Cardiovascular disease classification system and method of multi-source vibration spectrum fusion network based on common-private features, and application

Abstract

The invention relates to a cardiovascular disease classification system and method of a multi-source vibration spectrum fusion network based on common-private features and application. A cardiovascular disease classification system of a multi-source vibration spectrum fusion network based on common-private features comprises a data collection module for collecting Raman spectrum data and infrared spectrum data of serum; the data preprocessing module is used for preprocessing the Raman spectrum data and the infrared spectrum data; the multi-scale private feature extractor module is used for extracting multi-scale private features of the preprocessed Raman spectrum data and infrared spectrum data; the dynamic sharing characteristic generation module is used for extracting dynamic sharing characteristics of the preprocessed Raman spectrum data and infrared spectrum data; the bidirectional attention fusion module is used for performing deep interaction and adaptive fusion on the multi-scale private features; and the classifier module is used for splicing and classifying the fused multi-scale private features and the dynamic sharing features. According to the technical scheme, the classification performance is good.

Description

Technical Field

[0001] This invention pertains to medical data processing, specifically relating to a cardiovascular disease classification system and method based on a multi-source vibrational spectral fusion network with shared and private features, and its applications. Background Technology

[0002] Aortic dissection (AD) is a life-threatening cardiovascular emergency with a high global incidence. It has a rapid onset, progresses rapidly, and has an extremely high mortality rate; delayed treatment increases mortality by 1-2% per hour. Early, rapid, and accurate diagnosis is crucial to reducing mortality. Computed tomography angiography (CTA) is currently the "gold standard" for diagnosing AD, possessing extremely high sensitivity and specificity. Transesophageal echocardiography (TEE) and magnetic resonance angiography (MRA) are also important methods. However, CTA has limitations such as radiation exposure, expensive equipment, and difficulty in performing it quickly at the bedside. TEE is a semi-invasive procedure, and MRA is time-consuming; neither is suitable for extremely critically ill or unstable patients. D-dimer testing is often used to rule out other diagnoses, but its specificity is not high.

[0003] Aortic stenosis (AS) is a condition characterized by restricted aortic valve opening, leading to obstruction of the left ventricular outflow tract. Its severity can be classified into C0 (mild calcification) and C1 (severe calcification) types based on factors such as the degree of calcification. The disease typically progresses chronically, but severe stenosis can cause acute decompensation, leading to heart failure or even sudden death. Transthoracic echocardiography (TTE) is the preferred non-invasive method for assessing aortic stenosis. It measures transvalvular velocity, pressure gradient, and valve orifice area to determine severity and helps identify the degree of calcification. However, the accuracy of echocardiographic assessment is highly dependent on the operator's experience. For very mild or very severe lesions, there may be misjudgment, affecting the timing of early intervention or surgery.

[0004] Therefore, there is an urgent clinical need for a new method that can overcome the limitations of existing technologies and achieve rapid, non-invasive, objective, and accurate assessment of the degree of calcification in AD and AS.

[0005] It is noteworthy that the clinical manifestations of aortic aortic stenosis (AD) and severe aortic stenosis (AS) overlap, as both can present with chest pain, dyspnea, syncope, and even cardiogenic shock. This can lead to AD being misdiagnosed as an acute exacerbation of aortic stenosis. Especially when AD involves the aortic valve root and causes acute aortic regurgitation, the resulting murmurs and hemodynamic changes can easily be confused with severe aortic stenosis. Although both are cardiovascular emergencies, their treatment principles and urgency are drastically different. AD requires immediate surgery or endovascular repair, and thrombolysis or anticoagulation is contraindicated. While severe aortic stenosis (C1 type) ultimately requires surgical intervention, it usually allows for a period of pharmacological stabilization and elective evaluation. Blood pressure management strategies also need to be individualized, and hypotension should be avoided in patients with severe stenosis. Therefore, errors in differential diagnosis can directly lead to missed treatment opportunities and even catastrophic consequences due to inappropriate medication, such as incorrect anticoagulation in AD patients. Developing an objective diagnostic method that can rapidly and accurately differentiate the severity of AD and aortic stenosis (AS) is of urgent and significant clinical importance for optimizing clinical decision-making and improving patient prognosis.

[0006] Multi-source fusion is a key driving force for breaking through the performance bottleneck of single information sources and promoting the development of AI-assisted disease diagnosis. Its core lies in integrating complementary information from different sources to form a more comprehensive and accurate representation of the target object.

[0007] While numerous studies and applications have been conducted on multi-source spectral fusion technology, existing methods still have significant limitations, restricting the depth and breadth of their clinical application. For example, most fusion strategies remain at the level of low-level feature splicing or post-decision weighting, lacking explicit modeling of deep complementary relationships and differential features among multi-source spectra. This leads to insufficient information utilization and discriminative features being easily overwhelmed by noise or redundancy. Furthermore, current models generally lack explicit separation mechanisms for "common features versus private features." In addition, existing methods generally lack dynamic adaptability, failing to adjust fusion weights according to sample characteristics, resulting in poor generalization ability when facing individual differences, thus limiting their stability and reliability in real-world medical scenarios.

[0008] In view of this, this invention proposes a novel cardiovascular disease classification system and method based on multi-source vibrational spectral fusion of shared and private features, and its application. The system includes the MVSFNet model, which separates and fuses shared and private features from infrared and Raman spectroscopy. Through multi-scale feature extraction, dynamic weight sharing, and a bidirectional attention mechanism, it fully exploits the complementary information and discriminative features between the two modal data, effectively avoiding feature confusion and enhancing the model's discriminative ability. Summary of the Invention

[0009] The purpose of this invention is to provide a cardiovascular disease classification system based on a multi-source vibrational spectral fusion network with shared and private features. This system not only ensures the ability to learn the commonalities between modes to enhance generalization ability, but also gives it sufficient flexibility to adapt to the unique properties of different modes, thereby exhibiting stronger adaptability and robustness on complex and variable biomedical data.

[0010] To achieve the above objectives, the technical solution adopted is as follows:

[0011] A cardiovascular disease classification system based on a multi-source vibrational spectral fusion network with shared and private features, comprising:

[0012] The data collection module collects Raman and infrared spectral data of serum.

[0013] The data preprocessing module preprocesses the Raman spectral data and infrared spectral data.

[0014] The multi-scale private feature extractor module uses a deep convolutional neural network structure to extract multi-scale private features from the preprocessed Raman spectral data and infrared spectral data.

