Cross-view-angle electrocardiogram myocardial infarction diagnosis method, device and system and storage medium

Abstract

The invention discloses a cross-view electrocardiogram myocardial infarction diagnosis method, device and system and a storage medium. The cross-view electrocardiogram myocardial infarction diagnosis method comprises the steps that original multi-view electrocardiogram signals are obtained; the method comprises the following steps: respectively extracting multilevel time sequence characteristics from standard twelve-lead and Frank orthogonal lead signals by using a lead specific encoder; multi-scale fusion in the visual angles, semantic alignment between the visual angles and global fusion irrelevant to the visual angles are carried out in sequence; deep transformation is carried out on the fusion features, and a plurality of supervision tasks such as myocardial infarction existence judgment and focus part classification are constructed based on the deep transformation, and joint optimization is carried out; a multi-view multi-task model is used as a teacher, a lightweight student model only receiving standard twelve-lead input is trained, efficient migration of diagnostic knowledge is achieved, and a final deployment model is obtained; all stages are gradually expanded from bottom to top, and multi-task feature modeling and intelligent diagnosis of myocardial infarction are finally completed through multi-scale feature integration and cross-view semantic fusion from electrocardiograph characterization of different lead systems.

Description

Technical Field

[0001] This invention belongs to the field of time series processing technology, specifically relating to a cross-view electrocardiogram method, device, system, and storage medium for diagnosing myocardial infarction. Background Technology

[0002] Electrocardiography (ECG) is a crucial tool for identifying cardiac dysfunction and is widely used in the detection of diseases such as myocardial infarction (MI). In recent years, with the rapid development of deep learning in the medical field, researchers have begun to explore how to build higher-performance, more easily deployable intelligent ECG diagnostic models to adapt to multi-lead configurations, different device data, and complex classification tasks. Especially in the automatic identification of MI, the model not only needs to determine the presence or absence of MI but also needs to further locate its specific location to support refined clinical treatment. Therefore, how to fully extract representational differences from lead data has become a key issue. Simultaneously, when multi-lead / multi-view data is available for training but real-world applications may only provide standard 12-lead signals, balancing diagnostic accuracy with deployment limitations becomes a significant factor affecting model deployment.

[0003] In 2024, Park et al. from Harvard University proposed an electrocardiogram (ECG) diagnostic model based on multi-task learning. This model improved its discriminative ability on standard 12-lead signals through joint classification and signal reconstruction tasks, mitigating overfitting to some extent. However, this method uses a fixed weighting approach during task optimization, lacking a dynamic adjustment mechanism for task weights, making it difficult for the model to adapt to the learning requirements of different tasks at different stages. Qiang et al. from Stanford University constructed a 12-lead classification model based on lead subset encoding, performing feature learning across different lead combinations. By enhancing the expressive power of discriminative leads, they improved the model's recognition performance on typical and atypical MI samples. Although this method optimized data utilization, the lead types used are still limited to traditional planar lead combinations, and its ability to reflect cardiac electrical activity in the spatial dimension remains insufficient. Furthermore, the above methods mainly focus on modeling the standard 12-lead system, making it difficult to fully utilize the complementary information provided by lead structures with greater spatial expressive power.

[0004] To further enhance the coverage of electrophysiological information in spatial directions, some studies have begun to focus on introducing Frank orthogonal lead isovector electrocardiogram (VCG) structures. Frank orthogonal leads provide richer perspectives on electrical signal changes through X, Y, and Z three-dimensional spatial projection, theoretically compensating for blind spots in lesion localization within traditional 12-lead systems and improving classification accuracy and regional discrimination capabilities. Existing studies have preliminarily validated the spatial supplementary value of Frank orthogonal leads in specific tasks, but a systematic modeling method is still lacking for its application in intelligent diagnosis of myocardial infarction. Current mainstream methods, when fusing multi-lead information, still primarily rely on splicing or static fusion, lacking deep modeling of spatial relationships between leads. This makes it difficult to fully extract the potential information from Frank orthogonal leads, and the constructed multi-lead models are complex and computationally expensive, making them difficult to directly apply to conventional 12-lead devices or edge deployment scenarios. Furthermore, multi-view features differ in distribution, scale, and expression. Without effective intra-view modeling and inter-view alignment mechanisms, feature redundancy or complementary information after fusion may not be fully utilized, thus limiting model performance improvement.

[0005] Against this backdrop, lightweight techniques have gradually become a research hotspot to balance model performance and practical deployment needs. In 2024, Borah et al. from the Indian Institute of Technology proposed compressing ECG classification models using a teacher-student structure. This transfers the capabilities of the teacher model trained with complete input information to at least the student model in the lead, achieving rapid diagnosis during the inference stage. While this method significantly reduces model size and computational load for specific tasks, when dealing with complex models containing multi-task structures and multi-perspective information, a fixed distillation strategy struggles to effectively transfer key features. The performance of the student model is easily affected by structural weakening, limiting the transfer effect. Especially when the teacher model simultaneously incorporates multi-perspective fusion and multi-task constraints, distilling only the final classification output may be insufficient to transfer key discriminative features; a combination of intermediate features and dynamic strategies during training is needed for more effective knowledge transfer. On the other hand, in 2025, Makhmudov et al. from the University of Zurich, Switzerland, designed a multi-task ECG structure that integrates classification and multi-parameter prediction, achieving feature sharing between concurrent tasks and demonstrating the role of multi-task modeling in improving diagnostic efficiency. However, the structure is not optimized for potential interference between tasks. The auxiliary tasks suppress the performance of the main task in some stages, and there is a lack of dynamic adjustment mechanism for the relationship between tasks.

[0006] In summary, current intelligent ECG diagnostic models face three main challenges: First, traditional lead structures struggle to fully represent the spatial electrophysiological characteristics of myocardial infarction, necessitating the introduction of structures with three-dimensional projection capabilities, such as Frank orthogonal leads, to improve diagnostic resolution. Second, while multi-task structures enhance model performance, they also introduce task conflicts and optimization instability, with current mainstream methods failing to achieve effective coordination between tasks. Third, as model structures become increasingly complex, existing distillation methods struggle to fully transfer the capabilities of multi-view, multi-task models, making it difficult for lightweight models to maintain diagnostic accuracy. Particularly under the practical condition that "multi-view information can be used during training, but only standard 12-lead input single-view information can be used during deployment," how to achieve effective transfer and reuse of cross-view knowledge is one of the key issues restricting the practical application of such methods. Summary of the Invention

[0007] To address the problems existing in the prior art, this invention provides a cross-view electrocardiogram method, device, system, and storage medium for diagnosing myocardial infarction. Under the premise of ensuring that the model is ultimately deployed in a clinical environment equipped only with a standard twelve-lead device, it can make full use of multi-view spatial information (such as Frank orthogonal leads) and multi-task supervision signals available during the training phase to achieve high-precision myocardial infarction identification and localization.

