A multi-view input tensor construction method and device for dynamic electrocardio analysis

Abstract

A multi-view input construction method for dynamic electrocardio analysis, obtains original electrocardio signals of a target object in a dynamic monitoring scene, and obtains auxiliary information corresponding to the original electrocardio signals; the original electrocardio signals are preprocessed to obtain denoised electrocardio signals, residual signals and preprocessing confidence; based on the original electrocardio signals, the denoised electrocardio signals, the residual signals, the auxiliary information and the preprocessing confidence, a multi-view input tensor corresponding to a time position is constructed; the multi-view input tensor is input into a dynamic electrocardio analysis model to perform at least one of electrocardio event detection, rhythm recognition and abnormal classification, which can take into account artifact recognition, shape fidelity and long-time robustness.

Description

Technical Field

[0001] This invention relates to the field of intelligent analysis of dynamic electrocardiogram (ECG) signals, and more specifically, to a method, apparatus, electronic device, and storage medium for constructing multi-view input tensors for dynamic ECG analysis. Background Technology

[0002] Currently, most intelligent analysis methods for dynamic electrocardiogram (ECG) signals employ one of two input methods: first, directly using the raw ECG signal as the model input; second, first filtering, denoising, or suppressing artifacts on the raw ECG signal, and then using the cleaned ECG signal as the model input. While these methods can accomplish routine ECG recognition tasks to some extent, they still have significant technical limitations in complex scenarios such as out-of-hospital dynamic monitoring, wearable continuous monitoring, and motion state monitoring.

[0003] In existing technologies, a common approach is to directly input the raw, acquired dynamic ECG signals into convolutional networks, recurrent neural networks, Transformer networks, or other time-series analysis models, hoping that the model can learn effective features and complete rhythm recognition or anomaly detection. However, in dynamic monitoring scenarios, raw ECG signals often contain a variety of non-ideal components: motion artifacts, such as low-frequency disturbances or transient spikes caused by limb movements, trunk vibrations, and wearer displacement; unstable lead contact, such as waveform abrupt changes caused by electrode loosening or contact impedance fluctuations; environmental interference and power frequency noise; electromyographic interference, high-frequency spikes, and baseline drift; and significant quality inhomogeneity across different time periods, meaning that the same long time series contains both high-quality ECG segments and severely contaminated segments.

[0004] In this case, if only a single raw signal is used as input, the following problems exist: 1. The model has difficulty explicitly distinguishing between "physiological abnormalities" and "acquisition artifacts". Existing models typically learn patterns passively from waveform surface morphology, without being able to definitively determine whether a particular abnormal deformation originates from a genuine arrhythmia or from motion artifacts, lead contact changes, or noise contamination. Therefore, models are prone to misidentifying artifacts as target events such as atrial fibrillation, premature ventricular contractions, or conduction abnormalities, leading to a higher false alarm rate.

[0005] 2. The model assumes that the inputs at each time slice are equally reliable, lacking quality perception capabilities. A single raw input scheme typically does not explicitly distinguish the quality differences between different time periods, causing the model to assume that all observed data have the same reliability during training and inference. However, dynamic ECG data actually has obvious time-varying quality characteristics, and low-quality segments often contain strong artifacts. If these are used in feature extraction and state propagation without differentiation, they can easily pollute the context representation and affect the judgment results of subsequent heartbeats.

[0006] 3. Raw input significantly increases the learning burden on the model. Under single raw input conditions, the model not only needs to learn the morphological and rhythmic features of the electrocardiogram itself, but also needs to learn how to resist body movement, contact fluctuations, and complex noise. This requires the model parameters to simultaneously undertake the tasks of "signal denoising" and "clinical discrimination," resulting in high training difficulty, strong sample dependence, and unstable generalization ability.

[0007] 4. In long-term dynamic monitoring scenarios, erroneous memories are prone to accumulate. For dynamic electrocardiogram sequences lasting several minutes, hours, or even days, if time-series modeling is performed solely based on the original input, some short-term strong artifacts may be mistakenly included in the long-term state representation, thereby affecting the stability of anomaly detection over a longer time range.