[0015] The dynamic shared feature generation module adopts a two-level structure of dual-modal projection-dynamic gating fusion to extract the dynamic shared features of the preprocessed Raman spectral data and infrared spectral data.

[0016] The bidirectional attention fusion module performs deep interaction and adaptive fusion on the aforementioned multi-scale private features;

[0017] The classifier module concatenates the fused multi-scale private features with the dynamically shared features and performs classification.

[0018] Furthermore, the multi-scale private feature extractor module includes a private feature extractor for Raman spectral data and a private feature extractor for infrared spectral data.

[0019] Furthermore, the private feature extractor of the Raman spectral data is a parallel multi-scale one-dimensional convolutional neural network consisting of three parallel branches. The output feature maps of the branches are concatenated into feature vectors after global average pooling, and finally dimensionality reduction and fusion are performed through a fully connected layer to output the Raman spectral private feature vector.

[0020] The private feature extractor for the infrared spectral data is a serial-parallel hybrid multi-scale convolutional network containing two main branches. The outputs of the two main branches are both subjected to global average pooling to obtain two feature vectors, which are then concatenated to obtain fused features. Finally, the feature vectors are projected onto the infrared spectral private feature vectors through a fully connected layer.

[0021] Furthermore, in the private feature extractor of the Raman spectral data, the branch kernel sizes are 3, 5, and 7, respectively.

[0022] Furthermore, in the private feature extractor of the infrared spectral data, the first main branch path first uses a convolutional layer with a kernel size of 5 to extract basic features, then reduces the sequence length through a max pooling layer; and then refines semantic features through a convolutional layer with a kernel size of 3.

[0023] The second main branch path first uses a large convolutional kernel with a kernel size of 15, then reduces the sequence length through an aggressive pooling layer; then it optimizes the coarse-grained features after pooling through a convolutional layer with a kernel size of 5.

[0024] Furthermore, the dynamic shared feature generation module maps infrared and Raman features to a feature space of the same dimension through an independent linear projection layer, generates sample-level weights through a gating network, and performs weighted fusion of the dual-modal projection features to finally output dynamic shared features.

[0025] Furthermore, in the aforementioned bidirectional attention fusion module, using infrared private features as the query and Raman private features as the key, the fusion process is as follows:

[0026] Attention employs a scaled dot product mechanism, supporting multi-head attention to capture multi-dimensional dependencies. For single-head attention, the attention score is... ,in It is a single-head dimension;

[0027] The scores are normalized to a probability distribution using Softmax, ensuring that the sum of the weights is 1. Applying attention weights to the value vector yields attention features that focus on key information. ; Concatenate the bidirectional attention features into The feature vectors are used to output gate weights through a two-layer network. ,in , ;

[0028] Perform weighted fusion.

[0029] Furthermore, the bidirectional attention fusion module employs residual path and layer normalization processing. The process is as follows: the mean of the original private features is used as the residual term. After retaining the basic feature information, the result of the residual addition is layer normalized. The formula is: ,in and These are the mean and variance of the feature, respectively. and For learnable scaling and offset parameters, To prevent the denominator from being 0.

[0030] Another objective of this invention is to provide a cardiovascular disease classification method based on a multi-source vibrational spectral fusion network with shared and private features. This method fully exploits the complementary information and discriminative features between the two modal data through multi-scale feature extraction, dynamic weight sharing, and bidirectional attention mechanism.

[0031] To achieve the above objectives, the technical solution adopted is as follows:

[0032] A cardiovascular disease classification method based on a multi-source vibrational spectral fusion network with shared and private features, employing the aforementioned cardiovascular disease classification system, includes the following steps:

[0033] (1) Collect Raman and infrared spectral data of serum and tears;

[0034] (2) Preprocess the Raman spectral data and infrared spectral data;

[0035] (3) Extract the multi-scale private features of the preprocessed Raman spectral data and infrared spectral data using a multi-scale private feature extractor;

[0036] A two-level dynamic shared feature generation method based on dual-modal projection-dynamic gating fusion is used to extract the dynamic shared features of the preprocessed Raman spectral data and infrared spectral data.

[0037] (4) After performing deep interaction and adaptive fusion on the multi-scale private features, they are concatenated with the dynamic shared features for classification.

[0038] Another objective of this invention is to provide the application of the above-mentioned cardiovascular disease classification system and method in cardiovascular disease auxiliary diagnostic devices.

[0039] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0040] This invention discloses a cardiovascular disease classification system and method based on a multi-source vibrational spectral fusion network with shared and private features, and its application. This novel model, MVSFNet, innovatively and explicitly separates and fuses shared and private features from infrared and Raman spectra. Through multi-scale feature extraction, dynamic weight sharing, and a bidirectional attention mechanism, it fully exploits the complementary information and discriminative features between the two modal data. Specific advantages include:

[0041] 1. The technical solution of this invention proposes a general framework MVSFNet that integrates "feature separation-fusion" and "multi-scale perception". This framework effectively avoids feature confusion and enhances the model's discriminative ability by explicitly decomposing the input spectral information into common features and private features.

[0042] 2. The technical solution of the present invention utilizes parallel multi-scale convolutional layers to capture robust local and global feature representations from different granular levels, providing a richer and more robust information source for subsequent fusion.

[0043] 3. The technical solution of this invention designs a bidirectional attention feature fusion module. This module abandons the simple feature concatenation or weighted addition strategy, and instead uses an attention mechanism to dynamically and bidirectionally evaluate and weight the importance of features from different sources. This enables the system, i.e., the model, to adaptively focus on the feature elements that are most discriminative for the current discrimination task, achieving accurate and efficient feature fusion, and significantly improving fusion efficiency and final classification performance.

[0044] 4. The technical solution of this invention introduces a dynamic weight sharing mechanism. This mechanism enables the network to adaptively adjust its parameter sharing strategy according to task requirements during training, rather than statically sharing all weights. This design ensures that the model can learn the commonalities between modalities to enhance generalization ability, while also giving it sufficient flexibility to adapt to the unique properties of different modalities, thus exhibiting stronger adaptability and robustness on complex and variable biomedical data.

[0045] 5. The technical solution proposed in this invention has achieved excellent performance on serum spectral datasets of aortic dissection (AD) and aortic stenosis (AS) from clinical practice. In the AD identification task, the accuracy reached 95.00% with an AUC value of 0.9955; in the AS calcification grading task, the accuracy was 91.30% with an AUC value of 0.9697, significantly outperforming baseline models such as tensor fusion, bilinear fusion, and decision-level fusion.