[0008] To achieve the above objectives, the present invention provides the following solution: A cross-view electrocardiogram diagnostic method for myocardial infarction includes: Step S1: Preprocess the acquired raw multi-view electrocardiogram signals; Step S2: Based on the preprocessed original multi-view ECG signals, use lead-specific encoders to extract multi-level temporal features from the standard twelve-lead and Frank orthogonal lead signals, respectively. Step S3: Based on the multi-level temporal features, perform multi-scale fusion within the viewpoint, semantic alignment between viewpoints, and global fusion independent of the viewpoint in sequence. Step S4: Perform deep transformation on the fused features, and based on this, construct multiple supervised tasks such as myocardial infarction presence / absence discrimination and lesion site classification, perform joint optimization, and obtain a multi-view multi-task model; Step S5: Using the multi-view, multi-task model as the teacher, train a lightweight student model that only receives standard twelve-lead input to achieve efficient transfer of diagnostic knowledge and obtain the final deployment model to realize cross-view electrocardiogram myocardial infarction diagnosis.

[0009] As a preferred option, in step S1, the original multi-view ECG signal is filtered, denoised, and the heartbeats are extracted and normalized to obtain standardized heartbeat samples.

[0010] As a preferred approach, the multi-view, multi-task model employs a joint loss function to weight and integrate the MI discrimination loss, part localization loss, and reconstruction loss, enabling multiple tasks to work collaboratively under the same optimization objective. ; Through weight , , The joint loss function is adjusted to guide the model to balance the optimization of different tasks in the shared feature space, so that the features simultaneously contain global discrimination, local localization and waveform structure information.

[0011] The present invention also provides a cross-view electrocardiogram myocardial infarction diagnostic device, comprising: The first processing module is used to preprocess the acquired raw multi-view electrocardiogram signals; The second processing module is used to extract multi-level temporal features from the standard twelve-lead and Frank orthogonal lead signals respectively, based on the preprocessed original multi-view electrocardiogram signals and using lead-specific encoders. The third processing module is used to perform multi-scale fusion within the viewpoint, semantic alignment between viewpoints, and global fusion independent of viewpoints in sequence according to multi-level temporal features. The fourth processing module is used to perform deep transformation on the fused features and construct multiple supervised tasks based on this, such as the determination of the presence or absence of myocardial infarction and the classification of lesion sites, and perform joint optimization to obtain a multi-view, multi-task model. The fifth processing module is used to train a lightweight student model that only receives standard twelve-lead input, using a multi-view, multi-task model as the teacher, to achieve efficient transfer of diagnostic knowledge and obtain the final deployment model to realize cross-view electrocardiogram myocardial infarction diagnosis.

[0012] As a preferred option, the first processing module filters, reduces noise, extracts and normalizes the original multi-view ECG signals to obtain standardized heartbeat samples.

[0013] As a preferred approach, the multi-view, multi-task model employs a joint loss function to weight and integrate the MI discrimination loss, part localization loss, and reconstruction loss, enabling multiple tasks to work collaboratively under the same optimization objective. ; Through weight , , The joint loss function is adjusted to guide the model to balance the optimization of different tasks in the shared feature space, so that the features simultaneously contain global discrimination, local localization and waveform structure information.

[0014] The present invention also provides a cross-view electrocardiogram myocardial infarction diagnostic system, comprising: a memory and a processor, wherein the memory stores a computer program executed by the processor, and the computer program, when executed by the processor, performs a cross-view electrocardiogram myocardial infarction diagnostic method.

[0015] The present invention also provides a storage medium storing a computer program that, when running, executes a cross-view electrocardiogram method for diagnosing myocardial infarction.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. By introducing multi-view ECG signals composed of standard 12 leads and Frank orthogonal leads, and designing a special feature fusion and alignment mechanism, we can model and fuse multi-dimensional spatial information of cardiac electrical activity, thereby overcoming the information limitations of a single planar projection perspective and improving the model's ability to express myocardial infarction features.

[0017] 2. By constructing a "multi-perspective teacher-single-perspective student" training paradigm that incorporates knowledge distillation, the rich diagnostic knowledge learned during the training phase using multiple perspectives and tasks is effectively transferred to a lightweight student model that relies solely on standard twelve-lead input. This resolves the contradiction between complex models and the need for lightweight deployment, enabling diagnostic performance approaching that of multi-perspective models to be achieved in clinical scenarios without the need for additional hardware support. Attached Figure Description

[0018] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart of the cross-view electrocardiogram method for diagnosing myocardial infarction according to an embodiment of the present invention; Figure 2 Here is a diagram of the encoder structure; Figure 3 Here is a structural diagram of the IRMB module; Figure 4 This is a schematic diagram of the in-view fusion module; Figure 5 This is a schematic diagram of the inter-view fusion module; Figure 6 This is a schematic diagram of the deep fusion feature extraction module; Figure 7 This is a schematic diagram of a multi-task output structure; Figure 8 This is a schematic diagram of the knowledge distillation structure. Detailed Implementation

[0020] 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] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0022] Example 1 like Figure 1 As shown, the present invention provides a cross-view electrocardiogram method for diagnosing myocardial infarction, comprising: Step S1, Data Preprocessing and Heartbeat Segmentation: The original multi-view ECG signals are filtered, denoised, and the heartbeats are extracted and normalized to obtain standardized heartbeat samples.

[0023] Step S2, Dual-view multi-scale feature encoding: Using a lead-specific encoder, multi-level temporal features are extracted from the standard twelve-lead and Frank orthogonal lead signals, respectively.

[0024] Step S3, Multi-view Feature Fusion: Perform multi-scale fusion within the viewpoint, semantic alignment between views, and global fusion independent of the viewpoint in sequence to generate a unified fused feature representation.

