fixed threshold adjusted according to the sensitivity of damage detection. Completed closed output: When high-irregularity element 3 is concentrated in specific rows and columns, it forms a focal cluster area 5, indicating the regional attribute of local myocardial conduction damage; when high-irregularity element 3 is scattered throughout the entire matrix and not concentrated locally, the entire time difference distribution matrix exhibits a diffuse global form 6, corresponding to ventricular fibrillation and other global ventricular conduction disorders. In the acute phase of myocardial infarction, the time difference irregularity of the lead pairs related to the infarct area increases first, and the time difference distribution matrix forms local high-value clusters in the corresponding rows and columns, while the elements of the lead pairs in the distal healthy myocardial area remain at low values. This non-uniform distribution marks the location of the infarct-related conduction area on the matrix. Note that the -1 filled in due to insufficient effective beats (less than 15 beats) is a placeholder rather than an estimated value. The time difference distribution matrix is ​​stored with row and column double indices identified by the lead number. The diagonal zero value and the -1 placeholder are masked together before spatial analysis. When reading the time difference distribution matrix, the -1 element must be filtered out before performing spatial analysis.

[0036] Anomalies in conduction were identified by assessing the conduction velocity and time difference distribution matrix. In each off-diagonal significant element of the matrix, the 75th percentile of all significant elements was used as the anomaly threshold. Lead pairs exceeding this threshold were designated as anomalous conduction pairs. Anomalous conduction segments were defined by their clustering range within the matrix, with each segment defined as a connected sub-block of the anomalous lead pair. For each anomalous conduction pair, the local conduction velocity (cm / ms) was estimated by dividing the distance between the two lead electrodes by the initial QRS duration. Lead pairs below the normal lower bound were marked with a conduction slowdown. In this example, the distance between the lead electrodes was calculated using the frontal projection coordinates of a standard 12-lead electrode; however, measured coordinates of individual electrodes could also be used to improve the accuracy of velocity estimation. Using the beat number as an index, the abnormal lead pair set, local conduction velocity value, and conduction abnormality determination are encoded into a conduction abnormality label for each beat. Beats with a non-empty abnormal lead pair set or conduction velocities below the normal lower bound are marked as having abnormal conduction. The -1 placeholder element in the time difference distribution matrix is ​​not included in the abnormality threshold statistics. If the number of valid lead pairs in a certain beat is less than half of all lead pairs, the velocity assessment sample is insufficient. The corresponding beat for the conduction abnormality label is written with a low confidence label. Subsequent sinus rhythm residual tracking reads this label and then deweights the beat to avoid the velocity estimation of beats with insufficient samples interfering with the residual template. When myocardial ischemia causes conduction slowing, abnormal lead pairs are concentrated in the leads related to the ischemic area, and the conduction abnormality label marks the spatial range of local conduction abnormality in the corresponding beat. In ventricular fibrillation and other global ventricular disturbances, abnormal lead pairs are diffuse throughout the matrix, and the conduction abnormality label shows widespread abnormalities in all leads.

[0037] Step S103: Sinus rhythm residuals are traced beat by beat to form micropotential residual sequences from multi-lead ECG signals and conduction abnormality markers. Odd-even beat alternation amplitude screening is performed based on the micropotential residual sequences to obtain an episodic prodrome score. Episodic prodrome scores are then used to perform multi-lead repolarization dispersion beat by beat test to determine the warning triggering conditions.

[0038] Specifically, a micropotential residual sequence is formed by step-by-step tracking of sinus rhythm residuals in multi-lead ECG signals and conduction abnormality markers. The R-wave peak position is detected step-by-step in each lead of the multi-lead ECG signal using a sliding threshold of 1.5 times the root mean square amplitude of the past 200 ms. A QRS window is taken from 20 ms before the peak to 40 ms after the peak, and this window can be widened symmetrically or asymmetrically according to the QRS duration if necessary. Beats marked as abnormal conduction are excluded, and those marked with low confidence are downweighted in template construction due to insufficient velocity assessment samples. For the remaining normal conduction and non-low confidence QRS window segments, the median amplitude at each sampling point is used as the template value, constructing sinus rhythm reference templates point by point. Then, for all beats, the QRS window segments are subtracted from the templates point by point. The residual waveform time domain length is equal to the QRS window length. The residual waveforms are written into the micropotential residual sequence in ascending order of beat number and with the lead number as the internal index. Each record contains two fields: beat number and residual waveform of each lead. Abnormal conduction beats are marked with an abnormal beat label at the end of the corresponding record. The template is constructed using the median instead of the mean to suppress the high amplitude distortion waveforms of individual ectopic excitations, ensuring that the micropotential residual sequence is dominated by the true micropotential components rather than template estimation errors. The micropotential amplitude is usually no more than 20 μV. When the number of normal conduction beats marked by conduction abnormality markers is less than 15, the sample size is insufficient for stable median estimation. Instead, a replacement template is constructed point-by-point using the median of the entire waveform from the first 30 beats at each sampling point. This template is then used to replace the markers in the micropotential residual sequence for that segment. In the His-Purkinje system, conduction slows locally during the prodromal phase of ventricular tachycardia. The proportion of abnormal conduction beats marked by conduction abnormality markers increases during this period. This corresponds to high-frequency, low-amplitude oscillations at the QRS terminals of the multi-lead ECG signals in these beats. The amplitude of the residual sequence in the micropotential residual sequence at the tail segment of these beats is significantly higher than the baseline, corresponding to measurable micropotential enhancement.

[0039] In some embodiments, the step of obtaining a pre-seizure score by screening the alternating amplitude of odd and even beats based on the micropotential residual sequence includes: grouping the micropotential residual sequence into odd and even beats and aligning it with the residual waveform to generate odd and even beat residual groups; measuring the amplitude difference between the odd and even beat residual groups at corresponding times to obtain an inter-beat amplitude difference sequence; determining the regularity and amplitude of the inter-beat amplitude difference sequence to generate an alternation significance index; and performing a comprehensive assessment of seizure risk and pre-seizure segment based on the alternation significance index to obtain a pre-seizure score.

[0040] The micro-potential residual sequence is grouped into even-odd beats and aligned with the residual waveform to generate even-odd beat residual groups. For example... Figure 3As shown, odd and even beats are distinguished by the beat sequence label 9 for each ECG beat. The micropotential residual sequence is the residual after subtracting the sinus rhythm reference template 10 (constructed from normal conduction beats) from the QRS fragments of each beat. It is split into odd beat group and even beat group according to the parity of each recorded beat number. Within the odd beat group, the median value of the residual of each lead in each beat is taken as the odd beat alignment reference for each sampling point. The time shift with the minimum mean square error with the reference is searched within 5 sampling points before and after. The effective overlapping segment 13 is truncated according to the optimal time shift. The same method is used for the even beat group. For example, the time shift search range is taken within 5 sampling points before and after. In other implementations, this range can be expanded according to the sampling rate to accommodate larger inter-beat jitter. The three fields {beat number, odd beat micropotential residual waveform 11, even beat micropotential residual waveform 12} are included in the odd and even beat residual group for each beat. The two waveform fields are stored as a sampling point sequence with a length equal to the number of samples in the effective overlapping segment 13. During the prodromal phase of ventricular fibrillation, sympathetic tone increases, and amplitude separation occurs in the high-frequency oscillations at the end of the QRS complex of the micropotential residual sequence between odd and even beats. The oscillations at the tail of the odd-beat micropotential residual waveform 11 are more progressive than those of the even-beat micropotential residual waveform 12, and the amplitude difference 14 between beats increases pairwise, exceeding the dynamic noise baseline 15 until an alternation segment marker 16 is formed, indicating the beat number range where the odd-even amplitude separation is significant. Alignment significantly improves the comparability of this segment of odd and even waveforms, providing an accurate temporal baseline for subsequent amplitude difference determination. When a beat in the micropotential residual sequence carries an abnormal beat marker, the record for that beat in the odd-even residual group is retained but marked with a pairing failure; this beat is not included in the alignment reference estimation.