[0046] In addition to the main experiment, this embodiment also systematically conducted ablation experiments to verify the effectiveness of each module, including dynamic weight sharing, multi-scale feature extraction, and bidirectional attention fusion. Furthermore, the model also performed exceptionally well in the differential diagnosis of AD and AS, achieving an accuracy of 97.37% and an AUC of 0.9832, demonstrating excellent generalization ability and clinical applicability.

[0047] Therefore, the technical solution of the present invention can not only be used in a high-precision, non-invasive cardiovascular disease auxiliary diagnostic system, but also provides a novel deep learning method for the fusion of multi-source biomedical data, which has important clinical translation prospects. Attached Figure Description

[0048] Figure 1 shows the framework diagram of the MVSFNet model.

[0049] Figure 2 For C0 and C1 of AS, (a) the original average image of the infrared spectrum and (b) the original average image of the Raman spectrum.

[0050] Figure 3 (a) Average image after infrared spectral preprocessing and (b) Average image after Raman spectral preprocessing for C0 and C1 of AS.

[0051] Figure 4 Detailed analysis of the C0 and C1 infrared spectra of AS;

[0052] Figure 5 Detailed analysis of the C0 and C1 Raman spectra of AS.

[0053] Figure 6 For MVSFNet classification of aortic dissection and healthy control group, (a) confusion matrix on the test set, (b) loss curves on the training set and validation set, (c) ROC curve.

[0054] Figure 7 For MVSFNet classification of mild and severe aortic stenosis, (a) confusion matrix on the test set, (b) loss curves on the training and validation sets, and (c) ROC curve.

[0055] Figure 8 The dynamic weight module was removed from MVSFNet for the classification of aortic dissection and healthy control group. (a) Confusion matrix on the test set, (b) Loss curves on the training and validation sets, (c) ROC curve.

[0056] Figure 9 The multi-scale feature extraction module was removed from MVSFNet for the classification of aortic dissection and healthy control group. (a) Confusion matrix on the test set, (b) Loss curves on the training and validation sets, (c) ROC curve.

[0057] Figure 10 The bidirectional attention module was removed from MVSFNet for the classification of aortic dissection and healthy control groups. (a) Confusion matrix on the test set, (b) Loss curves on the training and validation sets, and (c) ROC curve.

[0058] Figure 11 For MVSFNet binary classification of AD and heavy AS (C1), (a) confusion matrix on the test set, (b) loss curves on the training and validation sets. Detailed Implementation

[0059] To further illustrate the cardiovascular disease classification system and method based on a multi-source vibrational spectral fusion network with shared-private features, and its applications, and to achieve the intended objectives of this invention, the following detailed description, in conjunction with preferred embodiments, provides a detailed explanation of the specific implementation, structure, features, and effects of the cardiovascular disease classification system and method based on a multi-source vibrational spectral fusion network with shared-private features proposed in this invention. In the following description, different "embodiments" or "embodiments" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable manner.

[0060] Before elaborating on the cardiovascular disease classification system and method based on a multi-source vibrational spectral fusion network with shared and private features, and its applications, it is necessary to further explain the relevant background mentioned in this invention to achieve better results.

[0061] Multi-source fusion is a key driving force for overcoming the performance bottleneck of single information sources and promoting the development of AI-assisted disease diagnosis. Its core lies in integrating complementary information from different sources to form a more comprehensive and accurate characterization of the target object. In the field of biomedical optics, infrared (IR) and Raman spectroscopy, as two complementary vibrational spectroscopy techniques, can non-invasively and efficiently acquire the "fingerprint" information of biological samples from the perspective of molecular bond vibrational energy levels, providing a new technical path for the early detection and accurate identification of diseases.

[0062] Raman spectroscopy is a scattering spectroscopy technique that obtains information about the structure, vibration, and rotation of a sample by inelastic scattering of incident light with the vibrations of molecules within the material. Xuechang Han et al. combined serum Raman spectroscopy with D-dimer detection for the early diagnosis of acute aortic dissection (AAD). Renxing Song et al. used confocal Raman spectroscopy combined with machine learning to assess myocardial hypertrophy and heart failure in mice, achieving an accuracy of 85% using a random forest model. Conversely, infrared spectroscopy utilizes the principles of absorption spectroscopy to identify chemical bonds and their characteristic vibrational modes by measuring changes in the intensity of infrared radiation in a material, thereby determining the composition and structure of the sample. The application of infrared spectroscopy in other diseases provides a basis and reference for the diagnosis of AD and AS. Mustafa Talip Sener et al. used attenuated total reflectance Fourier transform infrared spectroscopy (ATR-FTIR) and multivariate analysis techniques to provide a rapid, non-destructive, and highly accurate method for early post-mortem diagnosis of acute myocardial infarction. Although the two are complementary in mechanism, designing an efficient fusion architecture to deeply explore the intrinsic correlation and complementary advantages between these two vibrational spectral data, and to efficiently diagnose and differentiate AD and AS, still faces significant challenges and technical bottlenecks. For multi-source spectral fusion technology, there have been many studies and applications. Chima Robert et al. fused infrared and Raman spectra at medium, high, and low levels, constructing a PLS model to quantitatively assess red meat quality, demonstrating the effectiveness of traditional chemometric fusion methods. With the rise of deep learning technology, its powerful data-driven characteristics and high-dimensional data processing capabilities have brought new breakthroughs to spectral analysis. Deep learning can not only automatically extract features, reducing reliance on tedious manual feature engineering, but also significantly improve noise suppression and irrelevant information filtering, thereby optimizing the spectral data analysis process. Haoze Li et al. constructed a one-dimensional convolutional neural network model with an embedded multi-head attention mechanism, combining Raman spectroscopy and ATR-FTIR to classify eight microplastics. In the field of disease diagnosis, Hongyong Leng et al. constructed low-level fusion and feature fusion networks, fusing FTIR and Raman spectroscopy to classify serum samples of non-small cell lung cancer, glioma, and esophageal cancer, achieving significant results. To further enhance intermodal interaction, Xuguang Zhou et al. proposed the Cross Branch Co-Attention Network (CBCAN), which demonstrated superior performance on various cancer datasets. Xiangnan Chen et al. introduced frequency band adaptive decomposition and a co-attention mechanism, effectively alleviating the modal overlap problem.