[0025] Step S4, Deep Feature Extraction and Multi-Task Modeling: Deep transformation is performed on the fused features, and based on this, multiple supervised tasks such as myocardial infarction presence / absence discrimination and lesion site classification are constructed and jointly optimized.

[0026] Step S5: Model transfer and deployment based on knowledge distillation: Using the complete multi-view, multi-task model as the teacher, train a lightweight student model that only receives standard twelve-lead input to achieve efficient transfer of diagnostic knowledge and obtain the final deployment model.

[0027] Each stage unfolds from the bottom up, starting with the electrocardiogram representations of different lead systems, and through multi-scale feature integration and cross-perspective semantic fusion, ultimately completing multi-task feature modeling and intelligent diagnosis of myocardial infarction.

[0028] As one embodiment of the present invention, step S1 includes: Step 1.1: Prepare the dataset The PTB diagnostic electrocardiogram database was used, containing 549 ECG records from 290 subjects, with durations ranging from approximately 30 seconds to 2 minutes, a sampling rate of 1000 Hz, and each record containing 15 leads (12 standard leads and 3 Frank orthogonal leads). In this invention, based on the clinical annotation information provided by the database, data from healthy subjects and myocardial infarction (MI) subjects were selected to construct a binary classification task to determine the presence or absence of MI. Simultaneously, within the MI samples, myocardial infarction was divided into six categories based on lesion location to construct a multi-class classification task for MI location, thereby simultaneously achieving the two diagnostic objectives of diagnosing the presence or absence of myocardial infarction and lesion location on a unified dataset.

[0029] In this invention, the data is divided using five-fold cross-validation across participants. Specifically, the participants in the dataset are randomly divided into five folds, each containing data from different participants. During each validation process, one fold is selected as the validation set, and the data from the remaining four folds are used for training. In this way, the data from each participant is used as part of the validation set, ensuring the stability of the results and the generalization ability of the model.

[0030] Step 1.2: Data Preprocessing First, the raw ECG signal was bandpass filtered from 0.5 to 150 Hz, and a bandstop filter centered at 50 Hz was used to suppress power frequency interference. Then, multi-layer wavelet decomposition and threshold contraction were combined to denoise the signals in each lead, along with median filtering and moving average smoothing. After filtering and denoising, the R-wave position was detected using an improved Pan-Tompkins algorithm. Using the R-wave as a reference, 0.8-second heartbeat segments were extracted at 0.3 seconds before and 0.5 seconds after the heartbeat, and each heartbeat was resampled into 800 sampling points along the time dimension. Next, the heartbeat samples were normalized lead-by-lead amplitude, and quality control was performed based on the proportion of zero values, amplitude range, and flat segments, removing abnormal heartbeats that did not meet the requirements. After preprocessing, the processed heartbeat samples were divided into 5 folds with roughly equal sample sizes using stratified sampling, and the distribution of categories in each fold was kept basically consistent. Following the five-fold cross-validation strategy, in each round of experiments, one fold (accounting for about 20% of the total samples) was selected in turn as the validation set, and the remaining folds were used for model training.

[0031] As one embodiment of the present invention, in step S2, the structure of the dual-view multi-scale encoder is as follows: Figure 2As shown, this module aims to extract discriminative temporal features from standard 12-lead signals and Frank orthogonal lead signals, providing high-quality input features for subsequent multi-view fusion modules. The encoder is based on the IRMBUnetV2 architecture, an improvement on the traditional U-Net structure, combining residual networks and attention mechanisms to enhance feature extraction capabilities.

[0032] This module independently constructs a convolutional feature extractor for each lead. First, for the input... and Each lead undergoes initial feature extraction via a convolutional block. This convolutional block contains a 7×1 convolutional layer (8 kernels), a batch normalization layer, and a ReLU activation layer. Subsequently, the output of each lead is fed into an iRMB-based encoding network for further deep feature extraction.

[0033] The encoding network employs a U-shaped structure and includes three downsampling stages. The internal structure of the iRMB block is as follows: Figure 3 As shown, this module is used to extract deeper ECG morphological features. The iRMB block consists of three main components: the channel attention module obtains a channel-level global representation by performing global average pooling on each channel of the feature map, and then generates channel weights through two fully connected layers; the spatial attention module obtains spatial features by average pooling and max pooling along the channel dimension, and then generates spatial weights through convolutional layers; the SDI module uses multi-scale dynamic convolution, and generates sample-adaptive convolutional kernel weights through global average pooling and a multilayer perceptron. The outputs of the three components are connected via residual connections to achieve feature fusion.

[0034] ; The encoding network progressively increases the number of channels and reduces the temporal dimension through three downsampling stages. Each downsampling stage uses a convolution with a stride of 2 to achieve downsampling, while residual connections ensure effective information transmission. After multi-layer processing, each lead obtains feature outputs at three different scales: shallow features... , Mid-layer features , Deep features , ,in > > This indicates the time series length at different scales.

[0035] Ultimately, the three scale features of each lead are unified to the same latent dimension through a projection layer, forming a multi-scale feature representation. , Multi-scale features from different leads are stacked into a unified multi-view feature representation.

[0036] ; in, This indicates splicing along the lead dimension.

[0037] Through the above design, the dual-view multi-scale encoder can effectively extract multi-level temporal features, providing high-quality feature input for myocardial infarction diagnosis tasks.

[0038] Furthermore, the encoder module used to construct lead-specific representations in step 2 can be replaced with other deep feature extraction structures with equivalent expressive power, while maintaining the basic idea of ​​"lead-specific modeling". For example, multi-scale convolutional units or lightweight feature extraction networks can be used to replace the existing IRMB-Unet structure.

[0039] As one embodiment of the present invention, step S3 includes: Step 3.1: Intra-view multi-scale feature fusion After completing the dual-view ECG feature encoding, this invention first constructs a multi-scale feature fusion module within each viewpoint. This module performs unified mapping, discriminative weighting, and structural information integration on the encoded features at different scales to significantly enhance the representational capability within a single viewpoint. Since ECG changes related to myocardial infarction are typically reflected in local waveforms, inter-lead correspondences, and multi-frequency waveform details, single-scale or simply stacked features are insufficient to fully characterize temporal details. The multi-scale fusion module proposed in this invention employs a structure of scale-level mapping, attention filtering, and temporal weighting. In the final stage, scale alignment and feature concatenation form a complete single-view feature representation. The overall structure of this module is as follows: Figure 4 As shown, this module aims to address the problem that single-scale features are insufficient for characterizing multi-band ECG patterns. Through a hierarchical screening mechanism, it enables the focusing on key frequency bands and key time segments, thereby improving the discriminative quality of single-view features.

