Wearable device-based arrhythmia monitoring method and system

Abstract

The present application relates to the technical field of medical monitoring, and particularly relates to a heart rhythm disorder monitoring method based on a wearable device. The method collects and processes multi-lead electrocardio signals, extracts a morphology feature sequence and maps it to an adaptive topological space, identifies an abnormal path by tracking the evolution track of a morphology cluster to generate a hierarchical monitoring event, and uses the bidirectional association between the event and the electrocardio morphology to direct the shaping of the topological structure. The present application realizes dynamic and accurate identification and monitoring of the evolution mode of heart rhythm disorders.

Description

Technical Field

[0001] This invention relates to the field of medical monitoring technology, and in particular to a method and system for monitoring cardiac arrhythmias based on wearable devices. Background Technology

[0002] In the field of cardiovascular health monitoring, continuous monitoring of cardiac arrhythmias is crucial for early warning and diagnosis. Currently, wearable device-based cardiac rhythm monitoring has become a routine technique. These methods typically acquire the user's electrocardiogram (ECG) signals through single-lead or multi-lead sensors, followed by preprocessing such as filtering and noise reduction to improve signal quality. The preprocessed ECG signals undergo feature extraction, commonly involving calculating time-domain features such as RR intervals, frequency-domain features, or template-matched morphological features. The extracted features are then input into classification models or rule engines to detect specific types of arrhythmias such as atrial fibrillation and premature ventricular contractions (PVCs). The entire process aims to achieve long-term, dynamic daily monitoring, providing reference information for users and healthcare professionals.

[0003] However, these conventional monitoring methods still have limitations in practical applications. ECG signals are acquired during natural activity and are highly susceptible to interference from motion artifacts and electrode contact noise, leading to unstable signal quality. Although quality assessment is included, conventional methods often judge signal quality based on fixed thresholds or simple statistics, making it difficult to precisely distinguish between subtle residual artifacts and genuine physiological variations. This can result in the incorrect rejection of some valid but noisy signal segments, or the misuse of some low-quality signals, thus affecting the reliability of subsequent feature extraction and causing missed or false alarms in event detection.

[0004] Another significant drawback lies in the insufficient characterization of the temporal evolution patterns of electrocardiogram (ECG) morphology. The occurrence and development of arrhythmias is a dynamic process; their ECG morphological features do not exist in isolation but exhibit specific evolutionary patterns over time. Conventional feature extraction methods often focus on extracting static features from individual heartbeats or short segments, lacking modeling of the continuous trajectory of morphological feature changes over long time scales. Even when using temporal models, they typically rely on pre-defined fixed feature spaces and distance metrics, making it difficult to adaptively capture individualized, nonlinear ECG morphological manifold structures. This inadequacy in capturing deep temporal evolution patterns limits the system's ability to identify latent and progressive arrhythmias early, and also makes it difficult to accurately assess the severity and development trend of abnormal events. Summary of the Invention

[0005] The present invention provides a method and system for monitoring cardiac arrhythmias based on wearable devices, which can solve the problems in the prior art.

[0006] A first aspect of the present invention provides a method for monitoring cardiac arrhythmias based on a wearable device, comprising: Multi-lead electrocardiogram (ECG) signals of the target object are collected during a continuous monitoring period. The multi-lead ECG signals are timestamped and the signal quality is evaluated. Valid ECG signal segments that meet the quality threshold are selected. Multi-scale morphological decomposition is performed on the valid ECG signal segments to extract multi-dimensional morphological feature vectors and perform temporal correlation encoding to generate morphological feature sequences. An adaptive topological mapping space based on the morphological feature sequence is constructed. By calculating the geodesic distance and manifold embedding relationship between the morphological feature sequences, the original dimensional morphological feature sequence is projected onto the compressed dimensional mapping space to form a dynamically evolving morphological cluster. State transition paths are defined in the mapping space of the compressed dimension. By tracking the temporal migration trajectory and cross-cluster jump behavior of the morphological clusters, abnormal path segments that deviate from the normal evolutionary pattern are identified. Hierarchical monitoring events are generated based on the topological characteristics and duration of the abnormal path segments. A bidirectional tracing association is established between the graded monitoring events and the multi-temporal electrocardiogram morphological evolution. The bidirectional tracing association is used to selectively enhance the local curvature weight of geodesic distance and the topological sensitive domain of abnormal path determination in the adaptive topological mapping space, thereby forming a topological structure directional shaping.

[0007] The multi-lead ECG signals are timestamped and their signal quality is assessed. Valid ECG signal segments that meet the quality threshold are selected. Multi-scale morphological decomposition is performed on these valid ECG signal segments to extract multi-dimensional morphological feature vectors and perform temporal correlation encoding, generating a morphological feature sequence including: Cross-lead benchmark alignment is performed on the acquisition time of each lead in the multi-lead electrocardiogram signal to construct a global time synchronization index, and the instantaneous phase correspondence between each lead is established based on the global time synchronization index; The instantaneous phase correspondence is used to calculate the morphological consistency deviation and translead waveform propagation delay distribution of each lead signal at the synchronization moment. The morphological consistency deviation and translead waveform propagation delay distribution are used as topological evaluation indicators of signal quality, and effective ECG signal segments that meet the quality threshold of the topological evaluation indicators are screened out. The effective ECG signal segment is morphologically decomposed under multiple scale windows. The nonlinear coupling features of waveform curvature tensor and interwave period are extracted in each scale window, and the manifold distance of the above features between adjacent scale windows is calculated. The curvature tensor, nonlinear coupling features and manifold distance between scale windows are combined to form a multidimensional morphological feature vector. The multidimensional morphological feature vectors are arranged in a time series. By calculating the co-evolution trajectory of the continuous gradient of the inter-scale manifold distance and the curvature tensor between adjacent multidimensional morphological feature vectors, a temporal association identifier encoding the topological dependency relationship between each multidimensional morphological feature vector and its predecessor vector is generated. The temporal association identifier is embedded into the multidimensional morphological feature vector to form a morphological feature sequence.

[0008] The process of aligning the acquisition times of each lead in a multi-lead ECG signal across leads to a global time synchronization index, and establishing the instantaneous phase correspondence between leads based on the global time synchronization index, includes: Identify the characteristic peak times of each lead in a multi-lead electrocardiogram signal, align the characteristic peak times with the same physiological rhythm in each lead signal on the time axis, and determine the time offset between each lead signal; Based on the time offset, the acquisition time of each lead signal is compensated and corrected so that the signal characteristics of each lead signal corresponding to the same physiological event are mapped to a unified global time on the time axis, and a unique time identifier is assigned to each global time to establish a global time synchronization index covering all leads. The instantaneous signal phase of each lead signal at each global moment is extracted using the global time synchronization index. By calculating the distribution of the difference in instantaneous signal phase between different leads and the evolution trajectory of the phase difference over time, an instantaneous phase correspondence describing the instantaneous phase difference between each lead and its time dependence is constructed.

[0009] Constructing an adaptive topological mapping space based on the morphological feature sequences, and by calculating the geodesic distance and manifold embedding relationship between the morphological feature sequences, projecting the original-dimensional morphological feature sequences onto the compressed-dimensional mapping space to form dynamically evolving morphological clusters includes: In the original dimensional space, a set of candidate geodesic paths connecting each pair of morphological feature vectors is constructed. By calculating the coupled functional of the Riemann curvature tensor integral of each candidate geodesic path along the manifold section and the covariant derivative energy of the path crossing region, the candidate geodesic path that reaches the extreme value of the coupled functional is selected as the optimal geodesic path. The arc length measure of the optimal geodesic path in the manifold metric space is defined as the geodesic distance between each pair of morphological feature vectors. Based on the geodesic distance, a geodesic neighborhood topology map of morphological feature sequence is constructed. The geodesic neighborhood topology map is subjected to Laplace-Beltramian operator eigenvalue decomposition to extract an embedded manifold coordinate system that preserves the local isometry of geodesic distance. The embedded manifold coordinate system is used as a manifold embedding relation. The original dimension morphological feature sequence is projected to a compressed dimension mapping space through a differential homeomorphism that preserves the local geodesic structure using the manifold embedding relation. In the compressed dimension of the mapping space, topological attractor cores are identified based on the geodesic density field distribution of the projected morphological feature vectors. By tracking the evolution trajectory of the tangent vector field of each topological attractor core on the time manifold and the topological invariant deformation characteristics of the attraction domain boundary, attractor cores with isomorphic tangent vector field evolution modes and topological homotopy equivalence are merged to form dynamically evolving morphological clusters.