[0063] While the aforementioned studies have fully demonstrated the potential of vibrational spectroscopy in the rapid diagnosis of complex diseases, existing methods still have significant limitations, restricting the depth and breadth of their clinical application. Most fusion strategies remain at the level of low-level feature splicing or post-decision weighting, lacking explicit modeling of the deep complementary relationships and differential features between infrared and Raman spectra. This leads to insufficient information utilization and discriminative features being easily overwhelmed by noise or redundancy. Current models generally lack explicit separation mechanisms for "common features versus private features." Although infrared and Raman spectroscopy can characterize biological samples from different molecular vibrational perspectives, they contain both common information reflecting the basic components of blood and their own private information. Furthermore, existing methods generally lack dynamic adaptability, failing to adjust fusion weights according to sample characteristics, resulting in poor generalization ability when facing individual differences, thus limiting their stability and reliability in real-world medical scenarios.

[0064] Therefore, to address the aforementioned challenges, this paper proposes a novel cardiovascular disease classification system and method based on multi-source vibrational spectral fusion of shared and private features, along with its application.

[0065] Having understood the relevant background mentioned in this invention, the following will provide a more detailed description of the cardiovascular disease classification system and method based on a multi-source vibrational spectral fusion network with shared and private features, and its applications, in conjunction with specific embodiments:

[0066] Aortic dissection (AD) and aortic stenosis (AS) are two fatal cardiovascular diseases with overlapping clinical manifestations but vastly different treatment strategies, making rapid and accurate differential diagnosis crucial. Existing imaging diagnostic methods suffer from limitations such as radiation exposure, invasiveness, and operator dependence. Vibrational spectroscopy techniques (infrared and Raman spectroscopy) can non-invasively acquire molecular "fingerprint" information from biological samples, providing a new avenue for disease diagnosis. However, the information provided by a single spectral modality is limited, and existing fusion methods struggle to deeply mine the complementary and differential features between multi-source data. Therefore, this paper proposes a multi-source vibrational spectroscopy fusion network (MVSFNet) based on shared and private features. This network extracts private features from infrared and Raman spectra separately through multi-scale convolutional modules, generates shared features using a dynamic weight sharing mechanism, and designs a bidirectional attention module to achieve deep interaction and adaptive fusion of dual-modal features. Experiments on clinical AD and AS serum spectral datasets show that MVSFNet achieves an accuracy of 95.00% and an AUC of 0.9955 in the AD diagnosis task, and 91.30% and an AUC of 0.9697 in the AS calcification grading task, significantly outperforming various baseline fusion models and single-modal models. In particular, the model achieves an accuracy of 97.37% in the differential diagnosis of AD and severe AS, demonstrating excellent performance and generalization ability. Ablation experiments validated the effectiveness of each core module. This invention provides a new solution for rapid, non-invasive, and accurate auxiliary diagnosis of cardiovascular diseases and offers a novel deep learning framework for multi-source biomedical data fusion.

[0067] Example 1.

[0068] A. Materials and Methods

[0069] 1. Materials

[0070] This embodiment was approved by the Ethics Committee of the People's Hospital of Xinjiang Uygur Autonomous Region to obtain the infrared and Raman spectroscopy serum samples required for the experiment. All subjects signed informed consent forms. Two disease datasets were used in the experimental design. The first dataset included patients with aortic dissection type A and healthy controls (without hypertension), with 127 and 107 serum samples, respectively. The second dataset included patients with mild and severe aortic stenosis, with 50 and 40 serum samples, respectively. 3 mL of fresh blood without anticoagulants was collected from each sample. To minimize interference from external factors such as diet, blood collection was performed in the morning after participants had fasted for 12 hours. The blood samples were then centrifuged at 4000 r / min at 4 °C, and the uppermost clear liquid was extracted to obtain serum. The serum was separated into centrifuge tubes and stored at -80 °C for analysis.

[0071] 2. Spectral data acquisition

[0072] All samples in this experiment were thawed and measured at room temperature (22°C). Serum Raman spectra were obtained using a high-resolution confocal Raman spectrometer (LabRAM HR Evolution, Gora Raman spectroscopy, IdeaOptics).

[0073] The spectrum was obtained from China, with a spectral range of 500-2000 cm⁻¹. -1 The excitation wavelength was 785 nm, the integration time was 15 s, the laser power was 160 mW, and the acquisition mode was set to continuous acquisition. The laser beam was focused onto the sample surface through a 10 X lens. To avoid errors during spectral acquisition, three spectra of each sample at different locations were recorded, and the average spectrum of each sample was obtained for further analysis. Finally, Dataset 1 contained 234 Raman spectra, and Dataset 2 contained 90 Raman spectra.

[0074] The serum infrared spectroscopy acquisition equipment was a Fourier transform infrared spectrometer (FTIR-850, Gangdong Scientific, China), with a spectral wavelength range of 700-4000 cm⁻¹. -1 4 cm resolution -1 The scans were performed 32 times under constant temperature and humidity. Before measurement, the samples were thawed at room temperature for 15 min. Then, 5 µl of sample was dropped onto the lens and placed on the spectrometer, where it was allowed to dry for 8 min. Although the spectra were acquired in a relatively enclosed environment, they were still affected by air content. Existing studies show that at 2350 cm⁻¹... -1 The wavelength range near the 2283-2397 cm⁻¹ represents CO₂-related substances. To reduce interference from atmospheric background on the spectrum, the experiment used the 2283-2397 cm⁻¹ band. -1 Feature removal was performed on the spectral bands. To avoid errors during spectral acquisition, three spectra were recorded at different locations for each sample, and the average spectrum of each sample was obtained for further analysis. Finally, Dataset 1 contained 234 infrared spectra, and Dataset 2 contained 90 infrared spectra.