[0040] In this module, the encoded features at the k-th scale for a given perspective (standard 12-lead perspective or Frank orthogonal lead perspective) are denoted as... Its dimensions are usually ,in For batch size, This represents the number of channels at this scale. The length is the time series length. To achieve effective fusion across different scales, features at each scale need to be mapped to a unified latent dimension. This invention uses a 1×1 convolution as a linear projection to unify the number of channels at different scales to a fixed latent channel dimension. (For example, =256), which allows subsequent attention weighting and scale fusion to be performed within a consistent feature dimension space.

[0041] ; After mapping to a unified channel dimension, this invention introduces a channel attention mechanism to highlight key channels within a scale. This is achieved by applying the mapped features along the time dimension. Global average pooling is performed to extract the statistical description of each channel over the entire time period, thus obtaining the channel description vector. This description vector is then processed through a two-layer fully connected network to extract nonlinear relationships, and finally, channel-level weights are generated through Sigmoid activation. These weights are used to recalibrate the importance of each channel, achieving dynamic enhancement of different frequency bands of the ECG waveform.

[0042] ; Following channel attention enhancement, this invention further introduces a spatial attention mechanism to strengthen key ECG segments in the temporal dimension. Features of myocardial infarction, such as ST-segment elevation and T-wave changes, typically exhibit highly concentrated and temporally specific localized changes; therefore, selective weighting of the temporal dimension is necessary. Spatial attention first generates two one-dimensional time series through average pooling and max pooling in the channel dimension, concatenates them along the channel direction, extracts the fused temporal features through 1×1 convolution, and finally obtains the temporal weights through Sigmoid activation.

[0043] ; Through the aforementioned channel attention and spatial attention-based stepwise filtering, this invention can significantly highlight the "high-contribution region" containing key diagnostic patterns within each scale without increasing the number of parameters. To unify the time series lengths at different scales, the outputs at each scale need to be upsampled or downsampled to achieve the same time dimension. The length-aligned outputs at each scale are spliced ​​together in the channel dimension or scale dimension to form a refined multi-scale feature representation of that viewpoint, thereby realizing information integration within a single viewpoint and providing complete basic features for subsequent feature alignment between viewpoints and multi-view fusion.

[0044] ; This hierarchical structure of "linear mapping - channel attention - spatial attention - scale alignment fusion" enables the model to comprehensively encode features from different frequency bands, internal structures of different leads, and different scales, effectively improving the ability to capture local and global ECG features within a single perspective. This provides more stable and diagnostically sensitive feature inputs for subsequent cross-perspective interaction and global fusion.

[0045] Step 3.2: Semantic alignment between perspectives (bidirectional cross-attention) To establish an effective semantic association between the standard twelve-lead (planar projection) and the Frank orthogonal leads (spatial vector), and to achieve the directional exchange of complementary information between the two, this invention further proposes a perspective semantic alignment module (corresponding to...). Figure 5 This module is used to establish bidirectional, fine-grained information interaction and feature compensation relationships between two perspectives. Electrophysiological changes related to myocardial infarction are naturally complementary in both perspectives: the standard twelve-lead module focuses on local myocardial activity in the precordial leads, while the Frank orthogonal leads provide ECG vector changes closer to three-dimensional space. This invention, through a bidirectional cross-attention mechanism, enables the two perspectives to achieve semantic association in a shared latent space, thereby obtaining a more stable and structured cross-perspective expression than a single perspective. The overall structure of this module is as follows: Figure 5 As shown.

[0046] In this module, the present invention first uses two refined perspective features from step 3.1. and By projecting linearly onto a unified latent space dimension, a sequence of queries, keys, and values ​​can be constructed. From a standard perspective, this query vector sequence is denoted as... The key-value sequence from Frank's perspective is denoted as... , Subsequently, this invention employs a scaled dot product-based multi-head cross-attention computation to calculate the dependencies between two perspectives, enabling the standard perspective to select and absorb semantically associated time segments and lead combination patterns from the Frank perspective.

[0047] ; Through the aforementioned attention calculation, the standard perspective forms an enhanced interactive feature representation, enabling the extraction of highly correlated latent structural information from the Frank perspective. Similarly, this invention constructs a dual process for the Frank perspective, using the standard perspective as the key and value, and the Frank perspective's own features as the query. Semantic compensation for "Frank's perspective absorbing the standard perspective" is obtained through reverse cross-attention, thereby forming a bidirectional interactive framework.

[0048] ; Through this bidirectional interactive attention, this invention can simultaneously model the correspondence and difference relationships of two perspectives in temporal features, enabling the two perspective links to establish a stable high-order semantic mapping within a shared latent space. To ensure the stability of information propagation and the smoothness of gradient flow, this invention employs residual connections and layer normalization operations after each cross-attention output to fuse the interactively enhanced features with the original perspective features, thereby preserving the integrity of the features of this perspective while introducing complementary information from the other perspective.

[0049] ; The aforementioned bidirectional cross-attention structure offers the following advantages: Firstly, the standard perspective can leverage the spatial vector information from the Frank perspective to enhance its ability to discriminate local areas of myocardial infarction. Secondly, the Frank perspective can obtain finer-grained temporal structure compensation through local detail patterns in the standard leads. The complementary characteristics of the two perspectives are fully integrated under the bidirectional interaction mechanism, enabling the model to maintain a highly stable and semantically consistent cross-perspective representation even when facing different individual differences, different lead combinations, and ECG noise disturbances.

[0050] Furthermore, this invention employs a multi-head structure in bidirectional interactive attention, dividing the latent space projection into multiple subspaces. This allows the model to independently model semantic interactions between perspectives in different frequency bands and sub-feature subspaces. The multi-head structure can capture different waveform details of the time-series curves, such as local changes in the ST segment, T wave, and R peak morphology. This enables each interactive attention path to uncover deep correlations between the two perspectives from different angles, providing high-resolution and highly complementary representation inputs for subsequent global fusion.

[0051] Through the aforementioned bidirectional cross-attention mechanism, this invention successfully constructs a highly compact semantic bridge structure between the standard twelve-lead and Frank orthogonal leads, enabling stable, directional semantic alignment and information exchange between features from both perspectives. The resulting interactive enhanced features... and will be as Figure 5 The output of the upper half path, together with the global fusion features of the second path, forms a complete multi-view fusion structure in step 3.3.