[0010] Based on the geodesic distance, a geodesic neighborhood topology map is constructed using a sequence of morphological features. The geodesic neighborhood topology map is then subjected to Laplace-Beltramian eigenvalue decomposition to extract an embedded manifold coordinate system that preserves the local isometry of the geodesic distance. This process includes: Using each morphological feature vector in the morphological feature sequence as a node, for each node, identify the set of neighboring nodes within the geodesic distance metric space, and calculate the kernel function decay value of the geodesic distance between each node and each neighboring node in the set of neighboring nodes as the edge weight. Based on the edge weights, connections are established between nodes, and a geodesic neighborhood topology graph is constructed. For each edge connecting two nodes in the geodesic neighborhood topology graph, the principal curvature change rate when the geodesic path connecting the two nodes traverses the local section of the manifold is calculated. The principal curvature change rate is added to the corresponding edge as a geometric tensor component, and the edge weights and the geometric tensor components are coupled to form a Laplace-Beltramian operator matrix that integrates the geodesic topology and the curvature characteristics of the manifold. The Laplace-Beltrami operator matrix is ​​decomposed to obtain the eigenvalue spectrum and the corresponding eigenfunction basis. The distribution characteristics of the interval between adjacent eigenvalues ​​in the eigenvalue spectrum are analyzed. Based on the position of the spectral gaps in the interval distribution characteristics, the dominant eigenmode subspace corresponding to the intrinsic geometry of the manifold is identified. The response components of each eigenfunction basis in the dominant eigenmode subspace in the original dimension space are extracted. The response components are combined in the sequence order of the eigenfunction basis to form an embedded manifold coordinate system.

[0011] State transition paths are defined in the compressed dimension mapping space. By tracking the temporal migration trajectory and cross-cluster jump behavior of the morphological clusters, abnormal path segments deviating from the normal evolutionary pattern are identified. Based on the topological characteristics and duration of the abnormal path segments, hierarchical monitoring events are generated, including: In a compressed dimension mapping space, a topological adjacency network between morphological clusters is constructed. For each morphological cluster, a set of adjacent morphological clusters that meet the adjacency condition under the geodesic distance metric is identified. By calculating the geodesic barrier height and the topological connectivity index between each morphological cluster and each adjacent morphological cluster in the set of adjacent morphological clusters, the transition channel connecting morphological clusters and the composite metric of the geodesic barrier height and the topological connectivity index are defined as the state transition path. The movement trajectory of the morphological feature sequence on the state transition path is recorded as a temporal migration trajectory. In the temporal migration trajectory, the transition event that crosses the geodesic barrier height and exceeds the dynamic threshold is identified as the cross-cluster jump behavior. The inter-cluster topological homotopy class change corresponding to the cross-cluster jump behavior is extracted as the jump topological feature. Construct a state transition path manifold for the normal evolutionary pattern, calculate the Hausdorff distance between the temporal migration trajectory and the state transition path manifold, identify trajectory segments whose Hausdorff distance exceeds the deviation threshold as abnormal path segments, and extract the cumulative homotopy invariant of the jump topological features and the duration of the abnormal path segments. The cumulative homotopy invariant and the duration are mapped to the monitoring level to generate a graded monitoring event.

[0012] Establishing a bidirectional tracing correlation between the graded monitoring events and the multi-temporal electrocardiogram morphological evolution, and using the bidirectional tracing correlation to selectively enhance the local curvature weights of geodesic distances and the topological sensitive regions for abnormal path determination in the adaptive topological mapping space, forming a topological structure directional shaping includes: For each graded monitoring event, the morphological feature sequence evolution chain corresponding to the abnormal path segment that triggered the graded monitoring event is identified by tracing back in the time dimension, and the morphological feature sequence recovery chain after the occurrence of the graded monitoring event is recorded by tracing back in the time dimension. By calculating the topological trajectory homotopy mapping relationship between the morphological feature sequence evolution chain and the morphological feature sequence recovery chain in the mapping space of the compressed dimension, a bidirectional tracing association is established. Based on the bidirectional tracing association, the set of state transition paths traversed by the morphological feature sequence evolution chain is extracted. The topological recovery frequency of each state transition path in the morphological feature sequence recovery chain is calculated. The local curvature weights involved in the geodesic distance calculation of each state transition path are gradient enhanced according to the topological recovery frequency. The topological region covered by the abnormal path segment in the compressed dimension mapping space is extracted as the topological sensitive region. The shrinking tensor field of the boundary of the topological sensitive region is calculated using the homotopy mapping relationship of the topological trajectory. The boundary range of the topological sensitive region and the deviation threshold of the abnormal path judgment are coordinated and adjusted according to the shrinking tensor field to form the topological structure orientation shaping.

[0013] A second aspect of the present invention provides a cardiac arrhythmia monitoring system based on a wearable device, comprising: The signal acquisition unit is used to acquire multi-lead electrocardiogram (ECG) signals of the target object during a continuous monitoring period, perform timestamp calibration and signal quality assessment on the multi-lead ECG signals, screen out effective ECG signal segments that meet the quality threshold, perform multi-scale morphological decomposition on the effective ECG signal segments, extract multi-dimensional morphological feature vectors and perform time-series correlation encoding to generate morphological feature sequences. The topology mapping unit is used to construct an adaptive topology mapping space based on the morphological feature sequence. By calculating the geodesic distance and manifold embedding relationship between the morphological feature sequences, the original dimension morphological feature sequence is projected onto the compressed dimension mapping space to form a dynamically evolving morphological cluster. An abnormal path unit is used to define state transition paths in the mapping space of the compressed dimension, identify abnormal path segments that deviate from the normal evolution mode by tracking the temporal migration trajectory and cross-cluster jump behavior of the morphological clusters, and generate hierarchical monitoring events based on the topological characteristics and duration of the abnormal path segments. The topological structural unit is used to establish a bidirectional traceability association between the hierarchical monitoring events and the multi-temporal electrocardiogram morphological evolution. The bidirectional traceability association is used to selectively strengthen the local curvature weight of geodesic distance and the topological sensitive domain of abnormal path determination in the adaptive topological mapping space, thereby forming a topological structure directional shaping.

[0014] A third aspect of the present invention provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0015] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0016] This method ensures the reliability and completeness of the data used in subsequent analysis through the acquisition and quality screening of multi-lead ECG signals. Multi-scale morphological decomposition and feature extraction of effective ECG signal segments can precisely depict the complex changes of ECG waveforms in the time domain, overcoming the limitations of traditional single-feature methods in describing morphological details. Temporal correlation encoding of multi-dimensional morphological feature vectors transforms isolated waveform features into feature sequences with temporal context information, laying the foundation for capturing the dynamic evolution of ECG signals.

[0017] Constructing an adaptive topological mapping space and performing dimensionality reduction projection enables effective compression and organization of high-dimensional features based on the inherent manifold structure of the data. Calculations using geodesic distance and manifold embedding relationships allow for a more accurate measurement of the essential similarities and differences between different ECG morphological feature sequences. The dynamically evolving morphological clusters formed in the compressed space intuitively reflect the aggregation and distribution patterns of ECG states in the low-dimensional space, providing a clear structured view for state recognition.

[0018] By defining state transition paths and tracking the temporal migration trajectories of morphological clusters, continuous monitoring of changes in electrocardiographic states was achieved. By identifying aberrant path segments deviating from normal evolutionary patterns, transient or persistent abnormalities in electrocardiographic activity can be sensitively detected. Combining the topological characteristics and duration of aberrant paths to generate graded monitoring events enables quantitative assessment and classification of the severity of arrhythmic events, providing a basis for differentiated early warning and intervention. Attached Figure Description

[0019] Figure 1 This is a schematic flowchart of the arrhythmia monitoring method based on wearable devices according to an embodiment of the present invention; Figure 2 This is a flowchart of the electrocardiogram morphological topology mapping and abnormal monitoring event identification according to an embodiment of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0022] Figure 1 This is a flowchart illustrating the arrhythmia monitoring method based on wearable devices according to an embodiment of the present invention. Figure 1 As shown, the arrhythmia monitoring method based on wearable devices includes: Multi-lead electrocardiogram (ECG) signals of the target object are collected during a continuous monitoring period. The multi-lead ECG signals are timestamped and the signal quality is evaluated. Valid ECG signal segments that meet the quality threshold are selected. Multi-scale morphological decomposition is performed on the valid ECG signal segments to extract multi-dimensional morphological feature vectors and perform temporal correlation encoding to generate morphological feature sequences. An adaptive topological mapping space based on the morphological feature sequence is constructed. By calculating the geodesic distance and manifold embedding relationship between the morphological feature sequences, the original dimensional morphological feature sequence is projected onto the compressed dimensional mapping space to form a dynamically evolving morphological cluster. State transition paths are defined in the mapping space of the compressed dimension. By tracking the temporal migration trajectory and cross-cluster jump behavior of the morphological clusters, abnormal path segments that deviate from the normal evolutionary pattern are identified. Hierarchical monitoring events are generated based on the topological characteristics and duration of the abnormal path segments. A bidirectional tracing association is established between the graded monitoring events and the multi-temporal electrocardiogram morphological evolution. The bidirectional tracing association is used to selectively enhance the local curvature weight of geodesic distance and the topological sensitive domain of abnormal path determination in the adaptive topological mapping space, thereby forming a topological structure directional shaping.