[0075] 3. Spectral data processing

[0076] Because spectral data is easily affected by various factors such as measurement conditions, detection environment, and instrument performance during acquisition, it may lead to problems such as baseline drift and random noise in the spectrum. In Raman spectral preprocessing, iterative adaptive weighted penalized least squares (airPLS) is first used for baseline correction to remove fluorescence background interference. The polynomial degree is 3, the maximum number of iterations is 100, and the gradient threshold is 0.001. Subsequently, the mean smoothing method is used to smooth the spectrum, and the smoothing window width is set to 5. Infrared spectral data only undergoes baseline correction preprocessing. All preprocessing processes are completed in the Matlab R2022a environment. In order to unify the feature scales between different data sources and reduce the impact of numerical range on model training, the min-max normalization method is used to uniformly scale the Raman spectral intensity to the [0, 1] interval, which helps the model run more efficiently and with higher accuracy. 4. Methods

[0077] To achieve high-precision differential diagnosis of cardiovascular diseases, this invention proposes a cardiovascular disease classification system and method based on a multi-source vibrational spectral fusion network with shared and private features, which is a novel multi-source deep learning fusion framework (MVSFNet). Its overall structure is as follows: Figure 1 As shown, this framework extracts private features from infrared (IR) and Raman spectra using parallel multi-scale convolutional neural networks, and generates shared features using a dynamic gating mechanism. A bidirectional attention mechanism is introduced to achieve deep interaction and adaptive fusion between the two sources of private features. Finally, the fused private features and shared features are concatenated and classified. This network framework mainly consists of four core modules: a multi-scale private feature extractor, a dynamic shared feature generation module, a bidirectional attention fusion module, and a classifier.

[0078] Specifically:

[0079] A cardiovascular disease classification system based on a multi-source vibrational spectral fusion network with shared and private features, comprising:

[0080] The data collection module collects Raman and infrared spectral data of serum.

[0081] The data preprocessing module preprocesses the Raman spectral data and infrared spectral data.

[0082] The multi-scale private feature extractor module uses a deep convolutional neural network structure to extract multi-scale private features from the preprocessed Raman spectral data and infrared spectral data.

[0083] The dynamic shared feature generation module adopts a two-level structure of dual-modal projection-dynamic gating fusion to extract the dynamic shared features of the preprocessed Raman spectral data and infrared spectral data.

[0084] The bidirectional attention fusion module performs deep interaction and adaptive fusion on the aforementioned multi-scale private features;

[0085] The classifier module concatenates the fused multi-scale private features with the dynamically shared features and performs classification.

[0086] A cardiovascular disease classification method based on a multi-source vibrational spectral fusion network with shared and private features, employing the aforementioned cardiovascular disease classification system, includes the following steps:

[0087] (1) Collect Raman and infrared spectral data of serum and tears;

[0088] (2) Preprocess the Raman spectral data and infrared spectral data;

[0089] (3) Extract the multi-scale private features of the preprocessed Raman spectral data and infrared spectral data using a multi-scale private feature extractor;

[0090] A two-level dynamic shared feature generation method based on dual-modal projection-dynamic gating fusion is used to extract the dynamic shared features of the preprocessed Raman spectral data and infrared spectral data.

[0091] (4) After performing deep interaction and adaptive fusion on the multi-scale private features, they are concatenated with the dynamic shared features for classification.

[0092] The data collection module, data preprocessing module and steps (1) and (2) in the cardiovascular disease classification method of the cardiovascular disease classification system are as described in “2. Spectral data acquisition” and “3. Spectral data processing”.

[0093] The key points of the cardiovascular disease classification system and methods are as follows:

[0094] (1) Multi-scale private feature extractor

[0095] To fully capture the unique and discriminative information from different spectral modalities, dedicated private feature extractors were designed for infrared (IR) and Raman spectroscopy, respectively. These modules employ a deep convolutional neural network structure, designed to automatically learn modality-specific representations from raw spectral data, thereby providing high-quality base features for subsequent feature fusion and classification.

[0096] Considering the characteristics of its Raman spectral data, this embodiment uses a Raman feature dimension of 312 and designs a parallel multi-scale one-dimensional convolutional neural network (Multi-scale 1D CNN), the structure of which is as follows: Figure 1 The Multi-scale Feature Extractor Raman module is shown in the diagram. This extractor consists of three parallel branches designed to simultaneously capture local to global features across different receptive fields. A one-dimensional convolutional layer with a kernel size of 3 focuses on capturing subtle spectral variations and local patterns of sharp feature peaks at adjacent wavenumbers. A convolutional layer with a kernel size of 5, possessing a slightly larger receptive field, integrates relationships between adjacent feature peaks, capturing a moderate range of spectral contextual information. A convolutional layer with a kernel size of 7, having the largest receptive field, is designed to perceive the overall contours and trends across a wider spectral range. The output feature maps from all three branches are concatenated into a 48-dimensional feature vector (16 channels × 3 branches) after global average pooling. Finally, a fully connected layer (Linear(48, 128)) is used for dimensionality reduction and fusion, outputting a 128-dimensional Raman spectral private feature vector, Raman_p. This multi-scale design ensures that the model can robustly represent both the local details and global morphology of the Raman spectrum simultaneously.

[0097] Infrared spectral data has a higher dimension, 1664 dimensions in this embodiment, and contains more complex spectral overlap information. To address this, a deeper serial-parallel hybrid multiscale convolutional network (MultiScaleCNN_IR) was designed to efficiently process high-dimensional inputs and extract hierarchical features. The extractor contains two main branches, each of which includes downsampling operations to progressively compress the data dimension and expand the receptive field. The first main branch path first uses a convolutional layer with a kernel size of 5 to extract basic features, and then uses a max pooling layer with a stride of 2 (MaxPool1d(2)) to reduce the sequence length to 832, significantly reducing computational complexity while retaining key information. It is followed by a convolutional layer with a kernel size of 3 to further refine higher-level semantic features. To quickly capture the global context, the second main branch path uses a large convolutional kernel with a kernel size of 15, followed by an aggressive pooling layer with a stride of 4 (MaxPool1d(4)), which rapidly reduces the sequence length to 416. Subsequently, a convolutional layer with a kernel size of 5 is used to optimize the coarse-grained features after pooling. The outputs of both main branches are subjected to global average pooling, resulting in 64-dimensional feature vectors. These two vectors are concatenated to obtain a 128-dimensional fused feature, which is then projected onto a 128-dimensional infrared spectral private feature vector ir_p through a fully connected layer (Linear(128, 128)). This design enables the extractor to process high-dimensional data through a deep hierarchical structure and to fuse spectral information of different scales and granularities through a dual-branch structure.