[0041] For the odd-even beat residual group, the amplitude difference at corresponding time points is measured beat by beat to obtain the inter-beat amplitude difference sequence. In each record of the odd-even beat residual group, the number of effective overlapping segments of the aligned residual waveforms for odd and even beats is equal, and both are the maximum common overlap length determined by alignment. Within the overlapping segment, the even beat alignment residual is subtracted from the odd beat alignment residual for each lead and each sampling point. A positive sign indicates that the amplitude of the even beat is higher than that of the odd beat at that time, and a negative sign indicates the opposite. The absolute value quantifies the degree of amplitude separation between odd and even beats at that time. The difference values ​​of each lead are arranged into a difference vector according to the time sequence of the sampling points within the overlapping segment. The inter-beat amplitude difference sequence is formed using the beat number and lead number as a composite index. Each record contains three fields: beat number, lead number, and the difference vector for each sampling point. Each valid lead under the same beat number generates an independent record. When a pairing failure label is attached to the odd-even beat residual group record, the sampling points of the difference vector of all leads in the inter-beat amplitude difference sequence for that beat are forcibly set to zero and an inherited failure label is attached. Subsequent difference sign extraction actively skips this label. When the absolute value of the peak of the difference vector is concentrated at the end of the QRS complex and close to zero at the beginning, the amplitude separation of odd and even beats originates from the early repolarization process rather than the depolarization process, and has independent temporal localization value. A uniformly high amplitude throughout the QRS complex indicates a holistic amplitude alternation between depolarization and repolarization. These two types of patterns can be distinguished by the peak time of the difference vector in the inter-beat amplitude difference sequence. The peak time of the difference vector is recorded simultaneously to identify the early repolarization and holistic amplitude separation. Repolarization abnormalities spread from high-risk leads to adjacent leads during the prodromal period of ventricular fibrillation. The interlead dispersion of the root mean square amplitude of the inter-beat amplitude difference sequence in each lead continuously decreases as it spreads. When the differences in multiple leads tend to be consistent, the inter-beat amplitude difference sequence in a single lead can no longer reveal the spatial distribution; at this point, multi-lead joint analysis must be used.

[0042] For example, the step of generating an alternation significance index by determining the regularity and amplitude of the inter-beat amplitude difference sequence includes: marking the difference symbols on a beat-by-beat basis to generate a difference symbol sequence; identifying the boundaries of alternation duration and interruption segments in the difference symbol sequence to obtain alternation segment markers; tracking the inter-lead consistency of the difference amplitude and confirming the number of continuous beats to form an alternation intensity quantity; and performing a joint threshold assessment of amplitude stability and duration based on the alternation intensity quantity to obtain an alternation significance index.

[0043] A difference symbol sequence is generated by marking the difference symbols on a frame-by-frame basis for the inter-frame amplitude difference sequence. The odd-even amplitude difference of each sample point in the difference vector of each record in the inter-frame amplitude difference sequence is taken. The root mean square of all sample points of the difference vectors of the most recent 20 frames is used as the dynamic noise reference b, which is updated every 20 frames. Differences exceeding 0.2b are marked with +1, those below -0.2b with -1, and absolute values ​​not exceeding 0.2b with 0. The symbols at each sample point are arranged into a symbol vector of the same length as the difference vector. This symbol is indexed by the frame number and lead number and included in the difference symbol sequence. Each record contains three fields: lead number, frame number, and symbol vector. Each element of the symbol vector takes only the values ​​+1, -1, and 0. Optionally, the noise threshold is 0.2b, the rolling window is 20 frames, or the threshold can be between 0.15b and 0.3b to compromise between sensitivity and noise resistance. The proportion of non-zero elements in the sign vector quantifies the proportion of effective directional information carried by the difference vector in that beat. A low proportion indicates insufficient directional information for that beat, which is then uniformly handled according to the non-zero proportion threshold during the alternation segment identification stage. +1 and -1, when alternating regularly over multiple consecutive beats, indicate the inter-beat periodic reversal of the oscillation direction at the end of the odd-even beat QRS complex, representing the symbolic representation of micro-potential repolarity dispersion modulation. During an acute T-wave alternation attack, the odd-even amplitude difference sign at the end of the QRS complex in a certain lead changes from a random distribution to a strict alternation beat by beat. The difference sign sequence forms a continuous string of alternating positive and negative signs in this segment, which is not on the same order of magnitude as the sporadic short strings of occasional sign alternation. When an inheritance failure label is attached to the inter-beat amplitude difference sequence entry, the corresponding difference sign sequence record's sign vector is set to 0 and an inheritance label is attached. This is actively skipped during subsequent identification of alternation segment boundaries and is not counted in the continuous or interrupted segment beat count.

[0044] The alternation segment markers are obtained by identifying the boundaries between alternation duration and interruption segments in the difference symbol sequence. The difference between the number of +1 and -1 elements in the symbol vector of each beat in the difference symbol sequence is calculated for each lead. A positive difference is marked as +1 for the dominant symbol of the beat, a negative difference as -1, and a zero difference as 0 (a difference of 0 indicates that the number of positive and negative symbols in that beat is equal, and the direction is not dominant). Adjacent beat pairs in the dominant symbol sequence are traversed; the current beat is marked "alternating" if the sign is opposite to the previous beat, "continuous" if the sign is the same, and "interrupted" if the dominant symbol of any beat is 0. Continuous "alternating" marked segments are merged into continuous segments, and continuous "continuous" or "interrupted" marked segments are merged into interruption segments. The alternation segment marker is constructed using four fields: {lead number, segment type, start beat number, end beat number}. The beat number ranges of different segments in the same lead do not overlap. When a premature ventricular contraction disrupts normal parity pairing, the difference sign sequence experiences a break of 1 to 2 beats after the premature contraction where the dominant sign is 0. The alternation segment marker inserts an interruption at this point, and automatically resumes in the next duration when alternation resumes after the premature contraction. Alternatively, if the effective non-zero elements of the sign vector in a certain beat of the difference sign sequence are less than 30% of the total vector length, the dominant sign at that beat is forcibly set to 0, and the alternation segment marker writes a zero at that beat, causing a break in the preceding and following durations. This effective percentage threshold is 30%, and can fluctuate between 25% and 35% depending on the lead noise level. The duration beat number span indicates the number of consecutive alternation beats in the current lead, and the start and end beat positions of the interruption segment accurately locate the heartbeat number of the pairing break, which can help distinguish between brief interruptions caused by rhythm irregularities and the termination of a true alternation pattern.

[0045] Alternating segment markers are tracked for interlead consistency of amplitude differences and the number of sustained beats are confirmed to form alternating intensity. For alternating segment markers with a sustained segment type, the amplitude difference vector of each beat within the corresponding beat range is obtained according to the lead number and start / end beat number. For each beat, the mean of the absolute values ​​of each sampling point of the difference vector is taken as the beat amplitude value. The mean of all valid beat amplitude values ​​within the sustained segment is taken as the amplitude mean of that sustained segment in that lead. The standard deviation of the amplitude mean of each valid lead divided by its mean is the coefficient of variation. 1 minus the coefficient of variation is the interlead consistency. The number of sustained beats is the end beat number minus the start beat number plus 1. The sustained segment identifier, interlead consistency, number of sustained beats, and the mean amplitude of all leads are combined into a single record of alternating intensity, corresponding to one sustained segment for each alternating segment marker. When the number of effective leads within a sustained segment is less than 3, the sample size for calculating the coefficient of variation is insufficient. The interlead consistency field for alternation intensity is set to -1, while the sustained beat count and mean amplitude fields across all leads are written normally. In subsequent joint assessments, the interlead dimension is not included in the threshold comparison. Records with high interlead consistency and short sustained beat count correspond to transient synchronous alternation bursts across all leads, while records with long sustained beat count and low interlead consistency correspond to focal chronic alternation in a single lead. The quantitative differences in the alternation intensity fields between the two provide multidimensional distinguishing inputs for subsequent joint threshold assessments. Significant repolarization modulation deviations are observed in leads related to ischemic areas. The mean amplitude of the same sustained segment in the alternation region is much higher in ischemic leads than in healthy leads. In this case, the interlead consistency of alternation intensity is significantly lower, corresponding to the spatial range of focal repolarization heterogeneity.