[0023] In one optional implementation, the multi-lead ECG signal is timestamped and its signal quality is assessed to select valid ECG signal segments that meet a quality threshold. Multi-scale morphological decomposition is then performed on the valid ECG signal segments to extract multi-dimensional morphological feature vectors and perform temporal correlation encoding to generate a morphological feature sequence, including: Cross-lead benchmark alignment is performed on the acquisition time of each lead in the multi-lead electrocardiogram signal to construct a global time synchronization index, and the instantaneous phase correspondence between each lead is established based on the global time synchronization index; The instantaneous phase correspondence is used to calculate the morphological consistency deviation and translead waveform propagation delay distribution of each lead signal at the synchronization moment. The morphological consistency deviation and translead waveform propagation delay distribution are used as topological evaluation indicators of signal quality, and effective ECG signal segments that meet the quality threshold of the topological evaluation indicators are screened out. The effective ECG signal segment is morphologically decomposed under multiple scale windows. The nonlinear coupling features of waveform curvature tensor and interwave period are extracted in each scale window, and the manifold distance of the above features between adjacent scale windows is calculated. The curvature tensor, nonlinear coupling features and manifold distance between scale windows are combined to form a multidimensional morphological feature vector. The multidimensional morphological feature vectors are arranged in a time series. By calculating the co-evolution trajectory of the continuous gradient of the inter-scale manifold distance and the curvature tensor between adjacent multidimensional morphological feature vectors, a temporal association identifier encoding the topological dependency relationship between each multidimensional morphological feature vector and its predecessor vector is generated. The temporal association identifier is embedded into the multidimensional morphological feature vector to form a morphological feature sequence.

[0024] When processing multi-lead ECG signals, the first step is to address the time synchronization issue between different leads. During ECG signal acquisition by wearable devices, slight time shifts often exist between leads due to differences in hardware sampling circuits and signal transmission paths. To address this, a reference signal alignment strategy is employed. The lead with the highest sampling rate or best signal quality is selected as the reference lead, and its sampling time sequence is recorded. Then, interpolation methods are used to map the sampling times of other leads onto the same time axis. In practice, for a standard 12-lead ECG signal, lead II is typically chosen as the reference because its QRS complex morphology is the most typical. Cubic spline interpolation is used during alignment to ensure that the interpolated signal does not introduce significant waveform distortion. After cross-lead alignment is completed, a global time synchronization index is established. This index records the position of each sampling point on the global time axis with millisecond-level precision, enabling subsequent analysis to accurately obtain the synchronization data of each lead at any given time.

[0025] Based on the global time synchronization index, the instantaneous phase correspondence between leads is further established. A Hilbert transform is performed on the ECG signal of each lead to obtain an analytical signal form, from which instantaneous phase information is extracted. Ideally, at synchronization, the instantaneous phases of different leads should exhibit a specific correspondence; for example, the phase constraints derived from Einthoven's law should be satisfied between standard leads I, II, and III. By comparing the actual measured instantaneous phase with the theoretical phase correspondence, phase anomalies caused by poor electrode contact, electromyographic interference, or signal artifacts can be identified. In practice, the instantaneous phases of each lead at the same moment are combined to form a phase vector, and the angle between this phase vector and the theoretical phase pattern is calculated. A smaller angle indicates higher signal quality.

[0026] After obtaining the instantaneous phase correspondence, the morphological consistency deviation of each lead signal is further calculated. For spatially adjacent leads, their waveform morphology should exhibit gradual continuity, without abrupt changes or breaks. When calculating the morphological consistency deviation, signal segments within a 50-millisecond window before and after the synchronization moment of adjacent leads are selected. These segments are then dynamically time-normalized, and the Euclidean distance between the normalized signals is calculated. If the distance exceeds a preset threshold, a morphological consistency deviation is considered to exist at that moment. Simultaneously, the cross-lead waveform propagation delay distribution also needs to be calculated. When cardiac electrical excitation propagates within the myocardium, the characteristic waveforms (such as R-wave peaks) observed in different leads will exhibit a temporal sequence; this delay is typically within the range of 10 to 30 milliseconds. The time difference of R-wave peaks between adjacent leads is calculated through cross-correlation analysis, and a delay distribution histogram is plotted. Under normal circumstances, the delay distribution should be concentrated within a specific range. If abnormally dispersed delays or delays exceeding the physiologically reasonable range occur, it indicates a problem with signal quality.

[0027] This method combines morphological consistency deviation and translead waveform propagation delay distribution as a topology evaluation metric. When setting quality thresholds, morphological consistency deviation must be less than 15% of the signal amplitude, and the standard deviation of translead waveform propagation delay must be less than 8 milliseconds. Signal segments meeting these conditions are marked as valid ECG signal segments, while those not meeting the conditions are discarded or marked as low-quality segments requiring further processing. This topology-based quality evaluation method can more accurately identify hidden artifacts compared to traditional single-lead signal-to-noise ratio evaluation.

[0028] Multi-scale morphological decomposition was performed on the selected valid ECG signal segments. Multiple scale windows were set with lengths of 80 ms, 160 ms, 320 ms, and 640 ms, corresponding to the characteristic scales of different frequency components in the ECG signal. Within each scale window, the signal waveform was morphologically decomposed to extract the waveform curvature tensor. The curvature tensor describes the degree of bending of the waveform in two-dimensional space. It is calculated by performing a second-order derivative on the signal and then constructing a curvature function based on the first and second derivatives. For discrete sampled signals, the central difference method was used to approximate the calculation of the derivative. The extracted curvature tensor can capture the morphological details of the P wave, QRS complex, and T wave.

[0029] Simultaneously, nonlinear coupling features of interwave intervals are extracted within various scale windows. These intervals include the PR interval, QRS duration, and QT interval, and complex physiological regulatory relationships exist between them. When calculating the nonlinear coupling features, the mutual information method is used to quantify the dependencies between different interwave intervals. For example, the mutual information value between the PR interval and the QT interval is calculated, reflecting the degree of coordination between atrioventricular conduction and ventricular repolarization. Nonlinear coupling features can reveal intrinsic correlations in electrocardiogram signals that cannot be detected by traditional linear analysis.

[0030] Further calculations are performed on the manifold distance between features at adjacent scale windows. Unlike Euclidean distance, manifold distance considers the intrinsic geometric structure of the feature space. In the calculation, the curvature tensors and nonlinear coupling features extracted from windows at different scales are first combined into a high-dimensional feature vector. Then, a local neighborhood graph is constructed in the feature space, and the manifold distance is approximated by the shortest path distance on the graph. Manifold distance can measure the essential differences between features at different scales and capture cross-scale morphological evolution patterns. The curvature tensors, nonlinear coupling features, and inter-scale manifold distances of each scale window are concatenated dimensionally to form a multi-dimensional morphological feature vector containing 128 dimensions.

[0031] The extracted multidimensional morphological feature vectors are arranged according to time series to form a temporal feature matrix. To capture the temporal dependencies between feature vectors, the continuity gradient of the manifold distance between adjacent multidimensional morphological feature vectors is calculated. The continuity gradient reflects the rate of change of manifold distance over time; a gentle gradient indicates stable ECG morphological evolution, while a sharp gradient change corresponds to arrhythmic events. Simultaneously, the co-evolutionary trajectory of the curvature tensor is analyzed to track the synchronous change patterns of curvature tensors in different leads and at different scales over time. Principal component analysis is used to extract the dominant patterns of curvature tensor evolution, generating trajectory descriptors that reflect the co-evolutionary trend.

[0032] Based on continuous gradients and co-evolutionary trajectories, temporal correlation identifiers are generated for each multidimensional morphological feature vector.

[0033] In one optional implementation, cross-lead benchmark alignment is performed on the acquisition times of each lead in the multi-lead ECG signal to construct a global time synchronization index, and the instantaneous phase correspondence between leads is established based on the global time synchronization index, including: Identify the characteristic peak times of each lead in a multi-lead electrocardiogram signal, align the characteristic peak times with the same physiological rhythm in each lead signal on the time axis, and determine the time offset between each lead signal; Based on the time offset, the acquisition time of each lead signal is compensated and corrected so that the signal characteristics of each lead signal corresponding to the same physiological event are mapped to a unified global time on the time axis, and a unique time identifier is assigned to each global time to establish a global time synchronization index covering all leads. The instantaneous signal phase of each lead signal at each global moment is extracted using the global time synchronization index. By calculating the distribution of the difference in instantaneous signal phase between different leads and the evolution trajectory of the phase difference over time, an instantaneous phase correspondence describing the instantaneous phase difference between each lead and its time dependence is constructed.

[0034] When wearable devices acquire multi-lead ECG signals, the spatial differences in the physical distribution of the acquisition channels for each lead, coupled with slight frequency deviations in the sampling clocks between different channels, make it difficult to achieve complete synchronization of the signals across leads on the time reference. To solve this problem, cross-lead reference alignment is required for the acquisition times of each lead. Specifically, the characteristic peak times of each lead are extracted from the acquired multi-lead ECG signals. These characteristic peaks are typically selected from the R-wave peak point because it has the highest amplitude and stable morphology in the ECG signal, making it easy to automatically identify. For each lead signal, the first derivative of the signal is first obtained through differential operations. The zero-crossing point in the derivative signal is then searched, and combined with the amplitude constraints of the original signal, candidate positions of the R-wave peak are locked. To improve positioning accuracy, a parabolic fitting method is used within a window near the candidate position. By fitting a quadratic function to three to five sampling points near the peak, a sub-sampling point level estimate of the peak time is obtained.