[0098] (2) Dynamic shared feature generation module

[0099] Traditional methods such as fixed weights, static feature concatenation, or weighted averaging are prone to suppressing dominant modal information or causing interference from redundant modal information. The core idea of ​​this module is to abandon traditional methods and instead adopt a data-driven gating mechanism to dynamically and adaptively generate more discriminative shared feature representations, achieving "sample-level" dynamic fusion and providing robust global feature representations for subsequent classification tasks. For example... Figure 1 As shown in the Shared_feature section, the module adopts a two-level structure of "dual-modal projection-dynamic gating fusion". First, the infrared and Raman features are mapped to the same feature space through independent linear projection layers. Then, the sample-level weights are generated through a gating network to perform weighted fusion of the dual-modal projection features, and finally, the dynamic shared features are output.

[0100] Because the original infrared and Raman features have significantly different dimensions, a linear projection layer is needed to map them to a unified shared feature dimension, which is 256 dimensions by default, to eliminate the impact of dimensionality imbalance on fusion. For infrared features... , It is the infrared feature dimension. For batch size, Regarding Raman features , , Dimensions are Raman features. and These are the characteristic projection matrices for infrared and Raman spectroscopy, respectively. , This represents the projected bimodal features.

[0101] The gated network takes the concatenated original bimodal data as input and generates a weight vector with the same dimension as the shared features through a nonlinear transformation. Furthermore, the weight values ​​are constrained to the [0,1] interval by the Sigmoid activation function, thereby achieving dynamic weight allocation "per sample, per feature dimension".

[0102] Gated network input , Then, dynamic weights are generated through a linear layer and a sigmoid activation function. , Weights generated based on gating networks For the projected bimodal features and Weighted fusion is performed to ultimately output dynamically shared features. ,in Indicates elementwise operation, when When an element is close to 1, it indicates that the infrared mode contribution is dominant in that feature dimension; when it is close to 0, the Raman mode contribution is dominant, thus realizing "sample-level adaptive" cross-modal common feature extraction.

[0103] (3) Bidirectional attention fusion module

[0104] This embodiment introduces a bidirectional attention fusion module, a core component designed for the interactive fusion of private features from infrared (IR) and Raman dual-modal spectral data. This module overcomes the limitations of traditional unidirectional attention by achieving dynamic dependency modeling of cross-modal features through a dual-modal interaction query mechanism. It also introduces dynamic gating fusion and residual connections to address the issues of strong heterogeneity and insufficient information complementarity learning in dual-modal features. Ultimately, it outputs fused features with cross-modal collaborative information, providing high-quality feature support for subsequent classification tasks.

[0105] This module constructs symmetrical dual attention branches, using infrared private features as queries and Raman private features as keys / values, and Raman private features as queries and infrared private features as keys / values, respectively, to achieve bidirectional flow of cross-modal information. The attention mechanism employs the classic scaled dot product mechanism, supporting multi-head attention to capture multi-dimensional dependencies. For single-head attention, the attention score is... ,in For a single head dimension, This is used to mitigate the problem of excessively large scores caused by dimensionality growth. Then, Softmax is used to normalize the scores into a probability distribution, ensuring that the sum of the weights is 1. By applying attention weights to the value vector, attention features that focus on key information are obtained. The bidirectional attention features are concatenated as follows: The feature vectors are used to output gate weights through a two-layer network. ,in , Then, weighted fusion is performed to obtain preliminary fusion characteristics.

[0106] To enhance model stability, residual paths and layer normalization are introduced. The mean of the original private features is used as the residual term to preserve basic feature information. Residual addition, Then, layer normalization is performed on the residual addition results to mitigate internal covariate bias. ,in and These are the mean and variance of the feature, respectively. and For learnable scaling and offset parameters, To prevent the denominator from being 0.

[0107] Finally, the merged private and shared features are concatenated and then classified.

[0108] In the above technical solution, a bidirectional query mechanism allows infrared features to focus on key information in Raman features, and vice versa, achieving a synergistic effect. Simultaneously, the gating network can adjust the weights of the dual attention branches based on sample characteristics. Residual connections and layer normalization effectively alleviate the gradient vanishing problem in deep networks.

[0109] 5. Indicators

[0110] To comprehensively evaluate the performance of the proposed model across different tasks, six commonly used binary classification evaluation metrics were employed: Accuracy, Precision, Recall, Specificity, and Area Under the Curve (AUC). These metrics quantify the model's performance from multiple perspectives, including overall prediction accuracy and the precision and recall capabilities of each type of sample. All metrics were calculated independently on the test set, and the results depend on the model's generalization ability on unseen samples. The definitions of each metric are as follows:

[0111] (1) (2)

[0112] (3) (4)

[0113] The ROC curve is plotted with the model's true positive rate (TPR, or Sensitivity) on the vertical axis and the false positive rate (FPR) on the horizontal axis. The AUC (Area Under Curve) represents the area under the curve. The closer the AUC value is to 1, the stronger the model's ability to distinguish between positive and negative samples, and the better its generalization performance.

[0114] B. Results and Analysis

[0115] 1. Spectral analysis

[0116] To further investigate the degree of calcification in AS and the differences between mildly calcified C0 and severely calcified C1 at the molecular level, this embodiment systematically compares and analyzes Raman and infrared spectroscopy. Figure 2 This is the original average image of the infrared and Raman spectra of AS. Figure 3 This is the average image of the infrared and Raman spectra of AS after preprocessing. This embodiment provides a detailed analysis of the infrared and Raman spectra of AS, as follows: Figure 4-5 As shown. The peak matching table is shown in Table 1 (including Tables 1(a) and 1(b)).

[0117] Infrared spectroscopy primarily reflects the vibrational absorption characteristics of chemical bonds in molecules and is sensitive to polar groups. From Figure 4 As can be seen from Table 1(a), in the range of 1250 - 1020 cm -1 The absorption peaks within this range are mainly attributed to CN stretching vibrations, with the functional group being aliphatic amines, 1300-1150 cm⁻¹. -1 It exhibits CH rocking vibration, with the functional group being a haloalkane, and a wavelength range of 1320-1000 cm⁻¹. -1 This is a CO stretching vibration, representing functional groups such as alcohols, carboxylic acids, and esters, reflecting fluctuations in products related to carbohydrate metabolism and fatty acid oxidation. (1550-1475 cm⁻¹)-1 The asymmetric stretching of NO, representing nitro compounds, suggests that differences in absorption intensity may indicate varying levels of activity in inflammation-related metabolites. (1650-1580 cm⁻¹) -1 This represents the NH bending vibration, indicating amines. (1250-1020 cm⁻¹) -1 and 1650-1580 cm -1 Regional differences in uptake indicate alterations in the composition and conformation of protein or polypeptide metabolites. These regional differences reflect variations in the composition of proteins, lipids, and nitrogenous metabolites between the C0 and C1 groups, suggesting that the degree of calcification may be related to changes in the levels of specific metabolites in serum. Severe AS calcification is closely associated with chronic inflammation, oxidative stress, and tissue repair processes. Metabolites released into the bloodstream from these processes constitute the molecular basis for determining the degree of calcification using infrared spectroscopy.