[0046] The alternation significance index is obtained by combining the amplitude stability and duration threshold assessment based on the alternation intensity. For each alternation intensity, the translead consistency C, the number of beats N, and the mean amplitude A across all leads are taken. A is compared with the amplitude threshold t_A, and N is compared with the duration threshold t_N. When A is not lower than t_A and N is not lower than t_N, S = C × (A / t_A) × ln(N / t_N+1), where S is the alternation significance score. The scaling factor A / t_A is the amplitude over-threshold ratio, the logarithmic term increases S with the number of beats at a decreasing rate, and C applies a global weighting to S based on translead synchronicity. If A is lower than t_A or N is lower than t_N, S is set to 0. The duration identifier is used as the primary key, and S is stored in the alternation significance index in float32 format. Values ​​of S=0 are retained as placeholders and not removed. Records with higher S values ​​indicate large, prolonged, and synchronous alternating bursts across all leads, constituting the core scoring output for early warning of ventricular fibrillation precursors. A sustained early warning is triggered when five consecutive non-zero S values ​​in the alternation significance index are recorded, with the mean of each S value and the number of consecutive S values ​​in that segment written into the alternation significance index segment-level summary field. t_A and t_N are dynamically set based on the patient's individual baseline: t_A is the 80th percentile of the mean amplitude of alternation intensity across all leads over the past 5 minutes, and t_N is 1.5 times the median number of beats. In some embodiments, the sample window is the past 5 minutes, and the percentile is the 80th percentile and 1.5 times the median; in other embodiments, the window length and percentile can be adjusted as the monitoring duration increases. When there are fewer than 10 valid records within 5 minutes, t_A is fixed at 8μV and t_N at 6 beats to prevent extreme distortion of the individualized threshold under small sample sizes. For alternation intensity with translead consistency set to -1, C is fixedly replaced with 0.3 during calculation, systematically suppressing S values ​​and creating a difference in scoring level compared to complete translead records.

[0047] An attack prodrome score is obtained by comprehensively assessing the attack risk and prodrome segments based on alternating significant indicators. A sliding window of 30 records traverses all records of alternating significant indicators in order of duration segment identifiers. Records containing segment-level summary fields are weighted and accumulated using the product of the mean S and the number of consecutive records. Records with non-zero S but no segment-level summary are directly accumulated using the S value. Records with zero S are placed in the denominator but not in the numerator. The total weighted accumulation within the window is divided by the total number of records within the window to obtain the comprehensive score P. The last duration segment identifier within the window is used as the scoring anchor point. P exceeding P_H is marked "high", P between P_M and P_H is marked "medium", and P below P_M is marked "low". The total number of non-zero S values ​​within the window represents the number of prodrome duration segments, with the largest S value being the peak value S within the window. The current record of the attack prodrome score consists of the scoring anchor point duration segment identifier, the comprehensive score P, the risk level, the number of prodrome duration segments, and the peak value S within the window. The sliding window step size is 1 record, and adjacent scoring anchor points monotonically increase. A sliding window of 30 records with a step size of 1 record (the window length can be adjusted according to the density of aura events). P_H and P_M are dynamically determined by the 95th and 75th percentiles of the P-value of the aura score in the previous 5 minutes for the current patient. When there are fewer than 20 P-values, P_H=0.6 and P_M=0.3 can be fixed, and aura score records falling into this fixed threshold branch are separately marked. A low P value with an extremely high peak S value within the window corresponds to a short burst of alternation; when P remains high and the number of aura segments increases synchronously, high S records of the alternation significance index are densely stacked within the window, forming a diffuse continuous aura. In the minutes before ventricular fibrillation, the density of non-zero alternation significance index S records increases stepwise, and the comprehensive aura score P value rises accordingly. The temporal slope of P pushes the aura into the danger zone earlier than the absolute value, and can be used as a dynamic criterion for early initiation of subsequent tests.

[0048] The warning triggering conditions are determined by performing a beat-by-beat test of multi-lead repolarization dispersion based on the pre-attack aura score. Those with a "high" risk level in the pre-attack aura score are selected, and the high-risk period is identified by the duration of the score anchor point. Within this period, the QRS initiation sampling sequence number of each lead is retrieved at the start time of each lead. The T-wave amplitude window is then scanned within each lead channel of the multi-lead ECG signal, with the starting point 250 ms after the initiation and the ending point 400 ms. The maximum absolute value sampling point amplitude within the window is taken as the peak T-wave amplitude of that lead in that beat. For example, the T-wave amplitude window can be taken from 250 ms to 400 ms after the QRS initiation, or it can be adaptively shifted according to the heart rate-corrected QT interval. When the repolarization dispersion value D = T_max - T_min, where T_max and T_min are the maximum and minimum values ​​of the T wave peak amplitude in all leads, respectively; when the start time of a certain lead is marked as invalid, or when its T wave amplitude window falls within the motion artifact mark or invalid mark segment of the multi-lead ECG signal and no valid T wave peak amplitude can be obtained, the repolarization dispersion value of that lead cannot be reliably calculated, the trigger status is marked "invalid", and the repolarization dispersion value field is left empty; in other calculable leads, when the repolarization dispersion value exceeds the preset threshold for 3 consecutive leads and the corresponding pre-attack prodrome score P is not lower than P_H, the trigger status is marked "positive", where the lead number that first meets the conditions consecutively is the trigger reference lead number, and if the two conditions are not met simultaneously, it is marked "negative". The warning trigger conditions are registered for each lead, including the trigger reference lead number, trigger status, and repolarization dispersion value, and records are only generated during periods of "high" risk. Negative and invalid items are retained as placeholders and not removed. Positive entries are concentrated in the period when the comprehensive score P of the pre-attack prodrome score remains high and the repolarization dispersion value simultaneously exceeds the threshold; the combined triggering of the two indicators results in a lower false alarm rate than relying on either one alone; the pre-attack prodrome score is first pre-screened in the time domain, so that the warning triggering conditions are only activated for high-risk periods under high-load real-time monitoring, and this set of computationally intensive frame-by-frame analysis is initiated.

[0049] Step S104: Based on the warning triggering conditions, multi-lead amplitude isotropically measured beat by beat to obtain the amplitude coordination coefficient. The amplitude coordination coefficient is used to measure the contraction rate of the attenuation zone boundary to form a spatial heterogeneous distribution. The spatial heterogeneous distribution is used to locate the heterogeneous region and establish a record of the seizure type.

[0050] In some embodiments, the step of determining the amplitude coordination coefficient by measuring the amplitude in the same direction of multiple leads according to the warning triggering conditions includes: synchronously sampling the amplitude of each lead at each time step according to the warning triggering conditions to form a time-by-time amplitude sequence group; identifying the amplitude in the same direction decreasing step by step in the lead sequence group to obtain a same-direction decreasing mark; tracking the decreasing segment boundary and diffusion direction of the same-direction decreasing mark segment by segment to form an attenuation direction distribution; and confirming the attenuation segment and termination position step by step according to the attenuation direction distribution to generate the amplitude coordination coefficient.