[0008] Therefore, although the single raw input method retains all the original observation information, it has obvious limitations in complex dynamic scenes due to the lack of explicit expression of signal quality, artifact sources and preprocessing reliability.

[0009] Another common existing technical solution is to first clean the original ECG signal using filters, adaptive filtering, blind source separation, noise reduction networks, or artifact suppression algorithms, and then use the cleaned signal as the sole input to the subsequent model.

[0010] Compared to a single raw input, this approach can reduce noise interference with the model to some extent. However, the inventors discovered that using only a single clean input also has significant technical drawbacks.

[0011] 1. The cleaning process itself contains errors, but existing solutions typically do not explicitly express these errors. Any preprocessing algorithm inevitably has limitations in its applicability. In real-world dynamic scenarios, especially when there is intense body movement, unstable lead states, or noise overlapping with the target ECG frequency band, the preprocessing module may exhibit the following issues: insufficient artifact suppression, resulting in residual noise remaining in the clean signal; excessive suppression of real ECG components, leading to the accidental deletion of effective physiological information; and unstable performance at different time points, resulting in inconsistent output quality. However, in a single clean input solution, the downstream model can only see the "processed result" but cannot determine whether the result is reliable, whether the processing was excessive, or whether artifacts remain. In other words, the errors of the preprocessing module are hidden rather than explicitly passed to subsequent models.

[0012] 2. Noise removal may result in the loss of crucial clinical details. In Holter ECG analysis, some key clinical features are inherently small in amplitude and morphology, or locally overlap with noise bands. Examples include: P wave details; minor ST segment shifts; T wave morphological changes; fine-grained waveform clues of atrial flutter / fibrillation; and subtle features near abnormal pacing pulses. If the preprocessing module employs a strong cleaning strategy to achieve a higher signal-to-noise ratio, it may inadvertently delete or weaken these valuable components while removing noise. A single clean input scheme cannot allow the subsequent model to determine whether a component is absent or removed by the preprocessing module, potentially leading to missed detections or misclassifications.

[0013] 3. Downstream models over-rely on the cleaning results. Because the clean signal is smoother and more regular on the surface, downstream models tend to assume it is a "better input," thus weakening the verification of the authenticity of the original observations. Once there is a systematic bias in the preprocessing results, such as incorrectly retaining a certain type of artifact or incorrectly suppressing a certain type of weak waveform, this bias will be directly passed on to subsequent classification or detection results.

[0014] 4. Existing clean-only solutions fail to reflect "what is being removed." The essence of denoising is not simply generating a smoother waveform, but rather selecting and discarding different components from the original signal. However, under a single clean input framework, the suppressed information is completely lost, and the model has no way of determining whether these suppressed components are pure noise or contain some valid physiological information. This limits the ability of subsequent models to correct errors, compensate for, or identify risks in the preprocessing process. Therefore, while a single clean input reduces surface noise, it also introduces the problem of "preprocessing black box," making the model unable to perceive the quality of preprocessing and its potential side effects.

[0015] Besides the problems inherent in both single raw input and single clean input mentioned above, a deeper deficiency in existing technologies lies in the fact that most solutions do not explicitly model residual, SQI (Signal Quality Index), and confidence. Specifically: residual represents the difference between the original signal and the cleaned signal, or the residual characteristics suppressed by the preprocessing module; SQI represents the signal quality level of a given time slice or segment; and confidence represents the reliability, stability, or fidelity of the clean signal output by the preprocessing module at the current time position. The lack of explicit modeling of this information in existing technologies leads to the following drawbacks.

[0016] 1. The lack of explicit residual modeling prevents the model from understanding the "cleaned content." Without residual input, the model cannot directly access the differences between the raw and cleaned signals. Therefore, the model cannot determine: whether preprocessing primarily removes motion artifacts; whether the residual still retains valid physiological components such as QRS complexes, P waves, and ST-T signals; or whether the current clean signal carries a risk of over-cleaning. This makes subsequent models lack verifiability of preprocessing results and hinders the establishment of a consistent understanding between "raw," "clean," and "residual."