[0052] Step 3.3: Viewpoint-independent global fusion (global representation modeling) After achieving fine-grained semantic alignment between the two perspectives through bidirectional cross-attention, this invention further constructs a perspective-independent global fusion branch to obtain a highly stable potential unified feature representation under different lead configurations, individual differences, and signal-noise conditions. Compared to the "fine-grained compensation between perspectives" emphasized by cross-attention, this branch aims to abstract high-level semantics independent of specific perspectives from a global structure, forming a cross-perspective shared feature space and providing structured support for the final unified fusion representation. This part of the structure is shown in the lower half of the graph.

[0053] This invention first refines the internal features from the standard perspective and the Frank perspective. and By concatenating or weighting the features, local temporal patterns and overall electrophysiological changes from both perspectives can be simultaneously presented in the same feature space. Subsequently, a linear mapping module projects the concatenated features onto a unified representation space to construct the subsequent global attention structure. This linear mapping is implemented using a 1×1 convolutional or fully connected network, corresponding to the global branch initialization operation in the code of this invention.

[0054] ; Features after unified mapping Building upon this foundation, to capture holistic ECG rhythm and waveform trends beyond a single perspective and establish a stable semantic benchmark independent of any single perspective's quality, this invention employs a self-attention mechanism to construct a perspective-independent global representation. Self-attention can directly model the temporal dependencies of the entire ECG segment without perspective segmentation, enabling the model to capture global correlation patterns across scales and leads. This invention calculates internal attention weights by constructing query, key, and value sequences, thereby extracting global waveform trends, rhythmic variations, and long-range dependencies between different waveform segments within the overall structure.

[0055] ; Self-attention can enhance time segments with consistent structural features and suppress noisy segments, making the features obtained by the global branch more biased towards the overall waveform structure, such as the morphological changes in the ventricular repolarization process and the overall energy trend of ventricular depolarization. After the self-attention output, this invention also introduces residual structure and layer normalization, so that the global representation is structurally enhanced while maintaining the original information, improving the stability and expression consistency during subsequent fusion.

[0056] ; To further enhance the model's ability to comprehensively represent local details and global trends, this invention employs a feed-forward network (FFN) to perform nonlinear transformations on the features after global attention. The feed-forward network consists of two layers of linear transformations and activation functions, improving the model's nonlinear expressive power along the feature dimension, enabling it to capture more complex ECG structural relationships. This structure is similar to the FFN in the transformer, but with adaptive parameter settings based on the characteristics of ECG signals.

[0057] ; Through the synergistic effect of self-attention and feedforward networks, this invention constructs highly abstract latent global features without relying on viewpoint segmentation, enabling the model to extract structural patterns related to the extent of myocardial injury, rhythm changes, and overall vector direction changes from the overall electrocardiogram sequence. Unlike the bidirectional cross-attention in 3.2, this branch does not extract information from viewpoint comparison, but rather abstracts global semantics shared across viewpoints from a unified structure.

[0058] Specifically, the global branch of the present invention has the following advantages: First, it can automatically learn the overall waveform structure of ECG signals, including long-range dependence, trend changes, and implicit coupling between leads. Second, it provides a stable semantic space independent of perspective, so that the final fused representation does not depend on the quality of a particular perspective. Third, when faced with significant individual differences or when some leads are affected by noise from a particular perspective, the global branch can provide smooth compensation on the overall structure, making the model more robust to clinically diverse data.

[0059] Finally, this invention denotes the global features processed by the self-attention and feedforward networks as follows: This feature will serve as Figure 4 The output of the lower half-path, along with features from two directions of bidirectional cross-attention. and In step 3.4, they jointly participate in the final fusion to form a unified diagnostic representation that is consistent across perspectives and has structural stability.

[0060] Step 3.4: Unify and merge feature representation output To integrate two complementary types of information, "fine-grained interaction between perspectives" and "globally consistent abstraction," and form a unified representation that is both discriminative and robust, this invention constructs a unified fusion feature generation module. This module structurally integrates features from three sources to generate a final fusion representation that can be used for subsequent deep feature extraction and multi-task diagnostic prediction.

[0061] This invention follows two core objectives in designing its fusion strategy: first, to enable complementary information obtained from different perspectives and dimensions to be fully superimposed in a unified feature space; and second, to ensure that the fused representation maintains stability, consistency, and efficient distinguishability across individuals, across lead combinations, and under different noise levels. Based on this, this invention employs a fusion framework of "interactive features after structural alignment + global abstract features," combined with linear mapping, normalization, and nonlinear activation operations to achieve a smooth transition from a multi-branch structure to a single-path structure.

[0062] In its specific implementation, this invention uses two aligned features from the bidirectional cross-attention branch. , Viewpoint-independent features of global branch output By splicing the three features together, they can be represented as a unified joint feature expression in the same space according to the aligned feature dimensions. To ensure scale consistency between different types of features, this invention performs necessary channel reorganization and temporal dimension alignment operations on the features before splicing, giving them a unified latent space structure.

[0063] ; The concatenated joint features still contain structural differences from the three paths. To further enhance their expressive power in diagnostic tasks, this invention employs a linear mapping (typically implemented through 1×1 convolutions or fully connected layers) to compress and reshape the concatenated features, projecting them onto a unified latent representation dimension. This linear mapping not only reduces channel dimension expansion caused by concatenation but also further models the relationships between the three paths through learnable parameters, enabling the final representation to possess both cross-perspective consistency and global structural stability.

[0064] ; After completing linear projection, this invention further introduces layer normalization and nonlinear activation to improve the stability of the fused features in terms of distribution and enhance the model's plasticity in the feature space. Layer normalization ensures that different samples have consistent distribution across the fused feature dimensions, thereby reducing the interference of differences in the three features on subsequent tasks; nonlinear activation is used to increase the model's nonlinear expressive power during the fusion stage and improve the sensitivity of features to changes in local and global patterns.

[0065] ; The fused representation As the final output of this invention in step 3, it has the following advantages: First, thanks to the fine-grained semantic alignment mechanism of bidirectional interactive attention branches, It can simultaneously contain the interactive structural information between the standard twelve-lead and Frank orthogonal leads; Secondly, a unified feature representation is abstracted from global branches. It exhibits good robustness to individual differences, noise interference, and lead loss; Third, through the structural processing of linear mapping and normalization integration, the fused features have high discriminativeness and consistency, making them suitable for multi-task joint modeling structures.