[0051] Based on the warning trigger conditions, amplitude samples are synchronously sampled from each lead at each time step to form a time-by-time amplitude sequence group. Those triggering conditions with a "positive" trigger status are screened. The multi-lead ECG signal sampling frame converted from the first positive trigger reference beat number is taken as the starting frame, and the sampling frame corresponding to the last consecutive positive trigger reference beat number after 5 beats is taken as the ending frame. The sampling is then extended by 5 beats, and the number of extended beats can be increased or decreased according to the needs of cardioversion observation. Within the start and end frames, sampling values ​​from each valid lead channel are synchronously collected time-by-time. A two-dimensional matrix is ​​constructed using the time index as the primary key and the lead number as the column index, i.e., the time-by-time amplitude sequence group. The number of rows equals the total number of sampling points between the start and end frames, and the number of columns equals the number of valid leads. Each row stores a complete snapshot of the amplitude of all leads at that time step. The mean absolute value of the difference between two adjacent rows in the same column is used as the inter-row fluctuation. Rows that continuously exceed three times the root mean square amplitude of the corresponding column for more than 50 ms within the past 200 ms are marked as high fluctuation. Rows with "invalid" trigger conditions are marked as invalid, and invalid rows are skipped when accessing the time-by-time amplitude sequence group. During periods with high repolarization dispersion values ​​of the warning trigger conditions, the cross-column range of the corresponding row in the time-by-time amplitude sequence group increases accordingly. When the ST segment of the acute ischemia-related lead continues to sink, the mean of the column decreases monotonically, while the mean of the column in the non-ischemia lead remains stable. The horizontal distribution of the time-by-time amplitude sequence group shows a unilateral collapse. In contrast, during ventricular fibrillation, the entire column jumps synchronously and disorderly, and the column standard deviation increases diffusely. The two forms have different numerical ranges on the standard deviation distribution curve of the column direction of the time-by-time amplitude sequence group. High fluctuation rows and invalid rows are set separately for downstream recognition and skipping in the same direction.

[0052] For each time-series amplitude sequence, a frame-by-frame identification of the amplitude decreasing in the same direction is performed to obtain a "same-direction decrease" marker. The average of the absolute values ​​of the differences in each column within a 200ms row window of the time-series amplitude sequence is used as the dynamic noise reference b_n. The difference between each column of adjacent rows is calculated row by row. Columns with differences below -b_n are labeled "decreasing," and those with absolute values ​​not exceeding b_n are labeled "stable." All valid lead pairs (i,j) are enumerated. When the average markers of columns i and j simultaneously show "decreasing," that lead pair is recorded as a "same-direction decrease" event. Sampling segments where consecutive "same-direction decrease" events for each lead pair last for more than 20ms are considered "same-direction decrease intervals." The minimum duration required for merging is 20ms, and this value can be adjusted according to the sampling rate and target time resolution. Lead pair identifiers (ij), interval start and end time indices, and average descent rates are compiled into unidirectional descent markers. The average descent rate is the sum of the absolute values ​​of the differences between the two columns at each time point within the interval, divided by the number of sampling points in the interval. The total number of unidirectional descent marker records equals the sum of the number of segments that generate valid unidirectional descent intervals for each valid lead pair. The difference within the high-fluctuation marker coverage row of the time-by-time amplitude sequence group far exceeds the normal fluctuation of b_n, rendering the unidirectional descent determination physically meaningless. Therefore, the start and end lines of the high-fluctuation markers are used as truncation boundaries to forcibly split each lead pair interval into two segments, each segment being independently written into the unidirectional descent marker. The high-fluctuation row itself does not generate records. The earlier the interval start time index, the earlier the lead pair enters amplitude decay; the higher the average descent rate, the steeper the drop. In the acute phase of myocardial infarction, the average descent rate of lead pair records related to the infarct side is significantly higher, and the distribution of these lead pairs indicates the spatial direction related to the infarction.

[0053] The decay direction distribution is formed by segmentally tracking the decay segment boundaries and diffusion direction of the same-direction descent markers. All records of the same-direction descent markers are rearranged in ascending order by the interval start time index. The frontal plane coordinates of the standard 12-lead electrodes are used. The spatial direction vector is the difference between the surface projection coordinates of leads i and j in the lead pair identifier (ij), and the direction angle θ_ij is measured clockwise from the 0° axis of the frontal plane. A sliding time window of 100ms is used to traverse the same-direction descent markers. The average value of the circular calculation of θ_ij for each marker within the window, weighted by the average decay rate, is the synthesized decay direction angle for that window. Adjacent time windows with a synthesized angle deviation not exceeding 30° are merged into decay direction segments. In one implementation, the time window is 100ms and the merging angle threshold is 30°. In other implementations, the angle threshold can be relaxed or tightened based on the noise estimated by the direction. The start and end time indices of the direction segment, the synthesized decay direction angle, and the total number of lead pairs participating in the direction segment constitute a record of the decay direction distribution, corresponding to a spatially stable period of myocardial amplitude decay and diffusion. If fewer than 3 valid records are recorded within a 100ms time window, the reliability of the circular mean is insufficient, and this window is not written. When the interval between adjacent valid time windows exceeds 200ms, blank segments are inserted in the attenuation direction distribution to occupy the space, and the corresponding time range is covered by the start and end time index of the directional segments. Those with higher average descent rates have a greater contribution weight to the circular mean. The more lead pairs involved in the directional segment and the more convergent the synthesis angle, the more stable the attenuation diffusion direction is during that period. Within minutes of the onset of hyperacute myocardial infarction, the lead pairs on the infarcted side are concentrated in the same directional angle range, and the attenuation direction distribution produces dense records with stable directions. In contrast, during ventricular fibrillation, the directional vectors of all lead pairs are diffuse, and the attenuation direction distribution records are sparse.

[0054] The amplitude coordination coefficient is generated by tracking the attenuation segment and termination position step-by-step according to the attenuation direction distribution. For each valid record (excluding blank segments), the starting time index, ending time index, and total number of participating lead pairs of the direction segment are taken. All records whose starting time index falls within the range of the descending markers are filtered out, and the lead pair identifiers and ending time indices of each interval are extracted. The ending time indices of each interval are sorted in ascending order. The first terminating lead pair identifier is recorded as the attenuation leading edge lead pair identifier, and the last terminating lead pair identifier is recorded as the attenuation tail lead pair identifier. The standard deviation of the ending time indices of all filtered records is divided by the total number of participating lead pairs of the current entry in the attenuation direction distribution to obtain the step-by-step attenuation coordination degree. The starting time index of the direction segment, the step-by-step attenuation coordination degree, and the identifiers of the attenuation leading edge and attenuation tail lead pairs constitute a record of amplitude coordination coefficient, corresponding to one valid direction segment of the attenuation direction distribution. When only one valid record of the same downward descent marker is found in a certain direction segment, the standard deviation of the interval termination time cannot be calculated. The beat-by-beat attenuation coordination degree is set to -1 and a single invalid mark is attached. For blank segments in the attenuation direction distribution, placeholder records are marked with -1 in the amplitude coordination coefficient sequence. The smaller the beat-by-beat attenuation coordination degree of the amplitude coordination coefficient, the more concentrated the termination time of each lead in the same direction is, and the synchronous retreat of the attenuation wavefront. The larger the coordination degree, the more dispersed the termination time. When the excitation path of the returned wave passes through, the amplitude coordination coefficient coordination degree shows a periodic sharp increase, while the coordination degree increases slowly and continuously during diffuse attenuation.