[0017] 2. Without explicit SQI modeling, the model cannot perceive the time-varying quality of the signal, and the quality of dynamic ECG sequences is not constant throughout. Within the same monitoring record, the quality may be high during the resting phase and drop sharply during the active phase. However, without explicit SQI input, the model cannot structurally know "which time slices deserve higher trust and which should be downweighted." The results are: low-quality segments are treated with equal weight to high-quality segments; artifact clippings contaminate temporal state updates and attention propagation; and long-term models are prone to forming erroneous contexts during low-quality phases.

[0018] 3. The lack of explicit confidence modeling prevents the model from perceiving the reliability of the preprocessing results themselves. Even with a clean view introduced, without further confidence, the model cannot distinguish between two scenarios: the current clean signal is a reliable result obtained from high-confidence preprocessing; and the current clean signal is a low-confidence result barely generated under high-interference conditions. In other words, existing techniques often treat the clean view as a deterministic input, ignoring the underlying sources of uncertainty. This leads to a strong dependence on unreliable clean outputs, thus reducing overall robustness.

[0019] 4. The lack of a joint expression of residual, SQI, and confidence prevents the formation of a complete input semantic loop. Having only one of raw or clean can only express either "observation" or "processing result." Only by explicitly modeling residual, SQI, and confidence simultaneously can a complete technical semantic loop be established, as follows: raw: What are the original observations currently collected? clean: What are the candidate effective components retained after preprocessing? residual: What components were removed or suppressed? SQI: What is the overall quality of the current original observations? confidence: Is the current preprocessing output reliable?

[0020] For example, Chinese patent document CN119886268A discloses a method for generating electrocardiogram (ECG) signals. This method acquires multiple training samples and corresponding training labels for each sample. For each training sample, it inputs the sample into an initial model, outputting a simulated ECG signal. The training loss of the initial model is determined based on each training sample, and the model parameters are adjusted based on this loss until a preset training termination condition is met, resulting in a target ECG generation model. By introducing R-wave information into the model training, the simulated ECG signals generated by the trained model do not contain patients' personal privacy information. Furthermore, because the target ECG generation model incorporates R-wave information when generating simulated ECG signals, the generated simulated ECG signals have personalized characteristics, ensuring high data quality while better reflecting actual clinical patterns. While this method effectively guarantees high data quality, it struggles to guarantee artifact detection and long-term robustness.

[0021] The lack of the aforementioned closed-loop expression in existing technologies means that subsequent dynamic electrocardiogram models can only make inferences under the premise of incomplete information, making it difficult to simultaneously take into account artifact identification, morphological fidelity and long-term robustness. Summary of the Invention

[0022] In view of the technical problems existing in the prior art, the present invention aims to provide a multi-view input tensor construction technology solution for dynamic electrocardiogram analysis that can take into account artifact identification, morphological fidelity and long-term robustness.

[0023] Specifically, according to one aspect of the present invention, a method for constructing multi-view input tensors for dynamic electrocardiogram (ECG) analysis is provided, which acquires the original ECG time-series signal of a target subject in a dynamic monitoring scenario, and obtains a denoised ECG signal, residual signal, signal quality index, motion state index, and preprocessing confidence level from the original ECG time-series signal after preprocessing; characterized in that it includes: The original ECG time series signal, denoised ECG signal, residual signal, signal quality index, motion state index and preprocessing confidence are time aligned and dimension mapped to construct a joint input tensor corresponding to each time step; The joint input tensor is input into the dynamic electrocardiogram analysis model so that the dynamic electrocardiogram analysis model can perform electrocardiogram event detection, rhythm recognition or abnormal classification based on multi-view information.

[0024] Preferably, according to the method, a time synchronization and resampling mechanism is adopted: each view signal is aligned based on a unified sampling time base; interpolation, mask filling, and asynchronous sensor synchronization compensation can be added.