[0118] Raman spectroscopy is sensitive to changes in molecular polarizability, and is particularly suitable for nonpolar or highly symmetric chemical bonds. For example... Figure 5 As shown in Table 1(b), the Raman spectrum at 614 cm⁻¹ -1 and 702 cm -1 The peaks at these locations correspond to the characteristic vibrations of cholesterol esters and cholesterol, respectively, indicating abnormal lipid metabolism, particularly the accumulation of cholesterol and its esterified forms. (840 cm⁻¹) -1 Characteristic band of carbohydrates, 918 cm -1 Attributable to proline, hydroxyproline, glycogen, and lactic acid, 1048 cm -1 and 1112 cm -1 This is the overlapping zone of glycogen and carbohydrates, 1250 cm. -1 It is an amide III band, 1299 cm. -1 and 1317 cm -1 Corresponding to lipids and guanine respectively, 1440 cm -1 The cholesterol and fatty acid band, 1600 cm -1 The band is amide I. The C1 group generally showed higher intensities of cholesterol, lipid, and glycogen-related peaks than the C0 group, suggesting that severe calcification may be related to lipid metabolism disorders and glycogen accumulation. Lipid infiltration is one of the key initiating factors of cardiovascular calcification; cholesterol crystals can activate the NLRP3 inflammasome, promoting calcification progression.

[0119] Infrared spectroscopy is sensitive to polar groups, while Raman spectroscopy is more responsive to nonpolar skeletons. For example, in the infrared spectrum, the 1650-1580 cm⁻¹... -1 N-H bending with Raman at 1600 cm -1 The amide I bands together reflect protein conformational changes. In the infrared spectrum, 1320–1000 cm⁻¹ -1 C–O vibrations and Raman spectroscopy at 840–1112 cm⁻¹-1 The carbohydrate bands together reveal the state of carbohydrate metabolism.

[0120] Table 1 (a) Infrared spectral peak matching table.

[0121]

[0122] Table 1 (b) Raman Spectrum Peak Matching Table

[0123]

[0124] 2. Comparison Model

[0125] To verify the superiority of the proposed Multi-Source Vibrational Spectrum Fusion Network (MVSFNet) in cardiovascular disease diagnosis, this embodiment selects three representative classical models in the current multimodal fusion field as comparison benchmarks: tensor fusion model, bilinear fusion model, and decision-level fusion model. The core design and adaptation characteristics of each model are as follows.

[0126] Tensor Fusion Network: This model originates from the multimodal fusion framework proposed by Zadeh et al. Its core mechanism is to construct a high-dimensional fusion tensor from the single-modal features of infrared and Raman spectra through Kronecker product (outer product) to directly capture the cross-correlation information between modes.

[0127] Bilinear CNN Fusion: This model employs a two-level architecture of "single-modal feature extraction - bilinear interaction". First, a separate one-dimensional CNN is used to extract the basic representations of infrared and Raman spectra, respectively. Then, a bilinear layer is used to perform element-level interaction between the two types of features, strengthening the linear correlation between modalities.

[0128] Decision-Level Fusion: As a typical "post-fusion" paradigm, the core process of this model is to first use two independent multilayer perceptrons (MLPs) to perform single-modal classification of infrared and Raman spectra and output the probability distribution of the categories. Finally, the two probability results are fused by fixed weighted averaging or simple quadratic decision to obtain the final diagnostic conclusion.

[0129] 3. Results

[0130] (1) Classification results

[0131] Table 2 shows the comparison results of MVSFNet with other fusion models. In the AD diagnosis task, the MVSFNet model achieved an accuracy of 95.00%, precision of 95.06%, recall of 95.00%, specificity of 95.09%, and an AUC of 0.9955, significantly outperforming the comparison models. MVSFNet's accuracy is 1.67% higher than the tensor fusion network's accuracy of 93.33%, 5% higher than the bilinear CNN fusion's accuracy of 90%, and 3.33% higher than the decision-level fusion model's accuracy of 91.67%. The comparison models' AUC values ​​in the AD task were 0.9711, 0.9661, and 0.9855, respectively, which are lower than MVSFNet's AUC values ​​in the AD task, by 0.0244, 0.0294, and 0.0100, respectively. The confusion matrix is ​​shown below. Figure 6 As shown in (a), the model demonstrates a significant ability to distinguish between AD and healthy samples, with true positives and true negatives dominating and misclassified samples accounting for an extremely low percentage. Training process ( Figure 6 (b) indicates that the model converges stably, the training and validation loss curves fit closely, and no obvious overfitting is observed. The ROC curve is shown in Figure 1. Figure 6 As shown in (c), the excellent performance of the model is further verified. The AUC value reaches 0.9955, and the curve is close to the upper left corner, indicating that the model has both high sensitivity and high specificity.

[0132] In the AS calcification level classification task, the MVSFNet model achieved an accuracy of 91.30%, precision of 92.55%, recall of 91.30%, specificity of 90.91%, and an AUC of 0.9697, consistently leading across all evaluation metrics. MVSFNet's accuracy was 4.34% higher than Tensor Fusion Network, Bilinear CNN Fusion, and Decision-Level Fusion. Its AUC of 0.9697 was higher than Bilinear CNN Fusion by 0.0005 and Decision-Level Fusion by 0.0159, but slightly lower than Tensor Fusion Network by 0.0014. In the remaining metrics, MVSFNet's precision, recall, and specificity significantly outperformed or approached the best comparative models, demonstrating the most outstanding overall performance. Confusion matrix ( Figure 7 (a) shows that the model can effectively distinguish between mild and severe calcification samples. Although there are a few inter-class misclassifications, the overall classification accuracy is high. Training process ( Figure 7 (b) shows that the loss curve converges smoothly, indicating that the model has good learning stability on this task. ROC such as Figure 7 As shown in (c), MVSFNet achieved an AUC value of 0.9697, confirming that the model maintained excellent diagnostic efficacy in the AS calcification grading task and has reliable clinical translation potential.