[0055] A spatially heterogeneous distribution is formed by measuring the contraction rate of attenuation segment boundaries using amplitude coordination coefficients. In effective recording sequences where the attenuation coordination coefficient per beat is not -1, the coordinate distance between the midpoints of the projections of the leading edge leads of two adjacent effective recordings in the standard 12-lead frontal plane coordinate system is denoted as L. The time difference t is obtained by dividing the difference in the indices of the two directional segments by the sampling rate. The leading edge contraction rate v = L / t, where v is the leading edge contraction rate, L is the coordinate distance between the midpoints of the projections of the two adjacent attenuation leading edges, and t is the time difference (seconds). The tail-end contraction rate is obtained by dividing the coordinate distance between the midpoints of the tail-end leads of two adjacent effective recordings by the same time difference t. The projection midpoint is taken as the arithmetic midpoint of the frontal plane coordinates of the electrodes at both ends of the lead pair; in different embodiments, the midpoint can also be taken by weighting the lead sensitivity. The starting time index of the next directional segment, the leading and tail-end contraction rates, and the source per beat attenuation coordination coefficient are combined into a row of spatially heterogeneous distribution, corresponding to the boundary dynamic contraction characteristics between two adjacent effective recordings with amplitude coordination coefficients. When a -1 occupies a space between adjacent valid records of the amplitude coherence coefficient, the time interval between the two records exceeds the limit. The segment is split into two lines with the median time of the interval as the boundary and written into the spatial heterogeneous distribution and each attached blank segment label respectively. If there is no subsequent pairing of the last valid record of the amplitude coherence coefficient, no corresponding line is generated. The greater the difference between the leading edge contraction rate and the trailing edge contraction rate, the higher the spatial dispersion of the repolarization completion time of each region of the myocardium. After myocardial infarction, the local repolarization of the infarct area is completed early and the repolarization of the adjacent viable area is delayed. In the two adjacent records of the amplitude coherence coefficient, the leading edge of the attenuation recedes rapidly towards the infarct side while the trailing edge slowly advances on the opposite side. The difference between the two rates in the corresponding line of the spatial heterogeneous distribution continues to increase with the evolution of the infarction. In healthy sinus rhythm, both rates approach zero.

[0056] A heterogeneous region localization method is used to establish a record of seizure types based on spatial heterogeneous distribution. The absolute difference (dv) between the leading and trailing contraction rates is calculated for each row of the spatial heterogeneous distribution. A dynamic threshold is set at twice the average dv of the five preceding valid rows. Rows exceeding this threshold are marked as heterogeneous rows. The starting time index of the directional segment corresponding to consecutive high heterogeneous rows is used to locate the time range of the high heterogeneous region. The type is determined based on the average decay synergy of all rows within that interval: a mean below 0.3 is classified as "synchronous focal type," a mean above 0.7 as "dispersed diffuse type," and the rest, or alternating high and low heterogeneous rows, are classified as "mixed type." Optionally, the dual thresholds for classification are 0.3 and 0.7, and the dynamic threshold is twice the average dv of the five preceding valid rows. The dual thresholds and multipliers can be adjusted to achieve a compromise between accurate and comprehensive classification. The index of the start and end times of the high heterogeneous region's initiation and termination segments, the epidural type label, the number of high heterogeneous lines, and the mean of the source beat-by-beat decay synergy are compiled into one epidural type record, corresponding to a comprehensive qualitative output of a continuous high heterogeneous region period. When spatially heterogeneous distribution lines are labeled with cross-blank segments, the dv weight for high heterogeneity determination is reduced to 0.5 to prevent the estimated cross-blank segment rate from being misjudged as a continuous high heterogeneous region; when there are fewer than 5 effective lines, the dynamic threshold is changed to a fixed dv_0=2cm / s, and such small sample records are separately labeled in the epidural type record. "Synchronous focal" records are concentrated in the acute phase of myocardial infarction, with the amplitude of the infarct area leads decreasing synchronously, the leading edge of the spatial heterogeneous distribution rapidly receding and the tail stagnating, high dv, and low source beat-by-beat decay synergy; "dispersed diffuse" records are concentrated in the early stage of ventricular fibrillation, with both spatial heterogeneous distribution rates being low, dv approaching zero, and high source beat-by-beat decay synergy.

[0057] Step S105: Locate the seizure type and termination boundary from the seizure type record to determine the seizure conclusion parameters, perform termination frame wave group morphology tracing on the seizure conclusion parameters to generate termination cardioversion markers, and output a seizure identification report based on the consistency of the cardioversion process and type prediction by tracking the termination cardioversion markers.

[0058] Specifically, the seizure type and termination boundary are located from the seizure type record to determine the seizure conclusion parameters. In the entire seizure type record, several consecutive records with the same seizure type label and an interval of no more than 3 sampling frames between the first and last time indices of two adjacent high heterogeneous regions constitute a single seizure event. The first high heterogeneous region start time index is merged as the seizure start time index, and the last high heterogeneous region end time index is merged as the seizure termination time index. A weighted average of the source beat-by-beat attenuation synergy is calculated using the number of each high heterogeneous line as the weight, yielding the seizure period synergy representative value. Moving forward from the seizure termination time index, the first valid beat number is retrieved from the wave group time sequence and recorded as the termination beat number. The seizure type label, seizure start and termination time indices, termination beat number, and seizure period synergy representative value are combined into a single record for the seizure conclusion parameters. Each record is arranged in ascending order by the seizure start time index. Multiple seizure events within the same monitoring period each occupy an independent record. The merging interval threshold is 3 sampling frames, or it can be converted to a fixed duration based on the sampling rate and then the number of frames is calculated. When two adjacent episode type records are inconsistent, they should not be merged; each should be written independently into the episode conclusion parameter. At the position where the type changes, a spaced-out record labeled "mixed type" should be added. After cardioversion of persistent ventricular fibrillation by external shock, the myocardium is in a brief electrophysiological refractory period, and the risk of a second episode within tens of seconds is high. When the time interval between two highly heterogeneous areas in the episode type record is less than 5 beats, they should be recorded as the same episode conclusion parameter with a recurrence label. When the interval is not less than 5 beats, they should be split into two independent records to avoid the cardioversion interval being included in the duration of the first episode and thus underestimating the independence of the recurrence event.

[0059] In some embodiments, the step of performing a morphological tracing determination of the termination frame to generate a cardioversion termination marker on the epidural conclusion parameters includes: locating the termination frame waveform segment from the boundary of the waveform segment three beats before and one beat after the termination frame from the epidural conclusion parameters; performing a lead-by-lead comparison of the termination frame waveform segment with the cardioversion waveform to determine the termination frame waveform deviation; performing a historical template-by-class matching residual assessment on the termination frame waveform deviation to obtain a tracing matching result; and performing a comprehensive assessment of matching confidence and type attribution based on the tracing matching result to generate a cardioversion termination marker.

[0060] The termination frame waveform fragment is obtained from the three beats before and one beat after the termination frame, based on the termination beat number of the attack conclusion parameters. The cardioversion boundary of this attack is located according to the termination beat number of the attack conclusion parameters. Using this beat number as a reference, three beats are searched forward and one beat backward in the waveform boundary marker, extracting a total of five beats of waveform start and termination sample numbers. Sampling values ​​are extracted from the corresponding channels of the multi-lead ECG signal according to the sample range of each beat and lead. The five items—beat number, lead number, waveform fragment sampling value sequence, and fragment start and end sample numbers—are recorded one by one as the termination frame waveform fragment. In some embodiments, three beats are taken forward and one beat backward, for a total of five beats. In other embodiments, the number of beats before and after can be adjusted according to the morphological observation needs before and after cardioversion. The termination frame waveform fragment is recorded in five beats. The fourth beat corresponds to the termination beat, and the fifth beat corresponds to the first beat after cardioversion. The waveform fragment sampling value sequence at the missing beat position is set to all zeros and a missing beat label is added. When the boundary marker entry corresponding to the termination beat number in the episodic conclusion parameter is occupied by a missing beat marker, the 5-beat window is repositioned using the termination beat number followed by one beat as a replacement reference. The metadata field of the termination frame waveform fragment records the replacement beat number and the replacement reason annotation. At the moment of arrhythmia termination, the intraventricular excitation switches from the abnormal path of the episodic phase to the normal conduction of the His-Purkinje system. The waveforms of each lead in the 4th and 5th records of the termination frame waveform fragment show transitional deformation. The first three beats retain the QRS morphological characteristics of the episodic phase, and the last beat reflects the morphology of the cardioversion phase. The difference in morphology across beats quantifies the waveform jump amplitude of the cardioversion boundary for subsequent source tracing and matching.