[0025] Preferably, according to the method, the source of each component is limited: For the original ECG, To suppress artifacts in the output, for Or it can be output by the residual decomposition module. for , for Derivative motion state, Let the confidence level of the preprocessing module be . Then it satisfies the formula:

[0026] Preferably, according to the method, the dimension mapping and encoding method is limited as follows: each component is mapped to a unified feature space through linear projection, convolutional projection or embedding and then spliced ​​or weighted fused.

[0027] Preferably, according to the method, the multi-view missing fault tolerance is limited to: when any view is missing, tensor construction is completed by using a missing marker, a confidence mask, or an alternative estimate.

[0028] According to the present invention, the original electrocardiogram signal, denoised signal, residual signal, signal quality index, motion state index and preprocessing confidence obtained by dynamic acquisition are synchronously organized to form an input tensor for subsequent dynamic electrocardiogram event recognition.

[0029] According to the present invention, by using technical data with clear physiological / physical meaning as input objects, and by constructing results that directly affect the signal interpretation, artifact recognition, and event discrimination processes of subsequent electrocardiogram detection models, rather than abstract data rearrangement detached from the context, the perfect mutual support between algorithms and technical features is achieved, enabling them to jointly solve technical problems.

[0030] In contrast, existing technologies, such as traditional end-to-end CNNs / RNNs, often treat dynamic heart-view television as a single waveform input, failing to explicitly distinguish between "real physiological morphological changes" and "motion artifacts / preprocessing uncertainties," leading to erroneous learning. This invention, however, encodes the original, cleaned, residual, quality, motion, and confidence scores into a multi-view input at the same moment, enabling the model to achieve stronger artifact detection capabilities and more stable dynamic scene generalization.

[0031] Compared with the prior art, the present invention has at least the following beneficial effects: By jointly introducing three types of signal views—raw, cleaned, and residual—the model can explicitly perceive which components are preserved and which are suppressed, thereby more accurately distinguishing between real ECG abnormalities and motion artifacts and improving the discriminability between noise and physiological signals.

[0032] By introducing SQI, motion states, and preprocessed confidence, the model no longer assumes that all time-slice observations are equally reliable, thereby improving its robustness in dynamic scenarios and enhancing its ability to perceive the confidence of inputs.

[0033] Multi-view tensors provide a more stable context for long-term modeling, reducing the risk of erroneous learning caused by single-view input, thereby enhancing the generalization ability of long-term dynamic monitoring scenarios. They also provide a unified input interface for subsequent quality-aware attention, gating state updates, residual proof by contradiction, and self-supervised pre-training. This input construction method has platform-like significance and can serve as the underlying standard input for multiple subsequent dynamic ECG intelligent analysis modules. Attached Figure Description

[0034] Figure 1 is an overall flowchart illustrating a method for constructing multi-view input tensors for dynamic electrocardiogram analysis according to a specific embodiment of the present invention; (Original ECG / IMU / SQI → Preprocessing → Clean / Residual / Confidence → Time Alignment → Multi-view Tensor → Downstream Model) Figure 2 is a schematic diagram of the multi-view tensor structure illustrating a multi-view input tensor construction method for dynamic electrocardiogram analysis according to a specific embodiment of the present invention; (showing) ) Figure 3 is a flowchart illustrating the time synchronization and mapping process of a multi-view input tensor construction method for dynamic electrocardiogram analysis according to a specific embodiment of the present invention; (how data with different sampling rates are aligned to a unified time base). Figure 4 illustrates an application scenario of multi-view input tensor construction for dynamic electrocardiogram analysis according to a specific embodiment of the present invention. (Data flow in dynamic patch / sports wearing / daily monitoring) Detailed Implementation

[0035] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the following embodiments are for illustrative purposes only and are not intended to limit the scope of protection of the present invention.