[0066] Combining the above advantages, the final fusion representation obtained by this invention It retains complementary information from both perspectives while also possessing cross-perspective unity, making it the core input feature for the subsequent step 4 (deep feature extraction and multi-task modeling). This representation is in... Figure 5 The output position clearly shows the final merging form of the two branch paths, providing key technical support for the structural closed loop of the entire model from "dual-view input → multi-path encoding → single-path diagnostic output".

[0067] As one embodiment of the present invention, step S4 includes: Step 4.1: Deep Fusion Feature Extraction Module To further extract deep nonlinear patterns highly correlated with myocardial infarction pathology from the fused high-order features, this invention first maps the cross-view fused representation to a deep feature space. A linear projection layer is used to reorganize the channel dimensions and feature distribution, enabling the deep structure to process the multipath information contained in the fused representation within a unified latent dimension. After linear projection, this invention employs a feedforward network (FFN) structure to perform deep nonlinear transformations on the features, allowing the model to capture complex changes in the higher-order semantics of ECG signals, such as repolarization phase delay caused by myocardial ischemia and temporal shifts in the ST segment elevation region. The overall structure of this module is as follows: Figure 6 As shown.

[0068] To further enhance the model's sensitivity to local continuous waveform changes, this invention introduces a one-dimensional convolutional structure. Through local convolution operations with a fixed receptive field length, local continuous features are extracted from the time axis, capturing significant waveform segments related to myocardial infarction in ECG (such as ST segment abrupt changes and T wave inversion). After convolutional enhancement, residual connections and layer normalization structures are used to ensure the stability of features propagating in deep networks, thus forming the final deep fusion representation. This representation possesses complementary information from both perspectives, a global semantic structure, and deep feature relationships, making it the optimal potential representation for performing multi-task diagnosis.

[0069] ; Step 4.2: Multi-task output head and joint loss modeling In obtaining deep fusion representation Subsequently, this invention constructs a multi-task output structure on a unified deep semantic space, enabling the model to simultaneously perform diagnostic reasoning for myocardial infarction from both global and local perspectives. Since the deep features have already integrated complementary information from both perspectives, global ECG structure, and multi-scale local waveform details, this invention derives multiple parallel semantic interpretation paths from this shared representation through a multi-task framework. This allows the model to complete multi-dimensional tasks such as "whether myocardial infarction has occurred," "location of occurrence," and "cross-perspective structural consistency constraints" within a unified latent space. The multi-task structure not only enhances the discriminativeness and robustness of the features but also provides rich semantic supervision signals for the formation of a high-quality teacher model in step 5, the knowledge distillation stage. The overall structure of this module is as follows: Figure 7As shown.

[0070] This invention first constructs a task to determine the presence or absence of myocardial infarction based on deep features, and then uses a linear mapping structure to... Projecting the model onto a binary semantic space enables it to extract features reflecting the presence of myocardial injury from overall ECG waveform changes, such as ST-segment elevation, T-wave alterations, and Q-wave abnormalities. This task captures global pathological patterns and is one of the core outputs of the entire diagnostic framework of this invention.

[0071] ; Beyond the presence or absence of a lesion, this invention further constructs a lesion location task to characterize the spatial distribution information of myocardial infarction. Since the electrophysiological manifestations of myocardial injury in different anatomical regions exhibit significant lead dependence, the location classification task enables the model to retain locally sensitive features in its deep representation, allowing it to distinguish injury features in different locations such as the anterior wall, inferior wall, and lateral wall. This task allows the model to learn to pay attention to spatial differences and local structural changes between leads while sharing the backbone, achieving fine-grained inference of lesion location.

[0072] ; To further enhance the robustness and structural consistency of the deep feature space, in a preferred embodiment, this invention introduces an ECG signal reconstruction task as auxiliary supervision. This ensures that deep features retain sufficient temporal structural information and perspective correspondences to recover ECG waveforms of standard or Frank orthogonal leads from the latent space. In this way, this invention imposes strong structural constraints on the latent representation during training, prompting features to retain richer original waveform structural information. This not only enhances the stability of the feature space but also facilitates the transmission of more comprehensive waveform semantics to the student model during the knowledge distillation stage. In particular, within the distillation framework of this invention, the information obtained by the teacher model through multi-task learning is more comprehensive, helping to transmit richer soft-label structures to the student model.

[0073] ; To enable the three types of tasks to mutually reinforce each other during training, this invention employs a joint loss function to weightedly integrate the MI discrimination loss, site localization loss, and reconstruction loss, allowing multiple tasks to work collaboratively under the same optimization objective. The joint loss not only constrains the model to focus on diagnostic signals across different dimensions but also promotes the joint expression of global semantics, local spatial structure, and temporal reconstruction information at the feature level, thereby forming a stable and clinically meaningful unified representation. This joint optimization strategy is crucial for the distillation process in step 5, as the student model obtains structured supervision signals from the multi-task output of the teacher model, thus learning a more comprehensive cross-perspective semantic pattern.

[0074] ; in, For binary classification, cross-entropy loss, For multi-class cross-entropy loss, To reconstruct the loss; through weights , , The joint loss function, through adjustment, guides the model to balance the optimization of different tasks in the shared feature space, enabling features to simultaneously contain global discrimination, local localization, and waveform structure information, thus providing a rich source of supervisory signals for knowledge distillation.

[0075] Through the aforementioned multi-task output structure and joint loss modeling, this invention constructs a comprehensive diagnostic system covering "global discrimination - local localization - structural consistency" in a unified deep representation. This enables the model to automatically capture key waveform structures and translead relationships related to myocardial infarction without relying on hard-coded ECG rules, thus laying a complete semantic foundation for the subsequent construction of deployment models based on knowledge distillation.