[0035] After obtaining the characteristic peak time sequences of signals from each lead, it is necessary to align the characteristic peaks with the same physiological rhythm on the time axis. Since the R wave of the same cardiac cycle corresponds to the same ventricular depolarization event in different leads, theoretically they should occur simultaneously; however, time offsets often exist in actual data acquisition. Lead II is selected as the reference lead, and the characteristic peak times in other leads are paired with the corresponding characteristic peak times in the reference lead. The pairing process uses the minimum time distance criterion, matching the k-th R wave peak time t in the reference lead. ref，k Find the closest R-wave peak time t in other leads. i，k Where i is the lead number. Calculate the time offset. 7 represents the number of successfully paired R waves. This time offset reflects the systematic time delay of lead i relative to the reference lead.

[0036] Based on the calculated time offset, the acquisition time of each lead signal is compensated and corrected. Specifically, the corresponding time offset is subtracted from the sampling time label of lead i. This ensures that the corrected time labels are aligned with the reference leads on the same time base. After compensation and correction, the signal features corresponding to the same physiological event in each lead will be mapped to a unified global time on the time axis. To achieve efficient time alignment, a global time synchronization index structure is established. This index structure uses an ordered timestamp array, merging and sorting the corrected sampling times of all leads, removing duplicate times, and generating a monotonically increasing global time sequence. A unique time identifier is assigned to each global time, which is encoded using a 64-bit integer. The high 32 bits record the number of milliseconds since the start of monitoring, and the low 32 bits record the sub-millisecond subdivision of time, ensuring a time resolution on the microsecond level. The global time synchronization index also stores the lead activation state mask corresponding to each global time, allowing for quick determination of whether a valid sampling point for a specific lead exists at a given time through bitwise operations.

[0037] After establishing the global time synchronization index, it is necessary to extract the instantaneous signal phase of each lead signal at each global time. The instantaneous phase extraction employs the Hilbert transform method to construct an analytical signal representation for each lead signal. For the real-valued ECG signal x of lead i... i(t) Its orthogonal components are obtained through Hilbert transform. , constitute complex analytic signals Instantaneous phase is defined as , where j represents the imaginary unit. To avoid the ambiguity of the arctangent function's sign, a four-quadrant arctangent function is used in actual calculations, and the phase is continuously expanded to eliminate the 2π jump. For each global time point in the global time synchronization index, the precise instantaneous phase value of each lead at that time is obtained through linear interpolation. Even if a lead does not have an actual sampling point at that time, an estimated value can be obtained by interpolating the phase values ​​of adjacent sampling points in the time domain.

[0038] After obtaining the instantaneous phase sequence of each lead signal, the distribution of the phase difference between different leads is calculated. Lead pair (i, j) is selected, and its phase difference sequence is calculated. Statistical analysis was performed on the phase difference sequence, and its mean was calculated. and standard deviation The mean reflects the systematic phase shift between two leads, while the standard deviation reflects the stability of the phase relationship. Further analysis of the phase difference's evolution over time involves dividing the time axis into several analysis windows, each with a length of 10 cardiac cycles. Within each window, the local mean and fluctuation amplitude of the phase difference are calculated. When the local mean of the phase difference significantly shifts relative to the global mean, or when the fluctuation amplitude exceeds a preset threshold, it indicates a change in the phase coupling relationship between leads, suggesting an abnormal alteration in the cardiac conduction pathway.

[0039] A data structure describing the instantaneous phase correspondence between leads was constructed, using a sparse matrix to store the statistical characteristics of phase differences between lead pairs. The matrix row index corresponds to the source lead number, the column index corresponds to the target lead number, and the matrix elements store triples, including the mean, standard deviation, and time dependence coefficient of the phase difference. The time dependence coefficient is obtained through autocorrelation analysis of the phase difference sequence, reflecting the strength of the phase difference's memory over time. This instantaneous phase correspondence not only records the static phase difference distribution between leads but also characterizes the dynamic evolution of phase differences, providing a quantitative basis for subsequent arrhythmia pattern recognition based on the coordinated changes between leads.

[0040] In one optional implementation, an adaptive topological mapping space based on the morphological feature sequences is constructed. By calculating the geodesic distance and manifold embedding relationship between the morphological feature sequences, the original-dimensional morphological feature sequences are projected onto the compressed-dimensional mapping space to form dynamically evolving morphological clusters, including: In the original dimensional space, a set of candidate geodesic paths connecting each pair of morphological feature vectors is constructed. By calculating the coupled functional of the Riemann curvature tensor integral of each candidate geodesic path along the manifold section and the covariant derivative energy of the path crossing region, the candidate geodesic path that reaches the extreme value of the coupled functional is selected as the optimal geodesic path. The arc length measure of the optimal geodesic path in the manifold metric space is defined as the geodesic distance between each pair of morphological feature vectors. Based on the geodesic distance, a geodesic neighborhood topology map of morphological feature sequence is constructed. The geodesic neighborhood topology map is subjected to Laplace-Beltramian operator eigenvalue decomposition to extract an embedded manifold coordinate system that preserves the local isometry of geodesic distance. The embedded manifold coordinate system is used as a manifold embedding relation. The original dimension morphological feature sequence is projected to a compressed dimension mapping space through a differential homeomorphism that preserves the local geodesic structure using the manifold embedding relation. In the compressed dimension of the mapping space, topological attractor cores are identified based on the geodesic density field distribution of the projected morphological feature vectors. By tracking the evolution trajectory of the tangent vector field of each topological attractor core on the time manifold and the topological invariant deformation characteristics of the attraction domain boundary, attractor cores with isomorphic tangent vector field evolution modes and topological homotopy equivalence are merged to form dynamically evolving morphological clusters.

[0041] After acquiring valid ECG signal segments that meet the quality threshold and extracting multidimensional morphological feature vectors, these high-dimensional feature vectors need to be projected into a low-dimensional space to reveal the topological evolution of arrhythmias. For any two morphological feature vectors, there are infinitely many connecting paths in the original dimensional space. The geodesic path that truly reflects the gradual change in ECG morphology needs to be identified. In practice, a set of candidate geodesic paths is first established in the original dimensional space. This set is constructed by setting multiple connecting curves with different curvatures between the two feature vectors, and each candidate path is discretized into several path segments.

[0042] For each candidate path, the Riemann curvature tensor integral along the manifold cross section is calculated. This integral quantifies the severity of the path's bending deformation on the manifold surface; a larger curvature tensor integral value indicates that the path traverses a highly curved region on the manifold. Simultaneously, the covariant derivative energy of the region traversed by the path is calculated. This energy describes the rate of change of the morphological eigenvector along the path direction; a larger covariant derivative energy indicates a drastic change in the electrocardiogram morphology along that path. A weighted combination of the Riemann curvature tensor integral and the covariant derivative energy is used to construct a coupled functional. The weight coefficients of this functional are adaptively adjusted according to the physiological characteristics of the electrocardiogram signal. For gradually evolving arrhythmias, paths with smooth curvature are preferred, while for sudden arrhythmic events, larger covariant derivative energies are allowed.

[0043] By traversing all candidate geodesic paths and calculating their respective coupled functional values, the candidate path that reaches the minimum value of the coupled functional is selected as the optimal geodesic path. This optimal geodesic path truly reflects the intrinsic manifold distance between two morphological feature vectors. The arc length measure of the optimal geodesic path in the manifold metric space is used as the geodesic distance between the two morphological feature vectors. This geodesic distance differs from the straight-line distance in Euclidean space; it follows the intrinsic geometric structure of the manifold and can accurately characterize the degree of similarity of ECG morphology in the intrinsic topological space.

[0044] Based on the calculated geodesic distances between each pair of morphological feature sequences, a geodesic neighborhood topology graph is constructed. This graph uses morphological feature vectors as nodes and establishes edges connecting vector pairs whose geodesic distance is less than an adaptive neighborhood threshold. The geodesic neighborhood topology graph preserves the local geodesic structure of the morphological features in the original high-dimensional space, providing topological constraints for subsequent dimensionality reduction projection. Laplace-Beltramian eigenvalue decomposition is performed on the geodesic neighborhood topology graph. This operator is a generalization of the Laplace operator on Riemannian manifolds, and its eigenvectors correspond to the harmonic function basis on the manifold. By solving the eigenvalue problem of the Laplace-Beltramian operator, the eigenvectors corresponding to the first few smallest non-zero eigenvalues ​​are extracted. These eigenvectors constitute an embedded manifold coordinate system that preserves the local isometry of geodesic distances. The dimension of this coordinate system is much lower than that of the original morphological feature vectors, but it can preserve the distance relationships between feature vectors within the geodesic neighborhood to the greatest extent possible.

[0045] Using the embedded manifold coordinate system as the manifold embedding relation, the morphological feature sequence of the original dimension is projected to the compressed dimension mapping space through differential homeomorphism. This mapping process ensures the invariance of local geodesic structures; that is, morphological features with close geodesic distances in the original space remain adjacent in the compressed space, while features with greater geodesic distances are separated accordingly. The reversibility of differential homeomorphism ensures that no key topological information is lost during the projection process. The dimensions of the compressed dimension mapping space are typically set to three to five dimensions, which facilitates visualization while retaining sufficient topological distinguishability.