[0036] Figure 1 The diagram illustrates the overall flowchart of a multi-view input construction method for dynamic electrocardiogram (ECG) analysis according to Embodiment 1 of the present invention. This multi-view input construction method is particularly suitable for scenarios such as outpatient dynamic ECG monitoring, Holter long-term monitoring, wearable ECG patches, and exercise ECG monitoring.

[0037] In dynamic ECG scenarios, the raw ECG signal is typically affected simultaneously by body movement, lead contact changes, baseline drift, electromyography interference, and environmental noise. Traditional methods usually input a single raw waveform or a single filtered waveform directly into the recognition model. This approach has the following problems: the model struggles to distinguish between "true abnormal waveforms" and "abnormal morphologies caused by motion artifacts"; the model cannot perceive the reliability of the preprocessing module over a given time period; whether the information suppressed during denoising contains effective physiological components cannot be explicitly expressed at the input side; and for long-term dynamic monitoring scenarios, the model is prone to incorrectly learning noise patterns. To address these issues, this invention proposes jointly organizing the raw ECG signal, denoised ECG signal, residual signal, quality information, motion information, and preprocessing confidence into a time-aligned multi-view input tensor for use by the subsequent dynamic ECG analysis model. This allows the subsequent model to see not just "a single waveform," but a multi-perspective technical representation of the signal state at that moment.

[0038] This invention can be deployed in wearable devices, edge analytics terminals, host computer servers, or cloud analytics platforms. The system may include: a signal acquisition unit for acquiring raw ECG waveforms; an auxiliary sensing unit for acquiring information such as IMU, accelerometer, gyroscope, and lead contact status; a preprocessing unit for generating cleaned signals, residual signals, and preprocessed confidence scores; a tensor construction unit for alignment, mapping, splicing, and tensor output; and a dynamic ECG analysis unit for subsequent anomaly detection, rhythm classification, or event localization. The ECG acquired by the signal acquisition unit can be any of single-lead, dual-lead, or multi-lead. The auxiliary sensing unit may include a 3-axis or 6-axis IMU. The preprocessing unit may employ one or more methods such as digital filtering, Kalman filtering, blind source separation, adaptive denoising, and model-driven artifact removal.

[0039] The following section provides a detailed introduction to the acquisition of raw electrocardiogram signals and auxiliary information.

[0040] In a preferred embodiment, the original electrocardiogram (ECG) signal of the target object in a dynamic monitoring scenario is acquired, and auxiliary information corresponding to the original ECG signal is also acquired.

[0041] 1. Raw electrocardiogram signal The raw electrocardiogram signal is recorded as follows:

[0042] Where T represents the length of the time series. Represents the original ECG sample value or corresponding representation at time t.

[0043] 2. Auxiliary Information Auxiliary information may include one or more of the following: signal quality indicators Motion status indicators Lead contact status; electrode impedance change indication; device wearing stability information; sensor operating status information. Among these, signal quality indicators... It can be determined based on either of the following methods: waveform stationarity; Peak detectability; short-time signal-to-noise ratio; baseline oscillation; high-frequency noise intensity; period consistency; quality score of model estimation.

[0044] Motion status indicators Available from Raw data or its derived features, such as: acceleration magnitude; angular velocity magnitude; short-term motion intensity; attitude switching frequency; impact event markers.

[0045] The following section provides a detailed introduction to the preprocessing module's outputs: clean, residual, and confidence.

[0046] In a preferred embodiment, the raw ECG signal is preprocessed to output a denoised ECG signal, a residual signal, and a preprocessing confidence level.

[0047] 1. Noise-reduced ECG signal The denoised ECG signal is denoted as:

[0048] It can be generated by filters, artifact suppression networks, adaptive filters, model-driven reconstruction networks, etc.

[0049] 2. Residual signal The residual signal is denoted as: The residual signal is not limited to time-domain subtraction; it can also be the complement of a frequency-domain mask or the reconstructed difference vector in a variational autoencoder.

[0050] In a preferred embodiment:

[0051] In other embodiments, the residual signal may not be strictly subtracted point by point, but rather a "characterization of suppressed components" output by the preprocessing module, such as residual encoding in the feature space.