[0061] The termination frame morphology deviation is determined by comparing the termination frame morphology with the cardioversion morphology lead by lead in the termination frame waveform group segment. The sampled value sequence of each lead waveform segment from the 4th record (termination beat) of the termination frame waveform group segment is taken as the termination frame morphology, and the sampled value sequence of each lead waveform segment from the 5th record (first beat after cardioversion) is taken as the cardioversion morphology reference. After extracting the common effective sampling length of the two records, the deviation value sequence is obtained by subtracting the cardioversion morphology reference sampling value at the same position from the termination frame morphology sampling value for each lead and sampling point. This deviation value sequence is of equal length to the two input waveforms and corresponds one-to-one with each sampling point. The root mean square (RMS) of the absolute value of each sampling point in the deviation value sequence is calculated to obtain the root mean square (RMS) deviation amplitude for that lead. The lead number, deviation vector, and RMS deviation amplitude together constitute the termination frame morphology deviation for each lead. The number of records equals the number of effective leads, and the length of the deviation vector equals the common effective sampling length. In this example, the cardioversion morphology reference is taken as the first beat after cardioversion; alternatively, the second beat after cardioversion can be taken to avoid residual transitional deformation from the first beat. During an arrhythmia episode, the intraventricular depolarization pathway is highly abnormal. The switching of the activation pathway at the moment of termination causes an asymmetrical distribution of the deviation vector amplitude of the termination frame morphology deviation across leads. The deviation is concentrated in leads related to the conduction direction. This cross-lead deviation distribution pattern reflects the electrophysiological spatial properties of the cardioversion mechanism and can help distinguish between two types of termination mechanisms: spontaneous disruption of the re-entry path and external intervention. When the fifth record of the termination frame waveform segment (the first beat after cardioversion) has a missing beat annotation, the second beat after cardioversion is used as the cardioversion morphology reference, and the termination frame morphology deviation metadata field is written to the reference replacement annotation.

[0062] For example, the step of performing historical template class-by-class matching residual evaluation on the morphological deviation of the termination frame to obtain the source matching result includes: extracting historical template deviations of various types using the morphological deviation of the termination frame to form a template deviation library; comparing the termination frame deviation with the template deviation in the template deviation library to obtain class-by-class matching residuals; performing class-by-class evaluation of the matching residual distribution fit degree on the class-by-class matching residuals to generate a type matching degree sequence; and performing optimal type selection and confidence verification based on the type matching degree sequence to output the source matching result.

[0063] A template deviation library is formed by extracting historical template deviations of various types using terminating frame morphological deviations. The extraction range is defined by the lead numbers contained in the terminating frame morphological deviation. Records of confirmed terminating frame morphological deviations of each type are retrieved from the historical terminating frame morphological deviation library according to three categories: "synchronous focal," "dispersive," and "mixed." After filtering out abnormal records where the root mean square deviation amplitude exceeds three times the standard deviation of the historical record for the same type, the median of the historical deviation vectors of each lead corresponding to the terminating frame morphological deviation is taken point by point. The template deviation vector for that type and lead is output, along with the type label, lead number, template deviation vector, and number of template entries. This information is then added to the template deviation library. The total number of records equals the product of the number of types and the number of valid leads. Each lead of the same type occupies an independent record, and records are addressed using a combination of type label and lead number. The outlier removal threshold is three times the standard deviation (or a robust outlier criterion such as the absolute deviation of the median can be used). As the monitoring period extends and the historical termination morphology database continues to accumulate, the more template entries in the template deviation database, the more representative the template deviation vector is for the termination morphology of that type of individual. When the individual's historical records are insufficient during the initial monitoring phase, the template deviation database is filled with the population-wide median template. After an individual's records accumulate to more than 3, it switches to the individual's median template, while retaining the population template as a parallel reference to verify the rationality of the individual template. When there are fewer than 3 valid historical records for a certain type, the template deviation database records the actual number of template entries for each lead of that type in the numerical segment and adds a general template annotation, while the template deviation vector is filled with the population-wide value.

[0064] A point-by-point matching residual is obtained by comparing the termination frame deviation with the template deviation in the template deviation library. The deviation vector of each lead number in the termination frame morphological deviation is subtracted point-by-point from the template deviation vector of the same type and lead in the template deviation library, using the type label as the grouping key, after truncating the common sampling length. When the common sampling length is less than 80% of the original deviation vector length, it is padded with zeros to the original length before the entire segment is compared. Zeros are added to the padded bits, without changing the point-by-point difference of the non-padded segments. The difference in the zero-padded segment is not included in the point-by-point average absolute difference statistics. Optionally, the common length threshold can be 80% of the original length, with zero padding when insufficient, or endpoint extension and resampling alignment can be used to reduce the proportion of padded segments. The mean absolute value of each sampling point in the non-filled segment difference sequence is calculated to obtain the point-by-point mean absolute difference for that type of lead. The mean absolute difference of each effective lead is then used as the scalar residual value for the entire lead for that type. The type label, lead number, and point-by-point mean absolute difference are entered line by line for each type and lead to perform class-by-class matching residuals. For each type, a scalar residual value for the entire lead is then calculated based on the point-by-point mean absolute differences of each lead as the comprehensive matching deviation measure for that type. When a general template label is attached to a lead of a certain type in the template deviation library, the weight of the point-by-point mean absolute difference of that lead is reduced to 0.5 when calculating the scalar residual value for the entire lead. The lower the scalar residual value for the entire lead in the class-by-class matching residual, the closer the morphological deviation of the current termination frame is to the overall morphology of the historical template for that type. When the difference between the scalar residual values ​​of the two types of entire leads is less than 0.05, the cross-lead spatial pattern of the point-by-point mean absolute difference becomes the key distinguishing dimension for determining the optimal type. The fixed electrode position makes this pattern highly stable in the same patient.

[0065] A type matching degree sequence is generated by class-by-class assessment of the matching residual distribution fit. The full-lead residual scalar values ​​of the three summary records of the class-by-class matching residuals are collected. The initial matching degree value is obtained by subtracting the current type's full-lead residual scalar value from the sum of all type residual scalar values ​​and dividing by the sum of all residual scalar values. The three initial values ​​are normalized so that their sum is 1.0, and the normalized value is used as the final type matching degree. The number of leads with a type statistical weight of not less than 0.5 in the class-by-class matching residuals is used as the number of participating leads. The three fields {type label, type matching degree, number of participating leads} constitute the record of the type matching degree sequence, one record for each of the three types, arranged in descending order of type matching degree. When the scalar value of the residual of a certain type of all-lead in the class-by-class matching residuals is more than 3 times higher than the mean of the other two types, the matching degree of the corresponding type is forcibly set to 0 and placed at the end to prevent outlier residual types from compressing the matching degree discrimination of the other two types when participating in normalization. When the maximum matching degree of all three types is less than 0.4, a low-confidence label for all types is added to the metadata field of the type matching degree sequence, indicating that there is a large overall deviation between the morphological deviation of the termination frame and each type in the historical template library. The subsequent confidence degree verification will downgrade the overall matching conclusion accordingly. The competition judgment difference threshold is 0.05, the outlier multiple is 3 times, and the low confidence threshold is 0.4. The above thresholds can be adjusted according to the requirements of distinguishing reliability. When the difference between two adjacent type matching degrees is less than 0.05, a competition label is written in the metadata to indicate that the matching degree difference between the two types does not have reliable distinguishability. When the number of participating leads is less than 80% of the total number of effective leads, a low-coverage label is attached to the corresponding record. The type matching degree value is retained, but the subsequent confidence degree calculation is downweighted.