[0046] In the compressed-dimensional mapping space, the projected morphological feature vectors exhibit a non-uniform spatial distribution. Regions with densely clustered feature vectors correspond to stable states of ECG morphology, while sparsely distributed regions correspond to transitional states. The geodesic density field value at each location in the mapping space is calculated using kernel density estimation. Regions with high geodesic density field values ​​indicate that a large number of morphological feature vectors are close to each other in a geodesic sense. Local maxima of the geodesic density field are identified as topological attractor cores. Each attractor core represents a typical ECG morphological pattern; for example, normal sinus rhythm corresponds to a stable attractor core, while abnormal rhythms such as atrial fibrillation correspond to different attractor cores.

[0047] For each topological attractor core, its tangent vector field is calculated on the time manifold. This tangent vector field describes the instantaneous change direction and velocity of the attractor core over time. The evolution trajectory of the tangent vector field of each attractor core is tracked over consecutive monitoring periods. Stable attractor cores exhibit smooth trajectory curves, while unstable or soon-to-split attractor cores show violent oscillations in their trajectories. Simultaneously, the topological invariants of the attractor core's attraction domain boundary are calculated, including the Eulerian characteristic number and homotopy group structure of the boundary curve. These topological invariants remain constant under continuous deformation, changing only during topological abrupt changes. By comparing the evolution patterns of the tangent vector field of different attractor cores with the topological invariants of the attractor domain boundary, attractor cores with isomorphic tangent vector field evolution patterns (i.e., the evolution direction and velocity field distribution patterns are the same) and topological homotopy equivalence (i.e., consistent boundary topological invariants) are identified. These isomorphic and homotopic attractor cores are grouped into the same morphological cluster.

[0048] The morphological cluster merging process employs a hierarchical clustering algorithm. First, the spatial distribution of the tangent vector field is expanded using a Fourier series to extract the dominant mode coefficients as a quantitative representation of the evolutionary pattern. Then, the persistent cohomology characteristics of the attractor domain boundary are calculated. By calculating the bottleneck distance between the correlation coefficient of the evolutionary mode coefficients and the persistent cohomology characteristics, the isomorphic and homotopic relationships between attractor cores are determined. The merged morphological clusters dynamically evolve with the monitoring cycle, exhibiting drift in cluster center position, expansion and contraction of cluster boundaries, and splitting and merging between clusters. These dynamic evolutionary processes accurately reflect the temporal changes and state transition characteristics of the target object's electrocardiographic morphology.

[0049] In one optional implementation, constructing a geodesic neighborhood topology map based on the geodesic distance sequence, and extracting an embedded manifold coordinate system that preserves the local isometry of the geodesic distance by performing Laplace-Beltramian eigenvalue decomposition on the geodesic neighborhood topology map includes: Using each morphological feature vector in the morphological feature sequence as a node, for each node, identify the set of neighboring nodes within the geodesic distance metric space, and calculate the kernel function decay value of the geodesic distance between each node and each neighboring node in the set of neighboring nodes as the edge weight. Based on the edge weights, connections are established between nodes, and a geodesic neighborhood topology graph is constructed. For each edge connecting two nodes in the geodesic neighborhood topology graph, the principal curvature change rate when the geodesic path connecting the two nodes traverses the local section of the manifold is calculated. The principal curvature change rate is added to the corresponding edge as a geometric tensor component, and the edge weights and the geometric tensor components are coupled to form a Laplace-Beltramian operator matrix that integrates the geodesic topology and the curvature characteristics of the manifold. The Laplace-Beltrami operator matrix is ​​decomposed to obtain the eigenvalue spectrum and the corresponding eigenfunction basis. The distribution characteristics of the interval between adjacent eigenvalues ​​in the eigenvalue spectrum are analyzed. Based on the position of the spectral gaps in the interval distribution characteristics, the dominant eigenmode subspace corresponding to the intrinsic geometry of the manifold is identified. The response components of each eigenfunction basis in the dominant eigenmode subspace in the original dimension space are extracted. The response components are combined in the sequence order of the eigenfunction basis to form an embedded manifold coordinate system.

[0050] After the morphological feature sequence undergoes multi-scale morphological decomposition and temporal correlation encoding, it needs to be mapped from the high-dimensional morphological feature sequence to a low-dimensional manifold space using geodesic geometric methods. Each morphological feature vector in the morphological feature sequence is considered as an independent node in the geodesic metric space, and each node carries multi-dimensional features including P-wave morphological parameters, QRS complex features, and T-wave temporal morphology. For a specific node v... iIn the geodesic distance metric space, a geodesic neighborhood centered on the node is determined. The radius ε of this geodesic neighborhood is adaptively adjusted according to the local density of the morphological feature vector. A smaller radius is used in areas with higher density to preserve the fine topological structure, while the radius is appropriately increased in areas with sparse density to avoid the occurrence of isolated nodes.

[0051] Within the defined neighborhood of Earth, identify all nodes v. i The neighboring nodes whose geodesic distance is less than ε form a neighboring node set 7(v). i For any node v in the neighborhood node set. j compute node v i With nodes j Geodetic distance d between g (v i v j The geodesic distance is obtained by finding the shortest path length along the manifold surface in the morphological feature space. Substituting the geodesic distance into the Gaussian kernel function... The edge weight w is calculated. ij The bandwidth parameter σ controls the rate at which the distance decays with respect to the weights. The decay value of this kernel function reflects the similarity between two morphological feature vectors in the local neighborhood of the manifold. Nodes with smaller geodesic distances receive larger edge weights, ensuring that the topological graph prioritizes capturing the local continuous evolutionary relationships of morphological features.

[0052] Based on the calculated edge weights, a geodesic neighborhood topology graph is constructed by establishing connections between nodes. This graph uses nodes to represent morphological feature vectors and weighted edges to represent geodesic proximity relationships between nodes. The graph's connection pattern encapsulates the manifold structure of electrocardiogram morphological features in high-dimensional space. For the connection node v in the topology graph... i With node v j For any edge of the manifold, calculate the geometric deformation characteristics of the geodesic path connecting the two nodes as it traverses a local section of the manifold. Construct a local orthogonal coordinate system at the midpoint of the geodesic path. The tangent plane of this coordinate system is spanned by the tangent vector of the geodesic path and the two principal direction vectors of the manifold surface. As the manifold surface moves along the geodesic path, the curvature of the manifold surface in the two principal directions is determined by the principal curvatures. and Characterize and calculate the change in principal curvature of the geodesic path per unit path length to obtain the rate of change of principal curvature. , where s represents the arc length parameter along the geodesic path.

[0053] The calculated rate of change of principal curvature is used as a geometric tensor component and appended to the corresponding edge in the topological graph. This geometric tensor component captures the degree of curvature of the manifold within the local region represented by the edge, reflecting the rate of evolution of electrocardiogram morphological features in that direction. The edge weights are coupled to the geometric tensor component through a weighted combination to form modified edge weights. The adjustment coefficient α controls the strength of the curvature information's correction to the edge weights. The degree matrix D of the graph is constructed based on the corrected edge weights, and its diagonal elements... Represents node v i The total connection strength is calculated, and an adjacency matrix W is constructed, whose elements are... This represents the corrected edge weights between nodes. Calculate the Laplace-Beltrami operator matrix. This matrix integrates geodesic topology and manifold curvature characteristics, and can simultaneously maintain the local isometry of geodesic distances and the intrinsic geometric properties of manifolds in subsequent eigenvalue decompositions.

[0054] Solving the generalized eigenvalue problem by performing eigenvalue decomposition on the Laplace-Beltrami operator matrix. Obtain by eigenvalue The eigenvalue spectrum arranged in ascending order, and the corresponding eigenfunction basis. The distribution characteristics of the eigenvalue spectrum reflect the multi-scale structure of the manifold geometry, and the spacing between adjacent eigenvalues ​​reveals the relative importance of geometric modes at different scales. Calculating the spacing between adjacent eigenvalues... By analyzing the statistical distribution of the interval sequences, we can identify spectral gap locations that are significantly larger than the average interval. The appearance of a spectral gap indicates a jump in the eigenvalue spectrum at that location. The corresponding eigenfunction basis forms the dominant eigenmode subspace before this spectral gap, and this subspace captures the main variation patterns of the manifold's intrinsic geometry.

[0055] Extract the response components of each eigenfunction basis in the dominant eigenmode subspace in the original morphological feature vector dimension space. For the k-th eigenfunction basis... It is at node v i The value at the location This represents the response strength of the node in the k-th geometric mode. The response values ​​of all nodes under this eigenfunction basis are organized into a column vector, forming one dimension of the embedded coordinates.