[0052] 3. Preprocessing confidence level The preprocessing confidence level is denoted as:

[0053] The preprocessing confidence score is used to characterize the reliability of the preprocessed output at time t. It can be calculated by one or more of the following indicators: whether the morphological difference before and after denoising is too large; whether the position of the R peak remains consistent; whether the frequency band energy change is abnormal; whether the signal reconstruction error is too large; whether the artifact estimate is too strong; and the uncertainty of the preprocessing network's own output.

[0054] For example, in one embodiment, the preprocessing confidence level can be based on the reconstruction error. To confirm:

[0055] Where β is the scaling factor.

[0056] The following section provides a detailed explanation of time synchronization and the construction of multi-view input tensors.

[0057] After obtaining the above information, it is time-aligned and uniformly encoded to construct a multi-view input tensor.

[0058] 1. Time synchronization Since information such as ECG, IMU, SQI, and confidence levels may have different sampling rates, time synchronization is performed first. Synchronization methods include, but are not limited to: direct alignment based on a unified clock; resampling; interpolation; window aggregation; nearest neighbor mapping; and mask padding.

[0059] For example, if the ECG sampling rate is 250 Hz and the IMU sampling rate is 100 Hz, the IMU features can be mapped to the ECG time base to obtain the motion state index corresponding to each ECG time position.

[0060] 2. Unified Feature Mapping In one embodiment, the view information can first be mapped to a unified feature space, for example:

[0061]

[0062] in, These can be linear transformations, convolutional transformations, embedding maps, or normalization maps, respectively.

[0063] 3. Construction of multi-view tensors like Figure 2 As shown, the joint input tensor for each time position is constructed:

[0064] Furthermore, a complete sequence input can be formed:

[0065] If the embedding and splicing method is used, it can also be represented as:

[0066] The tensor construction method is not limited to simple concatenation; it can also be: Channel stacking; group stacking; weighted fusion; gated fusion; multi-branch input.

[0067] The multi-view input tensor constructed by this invention has at least the following technical advantages: 1. Original signal view

[0068] All sampling information is retained so that subsequent models can directly access real observations that have not been preprocessed or modified.

[0069] 2. Denoising signal view

[0070] Provide candidate physiological signals with higher signal-to-noise ratios so that the model can more easily learn rhythm and morphological features.

[0071] 3. Residual signal view Explicitly exposing the "cleaned-out components" to the model helps it determine whether the preprocessing has mistakenly deleted effective ECG components or whether strong motion artifacts exist.

[0072] 4. Quality Indicators

[0073] It is used to characterize whether the signal is reliable at the current moment, so that the model can learn to "reduce dependence when the quality is low and enhance utilization when the quality is high".

[0074] 5. Motion status indicators It is used to reveal the relationship between bodily movement events and abnormal waveforms, and to help the model distinguish between cardiac abnormalities and motion-induced pseudo-abnormalities.

[0075] 6. Preprocessing confidence level

[0076] This is used to indicate whether the preprocessed output is reliable, enabling the model to perceive the reliability boundaries of the clean view. Through this gradient input layout, the model can maintain basic artifact recognition capabilities based on the core three views (Raw, Clean, Resid) even if auxiliary sensors (such as IMU) fail.

[0077] Therefore, the present invention is essentially about constructing a multi-view technical input with signal physical semantics and processing semantics, rather than ordinary feature splicing.

[0078] The following reference Figure 3 , Figure 4 The following section explains the input-based dynamic electrocardiogram analysis model and its application scenarios.

[0079] like Figure 3 As shown, the constructed multi-view input tensor can be used to input any dynamic electrocardiogram analysis model.

[0080] Example 1: Input to Transformer Model The sequence tensor X is input into the Transformer encoder, which outputs a temporal representation, which is then used for anomaly classification.

[0081] Example 2: Input to a CNN + Transformer hybrid model First, local waveform features are extracted by convolutional layers, and then rhythmic relationships are modeled by Transformer.