[0076] As one embodiment of the present invention, step S5 includes: The structure of this knowledge distillation-based model transfer framework is as follows: Figure 8 As shown. In the dual-perspective teacher model constructed in this invention, the standard twelve leads and Frank orthogonal leads jointly participate in encoding, enabling the model to form a deep representation with high expressiveness and structural consistency in the cross-perspective semantic space. However, in real clinical deployment environments, most devices only have standard twelve-lead input and cannot provide Frank perspective signals. Therefore, directly deploying the dual-perspective teacher model is not feasible in terms of input conditions and computational cost. To solve this problem, this invention proposes a lead perspective transfer method based on knowledge distillation. By constructing a lightweight student model, it can still inherit the cross-perspective semantic capabilities of the teacher model under the condition of relying only on twelve-lead input, achieving diagnostic performance similar to the teacher model. The structure of this step corresponds to... Figure 7 As shown, this is the key design of the present invention to achieve the transition from "multi-view training" to "single-view deployment".

[0077] This invention first constructs a lightweight student model, whose network structure is significantly simplified compared to the teacher model, with fewer layers, fewer channels, and lower computational complexity, and it only relies on twelve-lead signals as input. Although the student model has a lightweight structure, it can learn its cross-perspective semantic structure through the teacher model's supervision signals during the distillation training phase. During distillation, the teacher model, as the builder of the multi-perspective feature space, has already learned rich multi-scale details, cross-lead structural relationships, and multi-task diagnostic semantics through steps 3 and 4. Therefore, this invention uses distillation loss to allow the student model to gradually approximate the teacher model's expression and decision-making process during training, thereby compensating for the representational deficiencies caused by the lack of Frank's perspective input in the student model.

[0078] In predictive distillation, this invention employs a temperature-based logits distillation method to ensure that the classification probability distribution of the student model aligns with the soft label distribution of the teacher model. By introducing temperature TTT into the softmax function, the output distribution of the teacher model is softened, enabling the model to present a more granular relative confidence relationship between different categories. This allows the student model to learn the diagnostic bias and inter-class structure of the teacher model in the MI discrimination task.

[0079] ; in, and Logits and temperature for teachers and students, respectively. Adjustments can be made during training based on data characteristics and to optimize stability.

[0080] In addition to predictive distillation, this invention further introduces feature distillation, enabling the student model to align with the latent representation of the teacher model in a deep semantic space. Feature distillation minimizes the differences between teacher and student features, allowing the student model to gradually learn the internal semantic organization of the teacher model, including important information such as cross-perspective fusion structure, global and local waveform patterns, and time-series structure. This mechanism is particularly crucial for the student model because its input lacks a Frank perspective, and feature distillation allows it to indirectly learn the latent semantics related to the Frank perspective.

[0081] ; In addition to MI discriminative distillation and feature distillation, this invention also utilizes the multi-task structure of the teacher model to impose additional semantic constraints on the student model, enabling it to learn not only the MI discriminative boundary but also the internal relationships between lesion location and optional reconstruction tasks. In step 4, the teacher model constructs a structured deep semantic space through multi-tasks, while the student model absorbs these semantics during distillation training through soft labels and task distillation loss, making the final latent representation more stable and clinically interpretable.

[0082] To enable the distillation loss and task loss to work together, this invention constructs a joint distillation loss, which integrates predictive distillation, feature distillation, and task loss according to weight parameters α, β, and γ. It is important to emphasize that the distillation weights in this invention are adjustable parameters, capable of being flexibly adjusted based on the characteristics of the training data, the signal strength of the teacher model, and the structural complexity of the student model, thereby ensuring the stability and efficiency of the distillation process. This characteristic strictly corresponds to the configurable parameter design in the training script.

[0083] ; By introducing the aforementioned multi-dimensional distillation mechanism, the student model of this invention can achieve cross-perspective diagnostic capabilities approaching those of the teacher model using only twelve leads as input, thereby realizing a low-cost, high-efficiency, and easily deployable myocardial infarction diagnostic system. This invention achieves a natural transition from a "dual-perspective training system" to a "single-perspective deployment framework," enabling the model to complete highly clinically valuable automated diagnostic tasks in practical use without relying on additional equipment.

[0084] This invention constructs a cross-view electrocardiogram intelligent diagnostic framework for myocardial infarction, characterized by "multi-view training and single-view deployment." Its innovations are reflected in the following two aspects: (1) Diagnostic system architecture based on multi-perspective fusion and knowledge distillation First, a multi-view teacher model is constructed, comprising a two-branch input encoding layer for processing standard 12-lead and Frank orthogonal leads, a cross-view fusion layer for fusing dual-view features, and a classification output layer for multi-task learning based on the fused features. Second, through knowledge distillation, the diagnostic capabilities of the teacher model are transferred to a lightweight student model. This student model only needs to input standard 12-lead signals to achieve myocardial infarction diagnosis with performance comparable to the teacher model. This architecture is the cornerstone of this invention, protecting the complete technical path from model training to clinical deployment.

[0085] (2) Achieving cross-perspective knowledge transfer Firstly, there is a multi-path cross-view feature fusion mechanism. Its characteristic is that it does not simply splice together features from two perspectives, but rather adopts a collaborative strategy that includes at least multi-scale fusion within a perspective, bidirectional semantic alignment between perspectives, and perspective-independent global modeling, in order to generate a unified feature representation with stronger discriminative power.

[0086] Secondly, it employs a multi-dimensional knowledge distillation strategy. Its key feature is that, when transferring knowledge from the teacher model to the student model, it not only distills the final classification prediction result but also constrains the student model to align with the teacher model's deep feature semantics through a loss function. Furthermore, it can integrate structured supervision signals from the teacher model's multi-task learning, thereby achieving deep and efficient diagnostic knowledge transfer.

[0087] The present invention has the following beneficial effects: (1) Improved diagnostic accuracy and robustness. Addressing the issue of incomplete spatial information in existing methods, this invention employs a dual-view input approach using standard twelve leads and Frank orthogonal leads, and designs a specialized mechanism for bidirectional semantic alignment and global fusion between viewpoints. This enables the model to effectively integrate three-dimensional spatial information and two-dimensional planar projection information of cardiac electrical activity. This not only overcomes the inherent limitations of a single viewpoint but also further enhances the representational ability of features through multi-task learning. Therefore, the method described in this invention exhibits superior discrimination accuracy and stronger robustness when facing complex lesions, individual anatomical differences, and signal interference.