[0066] The optimal type selection and confidence verification are performed based on the type matching degree sequence to output the source matching results. The type matching degree sequence is arranged in descending order of type matching degree. The first record is labeled as the optimal matching type and its type matching degree is the optimal type matching degree. The second record is labeled as the second optimal type and its type matching degree is the second optimal type matching degree. The matching discrimination ratio q = (m_1-m_2) / m_1, where q is the matching discrimination ratio, m_1 is the optimal type matching degree, and m_2 is the second optimal type matching degree. The initial matching confidence value is the product of the optimal type matching degree and the minimum number of participating leads in the type matching degree sequence divided by the total number of effective leads. When the metadata contains competing labels, the initial matching confidence value is first multiplied by 0.8 for downgrading, and then corrected by 1.2 depending on whether the matching discrimination ratio q exceeds 0.3. The corrected value is truncated to 1.0. The competition downgrading coefficient is 0.8, the discrimination ratio threshold is 0.3, and the correction coefficient is 1.2. In different embodiments, each coefficient can be adjusted according to the clinical tolerance for misjudgment. The optimal matching type, the best and second-best matching degree, and the matching confidence score constitute a single record in the source matching result, with one record output for each event. When all three records in the type matching degree sequence are labeled with low coverage, the matching confidence score is multiplied by a degradation factor of 0.7 before being written. The termination morphology recorded in the acute phase of myocardial infarction differs significantly from the historical chronic template. In the type matching degree sequence, the difference between the synchronous focal type and the other two types widens. The optimal matching type in the source matching result corresponds to the damage area attribute of the current event. The mean matching confidence scores for spontaneous termination and interventional termination are different, and their matching confidence scores are distributed differently between the two termination mechanisms.

[0067] Based on the source matching results, a comprehensive assessment of matching confidence and type attribution is performed to generate a termination cardioversion marker. The difference between the best and second-best type matching scores in the source matching results is used to calculate the matching discrimination score. The product of the matching confidence score and the matching discrimination score is used as the comprehensive credibility scalar. If the comprehensive credibility score exceeds the threshold C_H, the type attribution is judged as "clear"; between C_M and C_H, it is judged as "possible"; and below C_M, it is judged as "to be verified". The best matching type from the source matching results is used as the cardioversion type label. The termination beat number, cardioversion type label, comprehensive credibility score, and type attribution determination are combined to form the termination cardioversion marker. The termination beat number is used to trace back the epidural conclusion parameter value, and a record is generated for each epidural event. When the matching confidence score of the source matching results is below 0.3, the comprehensive credibility score will inevitably be lower than C_M. The termination cardioversion marker type attribution determination is forcibly set to "to be verified" and a low-confidence label is added. C_H and C_M are dynamically determined based on the 90th and 60th percentiles of the overall confidence level of the patient's most recent 10 cardioversion events. When there are fewer than 10 historical events, C_H is fixed at 0.6 and C_M at 0.3. Patients treated with fixed thresholds are further annotated with fixed thresholds. In some embodiments, the 90th and 60th percentiles are used as the percentiles and the review window is the most recent 10 events. In other embodiments, the review window can be extended as cardioversion events accumulate. The optimal matching type of spontaneously terminated ventricular tachycardia is usually consistent with the epidural type label. Cardioversion is mainly due to spontaneous disruption of the reentry path and has a high overall confidence level. During electrical cardioversion, the QRS complex is affected by energy decay, resulting in a slightly lower matching confidence level. The overall confidence level of this type of terminated cardioversion label is usually lower than that of spontaneously terminated tachycardia.

[0068] Based on the consistency between the cardioversion process and type prediction, a seizure identification report is output using the termination cardioversion marker tracking. The report retrieves the termination beat number, cardioversion type label, and type attribution determination for each termination cardioversion marker. It then traces back the seizure type label, seizure start time index, seizure end time index, and seizure period synergy representative value corresponding to the termination beat number in the seizure conclusion parameters. The cardioversion type label of the termination cardioversion marker is compared with the seizure type label in the seizure conclusion parameters. If they match, the consistency label is marked "match"; otherwise, it is marked "deviation." The prediction deviation is calculated by multiplying the seizure period synergy representative value by the overall reliability. The seizure type label, seizure start and end time indexes, cardioversion type label, type attribution determination, consistency label, and prediction deviation are summarized into a single record in the seizure identification report, corresponding to the full-process qualitative output of a complete seizure event. The prediction deviation is calculated as the product of the synergy representative value and the overall reliability; morphological residual terms can be introduced to refine the deviation quantification if necessary. For episode identification reports with a "clear" type attribution and a "consistent" consistency label, the entire identification process is internally consistent and can be directly pushed to the clinical early warning system. For episode identification reports with a "pending verification" type attribution or a "deviation" consistency label, a manual review label is added, prompting clinicians to make a secondary judgment based on the original electrocardiogram waveform. When the predicted deviation of an episode identification report with a termination cardioversion marker confidence level higher than C_H and a "deviation" consistency label exceeds 0.5, the historical template update process is triggered. The morphological deviation of the current termination frame is added to the historical termination morphology library to correct the corresponding type template deviation vector, so that the next round of extraction yields the updated template.

[0069] To implement the multi-lead electrocardiogram feature recognition method for arrhythmia episodes corresponding to the above method embodiments, and to achieve the corresponding functions and technical effects. See also Figure 4 , Figure 4 This diagram illustrates a structural block diagram of a multi-lead electrocardiogram (ECG) feature recognition system 400 for arrhythmia episodes according to an embodiment of this application. For ease of explanation, only the parts relevant to this embodiment are shown. The multi-lead ECG feature recognition system 400 for arrhythmia episodes according to an embodiment of this application includes: The signal acquisition module 401 is used to acquire multi-lead electrocardiogram signals and generate lead confidence weights based on the multi-lead electrocardiogram signals to perform respiratory modulation depth recognition of each lead. Conduction analysis module 402 is used to perform beat-by-beat localization of ventricular wave group boundaries on the multi-lead ECG signals and the lead confidence weights to generate wave group time sequence, use the wave group time sequence to perform beat-by-beat irregular tracking of cross-lead time difference to form a time difference distribution matrix, and perform time difference abnormality segment and conduction velocity assessment on the time difference distribution matrix to obtain conduction abnormality markers. The early warning module 403 is used to perform sinus rhythm residual tracking on a beat-by-beat basis on the multi-lead electrocardiogram signal and the conduction abnormality marker to form a micropotential residual sequence, perform odd-even beat alternation amplitude screening based on the micropotential residual sequence to obtain an episodic early warning score, and perform multi-lead repolarization dispersion test on a beat-by-beat basis according to the episodic early warning score to determine the early warning triggering conditions. The spatial positioning module 404 is used to measure the amplitude coherence of multi-lead amplitudes step by step according to the warning triggering conditions to obtain the amplitude coordination coefficient, use the amplitude coordination coefficient to measure the contraction rate of the attenuation segment boundary to form a spatial heterogeneous distribution, and perform heterogeneous region positioning of the spatial heterogeneous distribution to establish an attack type record. The conclusion output module 405 is used to locate the seizure type and termination boundary from the seizure type record, determine the seizure conclusion parameters, perform termination frame wave group morphology tracing on the seizure conclusion parameters to generate a termination cardioversion mark, and output a seizure identification report based on the consistency between the cardioversion process and the type prediction based on the termination cardioversion mark.

[0070] The multi-lead electrocardiogram feature recognition system 400 for arrhythmia attacks described above can implement a multi-lead electrocardiogram feature recognition method for arrhythmia attacks according to the above method embodiments. The options in the above method embodiments are also applicable to this embodiment and will not be detailed here. The remaining content of this application embodiment can be referred to the content of the above method embodiments, and will not be repeated in this embodiment.

[0071] The purpose of the above embodiments is to reproduce and derive the technical solution of the present invention by way of example, and to fully describe the technical solution, purpose and effect of the present invention. The purpose is to enable the public to have a more thorough and comprehensive understanding of the disclosure of the present invention, and not to limit the scope of protection of the present invention.