[0056] In one optional implementation, state transition paths are defined in the mapping space of the compressed dimension. By tracking the temporal migration trajectory and cross-cluster jump behavior of the morphological clusters, abnormal path segments deviating from the normal evolutionary pattern are identified. Based on the topological characteristics and duration of the abnormal path segments, hierarchical monitoring events are generated, including: In a compressed dimension mapping space, a topological adjacency network between morphological clusters is constructed. For each morphological cluster, a set of adjacent morphological clusters that meet the adjacency condition under the geodesic distance metric is identified. By calculating the geodesic barrier height and the topological connectivity index between each morphological cluster and each adjacent morphological cluster in the set of adjacent morphological clusters, the transition channel connecting morphological clusters and the composite metric of the geodesic barrier height and the topological connectivity index are defined as the state transition path. The movement trajectory of the morphological feature sequence on the state transition path is recorded as a temporal migration trajectory. In the temporal migration trajectory, the transition event that crosses the geodesic barrier height and exceeds the dynamic threshold is identified as the cross-cluster jump behavior. The inter-cluster topological homotopy class change corresponding to the cross-cluster jump behavior is extracted as the jump topological feature. Construct a state transition path manifold for the normal evolutionary pattern, calculate the Hausdorff distance between the temporal migration trajectory and the state transition path manifold, identify trajectory segments whose Hausdorff distance exceeds the deviation threshold as abnormal path segments, and extract the cumulative homotopy invariant of the jump topological features and the duration of the abnormal path segments. The cumulative homotopy invariant and the duration are mapped to the monitoring level to generate a graded monitoring event.

[0057] In the generated compressed-dimensional mapping space, the anchor point of each morphological cluster is determined. By traversing all morphological clusters in the mapping space, the geodesic distance between the current morphological cluster and other morphological clusters is calculated. The geodesic distance is obtained by integrating along the path of minimum local curvature in the manifold embedding space. Compared with directly using Euclidean distance, the geodesic distance can more accurately reflect the true proximity relationship of morphological clusters on the manifold structure. An adjacency determination threshold is set. When the geodesic distance between two morphological clusters is less than the threshold, the corresponding morphological cluster is marked as an adjacent morphological cluster, thus establishing a set of adjacent morphological clusters for each morphological cluster. This set of adjacency relationships constitutes the basic framework of the topological adjacency network, which uses morphological clusters as nodes and inter-cluster relationships that satisfy the adjacency conditions as edges.

[0058] For each edge in the topological adjacency network, the geodesic barrier height connecting two morphological clusters is calculated. This height is obtained by analyzing the local curvature changes and morphological feature density distribution gradient along the connection path. Specifically, multiple intermediate points are sampled along the geodesic path, and the manifold curvature tensor and morphological feature density at each sampling point are extracted. The peak curvature point or minimum density point on the path is identified as the barrier location, and the barrier height is defined as the difference in morphological feature density between this location and the initial morphological cluster location. Simultaneously, the topological connectivity index of the inter-cluster boundary is calculated. This index is quantified by statistically analyzing the area of ​​the overlapping region between the two morphological clusters in the mapping space and combining it with the manifold curvature continuity at the boundary. A higher topological connectivity index indicates a more stable transition channel and more continuous morphological evolution between the two morphological clusters. The geodesic barrier height and the topological connectivity index are weighted and combined, with weights determined statistically based on historical data of different arrhythmia types, to form a composite metric. The transition channel connecting morphological clusters and its composite metric are collectively defined as the state transition path. This path describes the channel attributes and difficulty of a morphological feature sequence migrating from one morphological cluster to another in the mapping space.

[0059] Based on the state transition path, the temporally arranged morphological feature sequence is tracked. The projection positions of the morphological feature sequence in the mapping space within a continuous time window are connected in chronological order to form a temporal migration trajectory. This trajectory reflects the evolution of the target object's electrocardiographic morphology over time, with each point on the trajectory corresponding to the morphological cluster affiliation at a specific moment. By detecting changes in morphological cluster affiliation between adjacent time steps, transition events where the morphological feature sequence moves from one morphological cluster to another are identified. For each transition event, the corresponding inter-cluster geodesic barrier height is extracted, and a kinetic threshold is set as the judgment criterion. When the geodesic barrier height exceeds the kinetic threshold, it indicates that the transition event needs to overcome a high energy barrier or morphological difference, and it is marked as an inter-cluster jump behavior. The kinetic threshold is determined based on the statistical distribution of transition events under normal cardiac rhythm, and is usually set as the 95th percentile of the geodesic barrier height of normal transition events.

[0060] For the identified cross-cluster jump behavior, inter-cluster topological homotopy class changes are extracted. Topological homotopy class describes the connectivity and closed-loop characteristics of morphological clusters in a topological sense, and is quantified by calculating the difference between the fundamental group or homology group of the morphological clusters in the mapping space before and after the jump. In specific implementation, a continuous homology method is used to construct the topological skeleton of the morphological clusters, and the Betti number sequence of each morphological cluster is extracted. The Betti number reflects the number of holes in different dimensions in the topological space. The change vector of the Betti number sequence before and after the jump is the inter-cluster topological homotopy class change, and this change vector is stored as a jump topological feature. This topological feature can capture whether there is a sudden change in the topological structure during the jump of the ECG morphology, such as from a simple connected morphology to a complex morphology with multiple independent peaks.

[0061] To identify anomalous path segments deviating from the normal evolutionary pattern, a state transition path manifold of the normal evolutionary pattern needs to be pre-constructed. Using historical monitoring data labeled as normal heart rhythms, corresponding temporal migration trajectories are extracted. The probability density of the distribution of these normal trajectories in the mapping space is estimated to generate the state transition path manifold of the normal evolutionary pattern. This manifold appears as one or more high-density tubular regions in the mapping space, representing the main channels of morphological evolution under normal heart rhythms. The Hausdorff distance between the actual temporal migration trajectory within the current monitoring cycle and this manifold is calculated. The Hausdorff distance is defined as the maximum distance from any point on the trajectory to the nearest point on the manifold; this distance metric effectively identifies the overall deviation of the trajectory. A deviation threshold is set; when the Hausdorff distance exceeds this threshold, it indicates that the current trajectory significantly deviates from the normal evolutionary pattern, and the corresponding trajectory segment is marked as an anomalous path segment.

[0062] For each anomalous path segment, the topological features corresponding to all cross-cluster jump behaviors are extracted, and the cumulative homotopy invariant of these topological features is calculated. The cumulative homotopy invariant is obtained by summing the Betti number change vectors of all jump topological features within the segment; this invariant reflects the cumulative degree of change in the topological structure of the anomalous path segment. Simultaneously, the duration of the anomalous path segment is recorded, i.e., the time interval between the start and end times of the segment. The longer the duration, the longer the anomalous state persists, and the more significant the clinical significance.

[0063] A mapping rule is established between cumulative homotopy invariants, duration, and monitoring levels. The numerical range of cumulative homotopy invariants is divided into multiple intervals, and duration is categorized into short, medium, and long duration levels. Different monitoring levels are mapped based on the combination of the cumulative homotopy invariant's interval and duration level. For example, abnormal pathway segments with small cumulative homotopy invariants and short durations are mapped to low-level monitoring events, triggering only a recording notification; segments with medium cumulative homotopy invariants and medium durations are mapped to medium-level monitoring events, triggering an alert notification; and segments with large cumulative homotopy invariants or long durations are mapped to high-level monitoring events, triggering an immediate alarm and pushing it to healthcare personnel's terminals. This mapping rule generates tiered monitoring events, enabling differentiated responses to arrhythmias of varying severity, avoiding alarm fatigue caused by excessive alarms, and ensuring timely handling of critical abnormal events.

[0064] In one optional implementation, a bidirectional tracing correlation is established between the graded monitoring events and the multi-temporal electrocardiogram morphological evolution. This bidirectional tracing correlation is then used to selectively enhance the local curvature weights of geodesic distances and the topologically sensitive regions for abnormal path determination in the adaptive topological mapping space, forming a topological structure-oriented shaping process, including: For each graded monitoring event, the morphological feature sequence evolution chain corresponding to the abnormal path segment that triggered the graded monitoring event is identified by tracing back in the time dimension, and the morphological feature sequence recovery chain after the occurrence of the graded monitoring event is recorded by tracing back in the time dimension. By calculating the topological trajectory homotopy mapping relationship between the morphological feature sequence evolution chain and the morphological feature sequence recovery chain in the mapping space of the compressed dimension, a bidirectional tracing association is established. Based on the bidirectional tracing association, the set of state transition paths traversed by the morphological feature sequence evolution chain is extracted. The topological recovery frequency of each state transition path in the morphological feature sequence recovery chain is calculated. The local curvature weights involved in the geodesic distance calculation of each state transition path are gradient enhanced according to the topological recovery frequency. The topological region covered by the abnormal path segment in the compressed dimension mapping space is extracted as the topological sensitive region. The shrinking tensor field of the boundary of the topological sensitive region is calculated using the homotopy mapping relationship of the topological trajectory. The boundary range of the topological sensitive region and the deviation threshold of the abnormal path judgment are coordinated and adjusted according to the shrinking tensor field to form the topological structure orientation shaping.

[0065] like Figure 2 As shown, the method includes: After the monitoring system detects a graded monitoring event, it is necessary to establish a complete correlation mechanism between the event and the morphological evolution of the electrocardiogram (ECG) signal. For a monitoring event of a certain grade, such as a suspected atrial fibrillation event, the system first traces back the continuous time window preceding the event's trigger time on the timeline. The length of this time window can be dynamically set according to the ECG signal sampling rate, typically ranging from 30 to 120 seconds before the event trigger. Within this time window, the morphological feature sequence evolution chain associated with the monitoring event is extracted. This evolution chain records the complete morphological change process from a normal rhythm state to an abnormal state, including features such as the gradual decay of the P wave morphology, amplitude fluctuations of the QRS complex, and irregular enhancement of the RR interval.