[0082] Example 3: Input to state-space model or Mamba model By inputting tensors into a state-space network, we can model trend changes and event dependencies in long-term monitoring data.

[0083] The steps for inputting into a multi-branch fusion model are explained in detail below.

[0084] Different views are sent to different sub-networks, and then joint decisions are made through the gating fusion module.

[0085] Downstream tasks may include: atrial fibrillation detection; premature ventricular contractions (PVCs) / premature atrial contractions (PACs) detection; conduction block identification; ST-T anomaly identification; rhythm typing; long-term event localization; and dynamic risk scoring.

[0086] In real-world wearable scenarios, some auxiliary information may be missing or sampled abnormally. For example: The IMU is temporarily unavailable; SQI is not generated in real time; preprocessing fails during certain periods; ECG and IMU are not fully synchronized. To address these issues, the present invention may also include robust processing mechanisms, such as: 1. Missing tagging mechanism For missing views, set the corresponding missing flag.

[0087] 2. Confidence masking mechanism The mask tensor indicates which components are available and which are not.

[0088] 3. Placeholder Filling Mechanism Use zero, mean, most recent value, or model estimate to complete the placeholder.

[0089] For example, in one embodiment, the final input can be written as:

[0090] in, This represents the availability indicator vector for each view at that moment.

[0091] like Figure 4 As shown below, an application example of dynamic patch ECG is given.

[0092] Scene setting ECG single-lead sampling rate: 250 Hz; IMU sampling rate: 100 Hz; The monitored individuals are in their daily activity state; Objective: Detection of atrial fibrillation and premature ventricular contractions.

[0093] step Acquire the raw ECG waveform; Calculate short-time signal quality score ; Motion intensity calculated by IMU ; Denoising was performed on the ECG to obtain ; Calculate the residual signal ; Based on the stability calculation of reconstruction before and after denoising ; Map all components to the ECG time base. Build:

[0094] will sequence Input the downstream dynamic electrocardiogram model; Output the exception type and its corresponding time interval.

[0095] In this example, when the user is exercising vigorously, Increase, and The model can use multiple views to jointly determine that the abnormal waveform is more likely caused by motion artifacts than by a true arrhythmia; while when and The difference is too large, and at the same time When significant QRS components are still retained, the model can further identify potential risks of over-cleaning during preprocessing. By employing a unified feature space mapping, redundant computation across multiple networks is avoided, resulting in a 10% reduction in overall inference latency compared to traditional asynchronous multimodal fusion.

[0096] In this invention, an electronic device is also provided, comprising: At least one processor; Memory; Computer programs stored in the memory.

[0097] When the processor executes the computer program, it implements the construction method in any of the above embodiments.

[0098] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, it implements the construction method in any of the above embodiments.

[0099] The present invention has been described in detail above through specific embodiments. However, this description is exemplary, and those skilled in the art can make various modifications and changes to it. As long as they do not depart from the spirit and purpose of the present invention, all such modifications and changes should fall within the protection scope of the present invention, which is defined by the appended claims.

Claims

1. A method for constructing multi-view input for dynamic electrocardiogram analysis, characterized in that, include: Acquire the raw electrocardiogram (ECG) signal of the target object in a dynamic monitoring scenario; The original electrocardiogram (ECG) signal is preprocessed to obtain a denoised ECG signal and a residual signal characterizing the difference between the original ECG signal and the denoised ECG signal; Based on the original ECG signal, the denoised ECG signal, and the residual signal, a multi-view input tensor corresponding to the time position is constructed. The multi-view input tensor is input into the dynamic electrocardiogram analysis model to perform electrocardiogram event detection, rhythm recognition, or abnormality classification.

2. The method according to claim 1, characterized in that, The construction of the multi-view input tensor also includes: Acquire auxiliary information that is synchronously acquired with the original electrocardiogram signal, the auxiliary information including at least one of signal quality indicators and motion state indicators; The auxiliary information is dimensionally fused with the original ECG signal, the denoised ECG signal, and the residual signal to construct the multi-view input tensor.