Claims

1. A method for identifying multi-lead electrocardiogram features during arrhythmia episodes, characterized in that, include: Collect multi-lead electrocardiogram (ECG) signals, and perform respiratory modulation depth recognition of each lead based on the multi-lead ECG signals to generate lead confidence weights; The ventricular wave group boundary is located step-by-step using the multi-lead ECG signal and the lead confidence weights to generate a wave group time sequence. The wave group time sequence is used to perform cross-lead time difference step-by-step irregular tracking to form a time difference distribution matrix. The time difference distribution matrix is ​​used to evaluate the time difference abnormal segments and conduction velocity to obtain conduction abnormality markers. The sinus rhythm residual is traced beat by beat to form a micropotential residual sequence by the multi-lead electrocardiogram signal and the conduction abnormality marker. Based on the micropotential residual sequence, the alternation amplitude screening of odd and even beats is carried out to obtain the pre-attack score. Based on the pre-attack score, the multi-lead repolarization dispersion test is performed beat by beat to determine the warning triggering condition. Based on the aforementioned warning triggering conditions, multi-lead amplitude isotropic beat-by-beat measurement is performed to obtain the amplitude coordination coefficient. The amplitude coordination coefficient is used to measure the contraction rate of the attenuation zone boundary to form a spatial heterogeneous distribution. The spatial heterogeneous distribution is then used to locate the heterogeneous region and establish an epidural type record. The seizure type and termination boundary are located from the seizure type record to determine the seizure conclusion parameters. The seizure conclusion parameters are used to perform termination frame wave group morphology tracing to generate termination cardioversion markers. Based on the termination cardioversion markers, the cardioversion process is tracked and the consistency with the type prediction is output to generate a seizure identification report.

2. The method according to claim 1, characterized in that, The step of generating lead confidence weights by performing deep respiratory modulation recognition of each lead based on the multi-lead ECG signals includes: Amplitude envelope sequences were obtained by measuring the amplitude envelope of each lead in the multi-lead electrocardiogram signal on a cycle-by-cycle basis. The respiratory modulation period is determined by hierarchical identification of respiratory frequency bands and residual fluctuations in the amplitude envelope sequence. Based on the breathing modulation cycle, the modulation deviation of each lead is identified step by step and the deviation rate is evaluated to generate a modulation depth index; Based on the modulation depth index, lead confidence is hierarchically sorted and weighted to obtain lead confidence weights.

3. The method according to claim 1, characterized in that, The step of generating a time series sequence by locating the ventricular wave group boundaries beat by beat using the multi-lead ECG signals and the lead confidence weights includes: The multi-lead ECG signals are subjected to differential suppression of the last lead and high-weighted fusion based on the lead confidence weights to generate a weighted fused ECG. The weighted fused ECG was subjected to beat-by-beat boundary detection of ventricular wave groups to obtain wave group boundary markers; The residual offset drift direction is identified frame-by-frame for alignment with the wave group boundary markers to form an offset direction sequence; Based on the offset direction sequence, perform beat-by-beat heart rate interval and rhythm analysis to generate a wavegroup time sequence.

4. The method according to claim 1, characterized in that, The step of using the wave group time series to perform interlead time difference frame-by-frame irregular tracking to form a time difference distribution matrix includes: The start time of each lead is obtained by step-by-step calibration of the start time of each lead using the wave group time sequence; For each lead-by-lead start time, the lead-pair time difference is calculated step-by-step to form a lead-pair time difference sequence; Perform a frame-by-frame progressive increase trend analysis on the time difference sequence of the leads to obtain the time difference irregularity. Based on the aforementioned time difference irregularity, a time difference distribution matrix is ​​formed by comprehensively evaluating the spatial distribution of irregularity.

5. The method according to claim 1, characterized in that, The pre-seizure aura score obtained by screening alternating odd and even beat amplitudes based on the micropotential residual sequence includes: The micropotential residual sequence is grouped into odd and even beats and aligned with the residual waveform to generate odd and even beat residual groups; For the odd and even beat residual groups, the amplitude difference at corresponding time points is measured beat by beat to obtain the inter-beat amplitude difference sequence; The inter-beat amplitude difference sequence is analyzed for multi-beat difference regularity and amplitude determination to generate alternating significant indicators; Based on the aforementioned alternating significant indicators, a comprehensive assessment of the attack risk and prodromal region is performed to obtain an attack prodromal score.

6. The method according to claim 1, characterized in that, The step of determining the amplitude coordination coefficient by measuring the multi-lead amplitude simultaneity step by step based on the aforementioned warning triggering conditions includes: Based on the aforementioned warning triggering conditions, the amplitude of each lead is synchronously sampled at each time step to form a time-by-time amplitude sequence group; For the time-by-time amplitude sequence group, conduct lead-based identification of amplitude decreasing in the same direction to obtain a same-direction decreasing marker; The descent markers in the same direction are tracked segment by segment along the boundary of the descent section and the diffusion direction to form an attenuation direction distribution; Based on the attenuation direction distribution, the attenuation segment and the termination position are tracked and confirmed step by step to generate the amplitude coordination coefficient.

7. The method according to claim 1, characterized in that, The step of performing termination frame group morphology tracing on the epidural conclusion parameters to generate a termination cardioversion marker includes: The termination frame group segment is obtained from the group boundary of the three beats before and the next beat of the termination frame located from the outbreak conclusion parameters. The termination frame morphology deviation is determined by comparing the termination frame morphology with the complex rhythm morphology lead by lead for the termination frame group segment. The source matching result is obtained by performing a historical template-by-class matching residual evaluation on the morphological deviation of the termination frame; Based on the source matching results, a comprehensive evaluation of matching confidence and type attribution is performed to generate a termination rule marker.

8. The method according to claim 5, characterized in that, The step of generating an alternation significance index by determining the regularity and amplitude of the multi-beat difference sequence includes: A difference sign sequence is generated by marking the difference sign frame by frame for the inter-frame amplitude difference sequence; The alternation segment markers are obtained by identifying the boundaries between alternation duration segments and interruption segments in the difference symbol sequence. The alternating segment markers are tracked for interlead consistency of difference amplitude and confirmed with continuous beat count to form the alternating intensity quantity; The alternating significance index is obtained based on the joint threshold evaluation of the stability and persistence of the alternating intensity magnitude.

9. The method according to claim 7, characterized in that, The process of performing historical template-by-class matching and residual evaluation on the morphological deviation of the terminated frame to obtain the source matching result includes: The termination frame morphological deviation is used to extract various types of historical template deviations to form a template deviation library. The template deviation library is compared item by item with the termination frame deviation to obtain the class-by-class matching residual; The matching residuals are evaluated class by class to determine the fit of the matching residual distribution, generating a type matching degree sequence. Based on the type matching degree sequence, perform optimal type selection and confidence verification to output the source matching result.

10. A multi-lead electrocardiogram feature recognition system for arrhythmia attacks, characterized in that, include: The signal acquisition module is used to acquire multi-lead electrocardiogram (ECG) signals and generate lead confidence weights based on the multi-lead ECG signals to perform respiratory modulation depth recognition of each lead. The conduction analysis module is used to perform beat-by-beat localization of ventricular wave group boundaries on the multi-lead ECG signals and the lead confidence weights to generate wave group time series sequences. The wave group time series sequences are used to perform beat-by-beat irregular tracking of cross-lead time difference to form a time difference distribution matrix. The time difference distribution matrix is ​​used to evaluate the time difference abnormal segments and conduction velocity to obtain conduction abnormality markers. The early warning module is used to perform sinus rhythm residual tracking on a beat-by-beat basis on the multi-lead electrocardiogram signal and the conduction abnormality marker to form a micropotential residual sequence, perform odd-even beat alternation amplitude screening based on the micropotential residual sequence to obtain an episodic prodrome score, and perform multi-lead repolarization dispersion test on a beat-by-beat basis according to the episodic prodrome score to determine the early warning triggering conditions. The spatial positioning module is used to measure the amplitude coherence of multi-lead amplitudes step by step according to the warning triggering conditions to obtain the amplitude coordination coefficient, use the amplitude coordination coefficient to measure the contraction rate of the attenuation segment boundary to form a spatial heterogeneous distribution, and perform heterogeneous region positioning of the spatial heterogeneous distribution to establish an attack type record. The conclusion output module is used to locate the seizure type and termination boundary from the seizure type record, determine the seizure conclusion parameters, perform termination frame wave group morphology tracing on the seizure conclusion parameters to generate a termination cardioversion mark, and output a seizure identification report based on the consistency between the cardioversion process and the type prediction based on the termination cardioversion mark.

Images