[0066] Simultaneously, the system traces the time period following the occurrence of the monitored event to capture the recovery process of the electrocardiogram (ECG) signal morphology. This recovery chain records the temporal characteristics of whether the abnormal rhythm spontaneously terminates, recovers to normal sinus rhythm, or continues to maintain the abnormal state. By comparing the trajectory characteristics of the evolution chain and the recovery chain in the compressed dimension mapping space, the topological homotopy mapping relationship between the two trajectories is calculated. This mapping relationship reflects the degree of reversibility of the ECG morphology before and after the abnormal state is triggered.

[0067] Specifically, the coordinates of each morphological feature sequence in the evolutionary chain are represented as a set of discrete points in the mapping space. These points are connected in chronological order to form a time-series curve. Similarly, the morphological feature sequence corresponding to the recovery chain also forms another time-series curve in the mapping space. By determining whether these two curves have the potential for continuous deformation in the topological space—that is, whether one curve can be gradually transformed into the other through a smooth topological transformation while keeping the curve endpoints fixed—the existence of a homotopy mapping relationship is determined. If such a homotopy mapping exists, it indicates that the arrhythmia corresponding to the monitoring event is highly reversible, and the electrocardiogram morphology can be recovered along a similar topological path. If no homotopy mapping exists, it suggests that the arrhythmia has caused irreversible changes in the electrocardiogram morphology, requiring an increase in the monitoring priority of this event.

[0068] Based on the established bidirectional tracing association, the set of state transition paths traversed by the evolutionary chain is further analyzed. A state transition path refers to the connection trajectory of a morphological cluster migrating from one stable region to another in the mapping space. The frequency of these paths being traversed again in the recovery chain is statistically analyzed for each state transition path traversed by the evolutionary chain and defined as the topological recurrence frequency. For example, if the evolutionary chain migrates from a normal sinus rhythm cluster to an atrial premature beat cluster via a certain path, and the recovery chain returns to the normal sinus rhythm cluster along the same or topologically similar path, then the topological recurrence frequency of that path increases.

[0069] State transition paths with high topological recurrence frequency have significant clinical implications, indicating that the corresponding electrocardiographic morphological changes are reproducible and predictable across different time phases. For these high-frequency paths, gradient enhancement is applied to the local curvature weights in the geodesic distance calculation process. Geodesic distance is a measure of the shortest path between two points in the mapped space along the manifold surface, while the local curvature weights determine the impact of the manifold surface's curvature in a specific region on distance calculation. By enhancing the curvature weights in the regions containing high-frequency paths, the geodesic distance in these regions becomes more sensitive to morphological changes, thereby enabling earlier identification of potential abnormal events evolving along these paths during subsequent monitoring.

[0070] Regarding the definition and optimization of the topology-sensitive region, the actual area covered by the anomalous path segment in the compressed dimensional mapping space is first extracted. This region is jointly formed by the positions of all morphological clusters traversed by the anomalous path segment and their neighborhood ranges, creating an irregular topological region, i.e., the topology-sensitive region. The boundary of this sensitive region is not fixed but dynamically adjusted based on the homotopy mapping relationship obtained from bidirectional tracing association.

[0071] Using homotopy mapping, a contraction tensor field is calculated on the boundary of the topologically sensitive domain. This contraction tensor field describes the contraction or expansion trend of the boundary in various directions. The calculation of this tensor field is based on the morphological distribution density of the evolutionary and recovery chains near the boundary. If a boundary region exhibits frequent morphological regression in the recovery chain, it indicates strong reversibility, and the sensitive domain boundary can be appropriately contracted to reduce the sensitivity of this region to trigger monitoring events and avoid excessive false positive alarms. Conversely, if a boundary region does not exhibit morphological regression in the recovery chain, or if the evolutionary chain shows an accelerated deviation trend in this region, the sensitive domain boundary should be expanded to improve the monitoring sensitivity of this region.

[0072] Based on the distribution characteristics of the contraction tensor field, the boundary range of the topological sensitive domain and the deviation threshold for abnormal path detection are adjusted collaboratively. The deviation threshold defines the minimum geodesic distance at which a morphological feature sequence deviates from its normal evolutionary pattern in the mapping space. When the actual geodesic distance exceeds this threshold, a monitoring event is triggered. In regions where the sensitive domain boundary contracts, the deviation threshold is increased accordingly to reduce the false alarm rate in those regions; in regions where the sensitive domain boundary expands, the deviation threshold is decreased to improve the ability to capture early abnormal signals. Through this collaborative adjustment mechanism of boundary range and deviation threshold, the topological structure is directionally shaped, making the topological characteristics of the mapping space more closely match the evolutionary patterns of actual ECG signals. This effectively controls the false positive rate while ensuring monitoring sensitivity, improving the practicality and reliability of wearable devices in long-term continuous monitoring scenarios.

[0073] A second aspect of the present invention provides a cardiac arrhythmia monitoring system based on a wearable device, comprising: The signal acquisition unit is used to acquire multi-lead electrocardiogram (ECG) signals of the target object during a continuous monitoring period, perform timestamp calibration and signal quality assessment on the multi-lead ECG signals, screen out effective ECG signal segments that meet the quality threshold, perform multi-scale morphological decomposition on the effective ECG signal segments, extract multi-dimensional morphological feature vectors and perform time-series correlation encoding to generate morphological feature sequences. The topology mapping unit is used to construct an adaptive topology mapping space based on the morphological feature sequence. By calculating the geodesic distance and manifold embedding relationship between the morphological feature sequences, the original dimension morphological feature sequence is projected onto the compressed dimension mapping space to form a dynamically evolving morphological cluster. An abnormal path unit is used to define state transition paths in the mapping space of the compressed dimension, identify abnormal path segments that deviate from the normal evolution mode by tracking the temporal migration trajectory and cross-cluster jump behavior of the morphological clusters, and generate hierarchical monitoring events based on the topological characteristics and duration of the abnormal path segments. The topological structural unit is used to establish a bidirectional traceability association between the hierarchical monitoring events and the multi-temporal electrocardiogram morphological evolution. The bidirectional traceability association is used to selectively strengthen the local curvature weight of geodesic distance and the topological sensitive domain of abnormal path determination in the adaptive topological mapping space, thereby forming a topological structure directional shaping.

[0074] A third aspect of the present invention provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0075] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0076] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0077] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for monitoring cardiac arrhythmias based on wearable devices, characterized in that, include: Multi-lead electrocardiogram (ECG) signals of the target object are collected during a continuous monitoring period. The multi-lead ECG signals are timestamped and the signal quality is evaluated. Valid ECG signal segments that meet the quality threshold are selected. Multi-scale morphological decomposition is performed on the valid ECG signal segments to extract multi-dimensional morphological feature vectors and perform temporal correlation encoding to generate morphological feature sequences. An adaptive topological mapping space based on the morphological feature sequence is constructed. By calculating the geodesic distance and manifold embedding relationship between the morphological feature sequences, the original dimensional morphological feature sequence is projected onto the compressed dimensional mapping space to form a dynamically evolving morphological cluster. State transition paths are defined in the mapping space of the compressed dimension. By tracking the temporal migration trajectory and cross-cluster jump behavior of the morphological clusters, abnormal path segments that deviate from the normal evolutionary pattern are identified. Hierarchical monitoring events are generated based on the topological characteristics and duration of the abnormal path segments. A bidirectional tracing association is established between the graded monitoring events and the multi-temporal electrocardiogram morphological evolution. The bidirectional tracing association is used to selectively enhance the local curvature weight of geodesic distance and the topological sensitive domain of abnormal path determination in the adaptive topological mapping space, thereby forming a topological structure directional shaping.

2. The method according to claim 1, characterized in that, The multi-lead ECG signals are timestamped and their signal quality is assessed. Valid ECG signal segments that meet the quality threshold are selected. Multi-scale morphological decomposition is performed on these valid ECG signal segments to extract multi-dimensional morphological feature vectors and perform temporal correlation encoding, generating a morphological feature sequence including: Cross-lead benchmark alignment is performed on the acquisition time of each lead in the multi-lead electrocardiogram signal to construct a global time synchronization index, and the instantaneous phase correspondence between each lead is established based on the global time synchronization index; The instantaneous phase correspondence is used to calculate the morphological consistency deviation and translead waveform propagation delay distribution of each lead signal at the synchronization moment. The morphological consistency deviation and translead waveform propagation delay distribution are used as topological evaluation indicators of signal quality, and effective ECG signal segments that meet the quality threshold of the topological evaluation indicators are screened out. The effective ECG signal segment is morphologically decomposed under multiple scale windows. The nonlinear coupling features of waveform curvature tensor and interwave period are extracted in each scale window, and the manifold distance of the above features between adjacent scale windows is calculated. The curvature tensor, nonlinear coupling features and manifold distance between scale windows are combined to form a multidimensional morphological feature vector. The multidimensional morphological feature vectors are arranged in a time series. By calculating the co-evolution trajectory of the continuous gradient of the inter-scale manifold distance and the curvature tensor between adjacent multidimensional morphological feature vectors, a temporal association identifier encoding the topological dependency relationship between each multidimensional morphological feature vector and its predecessor vector is generated. The temporal association identifier is embedded into the multidimensional morphological feature vector to form a morphological feature sequence.