3. The method according to claim 2, characterized in that, Also includes: Obtain the preprocessing confidence level, which characterizes the reliability of the preprocessing output; The preprocessed confidence level is used as an independent view dimension and connected to the multi-view input tensor.

4. The method according to claim 3, characterized in that, The multi-view input tensor satisfies the following form: ， in, This represents the original electrocardiogram signal. Characterization of denoised electrocardiogram signals To characterize the residual signal, As a signal quality indicator, As an indicator of motion status, For preprocessing confidence levels.

5. The method according to any one of claims 1-4, characterized in that, The residual signal The methods of obtaining it include: Calculate the point-by-point difference in the time domain between the original ECG signal and the denoised ECG signal; or, Obtain the residual representation of the suppressed component output by the preprocessing module in the feature space.

6. The method according to any one of claims 1-4, characterized in that, The construction of the multi-view input tensor corresponding to the time position includes: Multi-source components are time-synchronized and resampled to align to a unified sampling time base; The aligned components are mapped to a unified feature space through linear projection, convolutional projection, or embedding, and then spliced ​​or stacked according to a preset channel order.

7. The method according to any one of claims 1-4, characterized in that, Also includes: In the event of missing views or auxiliary information, a robust input representation is generated based on missing markers, confidence masks, or interpolation compensation.

8. The method according to claim 6, characterized in that, The multi-view input tensor satisfies the following form: ， in, This represents the raw electrocardiogram signal characterization at time t. This represents the denoised ECG signal characterization at time t. This represents the residual signal characterization at time t. The signal quality index at time t. The motion state index at time t. This represents the preprocessing confidence level at time t.

9. The method according to claim 6, characterized in that, The unified feature space is obtained through at least one of the following methods: linear projection, convolutional projection, embedding mapping, normalization mapping, and dimensionality scaling transformation.

10. The method according to claim 1, characterized in that, The dynamic monitoring scenario includes at least one of the following: Dynamic ECG patch monitoring scenarios, Holter long-term monitoring scenarios, exercise ECG monitoring scenarios, wearable continuous ECG monitoring scenarios, and outpatient ECG monitoring scenarios during daily activities.

11. The method according to claim 1, characterized in that, The multi-view input tensor is used to improve the ability of the dynamic electrocardiogram analysis model to distinguish at least one of the following situations: Changes in actual physiological waveforms, motion artifacts, poor lead contact, preprocessing errors, residual noise, and information loss due to excessive cleaning.

12. The method according to claim 1, characterized in that, Before or after constructing the multi-view input tensor, the following is also included: In cases of missing views, missing auxiliary information, or asynchronous sampling, robust input representations are generated based on missing markers, confidence masks, interpolation compensation, or default placeholders.

13. The method according to claim 1, characterized in that, The dynamic electrocardiogram analysis model includes at least one of the following network structures: Convolutional neural networks, Transformer networks, state-space model networks, recurrent neural networks, and time series analysis networks based on multi-branch fusion.

14. A multi-view input construction device for dynamic electrocardiogram analysis, characterized in that, include: The acquisition module is configured to acquire the raw electrocardiogram signals and auxiliary information of the target object in a dynamic monitoring scenario. The preprocessing module is configured to preprocess the original electrocardiogram signal to obtain a denoised electrocardiogram signal, a residual signal, and a preprocessing confidence level. The tensor construction module is configured to construct a multi-view input tensor corresponding to the time position based on the original ECG signal, the denoised ECG signal, the residual signal, the auxiliary information, and the preprocessing confidence. The analysis module is configured to input the multi-view input tensor into the dynamic electrocardiogram analysis model to perform at least one of electrocardiogram event detection, rhythm recognition, and abnormality classification.

15. An electronic device, characterized in that, include: At least one processor; and a memory communicatively connected to the at least one processor; The memory stores a computer program, which, when executed by the at least one processor, causes the electronic device to perform the method according to any one of claims 1 to 12.

16. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method according to any one of claims 1 to 12.

Images