[0088] (2) Breakthrough in Deployment Feasibility and Practicality. Addressing the challenge of deploying complex models in a lightweight manner, this invention constructs a "multi-perspective teacher-single-perspective student" distillation framework and optimizes it through multi-dimensional (prediction, feature, task) knowledge distillation loss, achieving efficient transfer of core diagnostic knowledge from complex models to lightweight models. This allows the final deployed model to achieve diagnostic performance approaching that of a multi-perspective fusion model with only standard twelve-lead input, while significantly reducing model complexity and computational resource consumption. Therefore, this invention fundamentally solves the bottleneck of applying high-performance diagnostic models to conventional clinical equipment, possessing outstanding practical value and broad prospects for promotion.

[0089] Example 2 The present invention also provides a cross-view electrocardiogram myocardial infarction diagnostic device, comprising: The first processing module is used to preprocess the acquired raw multi-view electrocardiogram signals; The second processing module is used to extract multi-level temporal features from the standard twelve-lead and Frank orthogonal lead signals respectively, based on the preprocessed original multi-view electrocardiogram signals and using lead-specific encoders. The third processing module is used to perform multi-scale fusion within the viewpoint, semantic alignment between viewpoints, and global fusion independent of viewpoints in sequence according to multi-level temporal features. The fourth processing module is used to perform deep transformation on the fused features and construct multiple supervised tasks based on this, such as the determination of the presence or absence of myocardial infarction and the classification of lesion sites, and perform joint optimization to obtain a multi-view, multi-task model. The fifth processing module is used to train a lightweight student model that only receives standard twelve-lead input, using a multi-view, multi-task model as the teacher, to achieve efficient transfer of diagnostic knowledge and obtain the final deployment model to realize cross-view electrocardiogram myocardial infarction diagnosis.

[0090] As one embodiment of the present invention, the first processing module filters, denoises, extracts and normalizes the original multi-view electrocardiogram signal to obtain a standardized heartbeat sample.

[0091] As one embodiment of the present invention, the multi-view multi-task model employs a joint loss function to weightedly integrate the MI discrimination loss, part localization loss, and reconstruction loss, enabling multiple tasks to work collaboratively under the same optimization objective. ; Through weight , , The joint loss function is adjusted to guide the model to balance the optimization of different tasks in the shared feature space, so that the features simultaneously contain global discrimination, local localization and waveform structure information.

[0092] Example 3 The present invention also provides a cross-view electrocardiogram myocardial infarction diagnostic system, comprising: a memory and a processor, wherein the memory stores a computer program executed by the processor, and the computer program, when executed by the processor, performs a cross-view electrocardiogram myocardial infarction diagnostic method.

[0093] Example 4 The present invention also provides a storage medium storing a computer program that, when running, executes a cross-view electrocardiogram method for diagnosing myocardial infarction.

[0094] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A cross-view electrocardiogram method for diagnosing myocardial infarction, characterized in that, include: Step S1: Preprocess the acquired raw multi-view electrocardiogram signals; Step S2: Based on the preprocessed original multi-view ECG signals, use lead-specific encoders to extract multi-level temporal features from the standard twelve-lead and Frank orthogonal lead signals, respectively. Step S3: Based on the multi-level temporal features, perform multi-scale fusion within the viewpoint, semantic alignment between viewpoints, and global fusion independent of the viewpoint in sequence. Step S4: Perform deep transformation on the fused features, and based on this, construct multiple supervised tasks such as myocardial infarction presence / absence discrimination and lesion site classification, perform joint optimization, and obtain a multi-view multi-task model; Step S5: Using the multi-view, multi-task model as the teacher, train a lightweight student model that only receives standard twelve-lead input to achieve efficient transfer of diagnostic knowledge and obtain the final deployment model to realize cross-view electrocardiogram myocardial infarction diagnosis.

2. The cross-view electrocardiogram method for diagnosing myocardial infarction as described in claim 1, characterized in that, In step S1, the original multi-view ECG signal is filtered, denoised, and the heartbeats are extracted and normalized to obtain standardized heartbeat samples.

3. The cross-view electrocardiogram method for diagnosing myocardial infarction as described in claim 1, characterized in that, The multi-view, multi-task model employs a joint loss function to weight and integrate the MI discrimination loss, part localization loss, and reconstruction loss, enabling multiple tasks to work collaboratively under the same optimization objective. ； Through weight , , The joint loss function is adjusted to guide the model to balance the optimization of different tasks in the shared feature space, so that the features simultaneously contain global discrimination, local localization and waveform structure information.

4. A cross-view electrocardiogram diagnostic device for myocardial infarction, characterized in that, include: The first processing module is used to preprocess the acquired raw multi-view electrocardiogram signals; The second processing module is used to extract multi-level temporal features from the standard twelve-lead and Frank orthogonal lead signals respectively, based on the preprocessed original multi-view electrocardiogram signals using lead-specific encoders. The third processing module is used to perform multi-scale fusion within the viewpoint, semantic alignment between viewpoints, and global fusion independent of viewpoints in sequence according to multi-level temporal features. The fourth processing module is used to perform deep transformation on the fused features and construct multiple supervised tasks based on this, such as the determination of the presence or absence of myocardial infarction and the classification of lesion sites, and perform joint optimization to obtain a multi-view, multi-task model. The fifth processing module is used to train a lightweight student model that only receives standard twelve-lead input, using a multi-view, multi-task model as the teacher, to achieve efficient transfer of diagnostic knowledge and obtain the final deployment model to realize cross-view electrocardiogram myocardial infarction diagnosis.

5. The cross-view electrocardiogram myocardial infarction diagnostic device as described in claim 4, characterized in that, The first processing module filters, denoises, extracts and normalizes the original multi-view ECG signals to obtain standardized heartbeat samples.

6. The cross-view electrocardiogram myocardial infarction diagnostic device as described in claim 5, characterized in that, The multi-view, multi-task model employs a joint loss function to weight and integrate the MI discrimination loss, part localization loss, and reconstruction loss, enabling multiple tasks to work collaboratively under the same optimization objective. ； Through weight , , The joint loss function is adjusted to guide the model to balance the optimization of different tasks in the shared feature space, so that the features simultaneously contain global discrimination, local localization and waveform structure information.

7. A cross-view electrocardiogram myocardial infarction diagnostic system, characterized in that, include: A memory and a processor, wherein the memory stores a computer program executed by the processor, the computer program, when executed by the processor, performs the cross-view electrocardiogram method for diagnosing myocardial infarction as described in any one of claims 1-3.

8. A storage medium, characterized in that, The storage medium stores a computer program that, when executed, performs the cross-view electrocardiogram method for diagnosing myocardial infarction as described in any one of claims 1-3.

Images