3. The method according to claim 2, characterized in that, The process of aligning the acquisition times of each lead in a multi-lead ECG signal across leads to a global time synchronization index, and establishing the instantaneous phase correspondence between leads based on the global time synchronization index, includes: Identify the characteristic peak times of each lead in a multi-lead electrocardiogram signal, align the characteristic peak times with the same physiological rhythm in each lead signal on the time axis, and determine the time offset between each lead signal; Based on the time offset, the acquisition time of each lead signal is compensated and corrected so that the signal characteristics of each lead signal corresponding to the same physiological event are mapped to a unified global time on the time axis, and a unique time identifier is assigned to each global time to establish a global time synchronization index covering all leads. The instantaneous signal phase of each lead signal at each global moment is extracted using the global time synchronization index. By calculating the distribution of the difference in instantaneous signal phase between different leads and the evolution trajectory of the phase difference over time, an instantaneous phase correspondence describing the instantaneous phase difference between each lead and its time dependence is constructed.

4. The method according to claim 1, characterized in that, Constructing an adaptive topological mapping space based on the morphological feature sequences, and by calculating the geodesic distance and manifold embedding relationship between the morphological feature sequences, projecting the original-dimensional morphological feature sequences onto the compressed-dimensional mapping space to form dynamically evolving morphological clusters includes: In the original dimensional space, a set of candidate geodesic paths connecting each pair of morphological feature vectors is constructed. By calculating the coupled functional of the Riemann curvature tensor integral of each candidate geodesic path along the manifold section and the covariant derivative energy of the path crossing region, the candidate geodesic path that reaches the extreme value of the coupled functional is selected as the optimal geodesic path. The arc length measure of the optimal geodesic path in the manifold metric space is defined as the geodesic distance between each pair of morphological feature vectors. Based on the geodesic distance, a geodesic neighborhood topology map of morphological feature sequence is constructed. The geodesic neighborhood topology map is subjected to Laplace-Beltramian operator eigenvalue decomposition to extract an embedded manifold coordinate system that preserves the local isometry of geodesic distance. The embedded manifold coordinate system is used as a manifold embedding relation. The original dimension morphological feature sequence is projected to a compressed dimension mapping space through a differential homeomorphism that preserves the local geodesic structure using the manifold embedding relation. In the compressed dimension of the mapping space, topological attractor cores are identified based on the geodesic density field distribution of the projected morphological feature vectors. By tracking the evolution trajectory of the tangent vector field of each topological attractor core on the time manifold and the topological invariant deformation characteristics of the attraction domain boundary, attractor cores with isomorphic tangent vector field evolution modes and topological homotopy equivalence are merged to form dynamically evolving morphological clusters.

5. The method according to claim 4, characterized in that, Based on the geodesic distance, a geodesic neighborhood topology map is constructed using a sequence of morphological features. The geodesic neighborhood topology map is then subjected to Laplace-Beltramian eigenvalue decomposition to extract an embedded manifold coordinate system that preserves the local isometry of the geodesic distance. This process includes: Using each morphological feature vector in the morphological feature sequence as a node, for each node, identify the set of neighboring nodes within the geodesic distance metric space, and calculate the kernel function decay value of the geodesic distance between each node and each neighboring node in the set of neighboring nodes as the edge weight. Based on the edge weights, connections are established between nodes, and a geodesic neighborhood topology graph is constructed. For each edge connecting two nodes in the geodesic neighborhood topology graph, the principal curvature change rate when the geodesic path connecting the two nodes traverses the local section of the manifold is calculated. The principal curvature change rate is added to the corresponding edge as a geometric tensor component, and the edge weights and the geometric tensor components are coupled to form a Laplace-Beltramian operator matrix that integrates the geodesic topology and the curvature characteristics of the manifold. The Laplace-Beltrami operator matrix is ​​decomposed to obtain the eigenvalue spectrum and the corresponding eigenfunction basis. The distribution characteristics of the interval between adjacent eigenvalues ​​in the eigenvalue spectrum are analyzed. Based on the position of the spectral gaps in the interval distribution characteristics, the dominant eigenmode subspace corresponding to the intrinsic geometry of the manifold is identified. The response components of each eigenfunction basis in the dominant eigenmode subspace in the original dimension space are extracted. The response components are combined in the sequence order of the eigenfunction basis to form an embedded manifold coordinate system.

6. The method according to claim 1, characterized in that, State transition paths are defined in the compressed dimension mapping space. By tracking the temporal migration trajectory and cross-cluster jump behavior of the morphological clusters, abnormal path segments deviating from the normal evolutionary pattern are identified. Based on the topological characteristics and duration of the abnormal path segments, hierarchical monitoring events are generated, including: In a compressed dimension mapping space, a topological adjacency network between morphological clusters is constructed. For each morphological cluster, a set of adjacent morphological clusters that meet the adjacency condition under the geodesic distance metric is identified. By calculating the geodesic barrier height and the topological connectivity index between each morphological cluster and each adjacent morphological cluster in the set of adjacent morphological clusters, the transition channel connecting morphological clusters and the composite metric of the geodesic barrier height and the topological connectivity index are defined as the state transition path. The movement trajectory of the morphological feature sequence on the state transition path is recorded as a temporal migration trajectory. In the temporal migration trajectory, the transition event that crosses the geodesic barrier height and exceeds the dynamic threshold is identified as the cross-cluster jump behavior. The inter-cluster topological homotopy class change corresponding to the cross-cluster jump behavior is extracted as the jump topological feature. Construct a state transition path manifold for the normal evolutionary pattern, calculate the Hausdorff distance between the temporal migration trajectory and the state transition path manifold, identify trajectory segments whose Hausdorff distance exceeds the deviation threshold as abnormal path segments, and extract the cumulative homotopy invariant of the jump topological features and the duration of the abnormal path segments. The cumulative homotopy invariant and the duration are mapped to the monitoring level to generate a graded monitoring event.

7. The method according to claim 1, characterized in that, Establishing a bidirectional tracing correlation between the graded monitoring events and the multi-temporal electrocardiogram morphological evolution, and using the bidirectional tracing correlation to selectively enhance the local curvature weights of geodesic distances and the topological sensitive regions for abnormal path determination in the adaptive topological mapping space, forming a topological structure directional shaping includes: For each graded monitoring event, the morphological feature sequence evolution chain corresponding to the abnormal path segment that triggered the graded monitoring event is identified by tracing back in the time dimension, and the morphological feature sequence recovery chain after the occurrence of the graded monitoring event is recorded by tracing back in the time dimension. By calculating the topological trajectory homotopy mapping relationship between the morphological feature sequence evolution chain and the morphological feature sequence recovery chain in the mapping space of the compressed dimension, a bidirectional tracing association is established. Based on the bidirectional tracing association, the set of state transition paths traversed by the morphological feature sequence evolution chain is extracted. The topological recovery frequency of each state transition path in the morphological feature sequence recovery chain is calculated. The local curvature weights involved in the geodesic distance calculation of each state transition path are gradient enhanced according to the topological recovery frequency. The topological region covered by the abnormal path segment in the compressed dimension mapping space is extracted as the topological sensitive region. The shrinking tensor field of the boundary of the topological sensitive region is calculated using the homotopy mapping relationship of the topological trajectory. The boundary range of the topological sensitive region and the deviation threshold of the abnormal path judgment are coordinated and adjusted according to the shrinking tensor field to form the topological structure orientation shaping.

8. A wearable device-based arrhythmia monitoring system for implementing the method as described in any one of claims 1-7, characterized in that, include: The signal acquisition unit is used to acquire multi-lead electrocardiogram (ECG) signals of the target object during a continuous monitoring period, perform timestamp calibration and signal quality assessment on the multi-lead ECG signals, screen out effective ECG signal segments that meet the quality threshold, perform multi-scale morphological decomposition on the effective ECG signal segments, extract multi-dimensional morphological feature vectors and perform time-series correlation encoding to generate morphological feature sequences. The topology mapping unit is used to construct an adaptive topology mapping space based on the morphological feature sequence. By calculating the geodesic distance and manifold embedding relationship between the morphological feature sequences, the original dimension morphological feature sequence is projected onto the compressed dimension mapping space to form a dynamically evolving morphological cluster. An abnormal path unit is used to define state transition paths in the mapping space of the compressed dimension, identify abnormal path segments that deviate from the normal evolution mode by tracking the temporal migration trajectory and cross-cluster jump behavior of the morphological clusters, and generate hierarchical monitoring events based on the topological characteristics and duration of the abnormal path segments. The topological structural unit is used to establish a bidirectional traceability association between the hierarchical monitoring events and the multi-temporal electrocardiogram morphological evolution. The bidirectional traceability association is used to selectively strengthen the local curvature weight of geodesic distance and the topological sensitive domain of abnormal path determination in the adaptive topological mapping space, thereby forming a topological structure directional shaping.

9. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 7.

Images