[0133] Table 2 Comparison of MVSFNet with other fusion models

[0134]

[0135] (2) Ablation test

[0136] To verify the effectiveness and necessity of each core module in the MVSFNet framework, this embodiment systematically conducted ablation experiments. The experiments evaluated the contribution of each module to model performance on two datasets: aortic dissection (AD) and aortic stenosis (AS), by sequentially removing the No dynamic weight sharing module (NDWS), the No multi-scale feature extraction module (NMFE), and the No Dual Attention Fusion Module (NDAF).

[0137] The results are shown in Table 3. The complete MVSFNet model achieved optimal or near-optimal performance, fully demonstrating the rationality of its architectural design and the effectiveness of inter-module collaboration. Removing the Dynamic Weight Sharing module (MVSFNet-NDWS) led to a significant decrease in model performance. In the AD task, accuracy decreased from 95.00% to 88.33%, precision from 95.06% to 88.57%, recall from 95.00% to 88.33%, specificity from 95.09% to 87.44%, and AUC from 0.9955 to 0.9559. In the AS task, the performance degradation was even more drastic, with accuracy dropping sharply from 91.30% to 78.26%, a decrease of 13.04%, and AUC from 0.9697 to 0.8304. This result indicates that the dynamic weight sharing mechanism is crucial for the model to adaptively learn shared representations between modalities and handle tasks such as AS calcification grading; its absence severely weakens the model's generalization ability and robustness. Removing the multi-scale feature extraction module (MVSFNet-NMFE) significantly negatively impacted both tasks. On the AD dataset, accuracy dropped to 86.67%, a decrease of 8.33%, and the AUC decreased sharply by 0.1082 to 0.8873. On the AS dataset, accuracy remained at 86.96%, but the AUC slightly decreased to 0.9692. This indicates that multi-scale convolutional structures are indispensable for capturing local details and global contextual information at different scales in vibrational spectra, especially in data with complex features and high dimensionality. As shown in the MVSFNet-NDAF results, removing the bidirectional attention module clearly degraded the model's performance on both tasks. In the AD task, accuracy dropped to 91.67%, a decrease of 3.33%, precision dropped to 92.11%, recall dropped to 91.67%, specificity dropped to 91.82%, and the AUC decreased to 0.9410. In the AS task, the performance degradation was even more pronounced, with accuracy decreasing by 8.69 percentage points to 82.61% and AUC decreasing by 0.0887 to 0.8810. This result strongly demonstrates the core value of the bidirectional attention mechanism in modeling the complex interactions between infrared and Raman modalities and achieving accurate feature fusion. Its absence caused a simultaneous decline in all metrics, indicating that this module plays an irreplaceable role in improving the model's overall discriminative performance.

[0138] Taking the AD dataset as an example, the performance of each ablation experiment on the AD task is as follows: Figure 8 , Figure 9 and Figure 10 As shown.

[0139] For the model without the dynamic weight sharing module, its loss curve is as follows: Figure 8As shown in (b), although the training loss and validation loss curves eventually converge, the entire convergence process is more volatile, and the final convergence value is higher than that of the complete model. The stability of the training process is reduced, suggesting that the lack of dynamic weights increases the difficulty of model optimization and weakens the generalization ability.

[0140] After removing the multi-scale feature extraction module, such as Figure 9 As shown in (b), the loss curve exhibits an oscillating and slow-converging trend, confirming that the verification loss is higher than the training loss throughout the entire training period, indicating that the model has a certain risk of overfitting. This is due to the lack of multi-scale perception capability, which causes the model to tend to memorize noise in the training data rather than learn universal features.

[0141] After removing the bidirectional attention fusion module, the convergence trend of the loss curve is closest to that of the complete model, but the final convergence value is slightly higher, such as... Figure 10 As shown in (b), the model can still be effectively optimized, but due to the reduced efficiency of feature interaction fusion, its final performance ceiling is lower than that of the full model. The ROC curve lies between the curves of the full model and the model with dynamic weight sharing removed, with an AUC value of 0.9410. Since the bidirectional attention module acts as a performance enhancement component, it improves the final classification accuracy and robustness of the model by refining feature interactions.

[0142] Table 3 Ablation experiments of MVSFNet on two datasets

[0143]

[0144] (3) Modal comparison

[0145] Table 4 systematically compares the performance of the proposed multi-source fusion model MVSFNet with various single-modal classification methods in the diagnosis of aortic dissection (AD) and aortic stenosis (AS). In the AD diagnosis task, MVSFNet achieved the best overall performance, with an accuracy of 95.00% and an AUC of 0.9955, ranking first among all compared models. In the AS calcification grading task, MVSFNet also outperformed most single-modal models with an accuracy of 91.30% and an AUC of 0.9697. This fully demonstrates that the complementary information provided by fusing infrared (IR) and Raman spectroscopy can significantly improve the model's ability to differentiate and diagnose complex cardiovascular diseases. Multi-source fusion effectively overcomes the limitations of the one-sidedness of single-modal information, enhancing the model's discriminative power through feature complementarity and interaction. In the AD task, the accuracy of the IR- and Raman-based single-modal multi-scale convolutional models (MSC) was 82.61% and 85.00%, respectively, which is about 10-12% lower than that of MVSFNet. On the AS task, the performance decline was more pronounced, with an IR accuracy of 69.57% and a Raman accuracy of 65.22%. Furthermore, both single-modal LSTM models performed poorly on both tasks, particularly in AS graded diagnosis, where LSTM-Raman's accuracy plummeted to 60.87% and its AUC was only 0.5294, almost completely losing its classification ability. Among the single-modal models, ResNet performed best, especially on the AD task, where ResNet-IR and ResNet-Raman achieved accuracies of 94.44% and 93.33%, respectively, with AUCs as high as 0.9933 and 0.9789, very close to the performance of the fusion model MVSFNet. This demonstrates the powerful feature extraction capability of deep residual networks, capable of learning highly discriminative features from a single modality. However, on the more challenging AS task, the performance of the ResNet model declined significantly, with a standalone IR accuracy of 85.71% and Raman accuracy of 73.91%, still lagging behind MVSFNet. Meanwhile, the information value of different modalities varies across different tasks, but in both AD and AS tasks, infrared spectroscopy shows slightly better diagnostic performance than Raman spectroscopy. These results strongly demonstrate the necessity and effectiveness of the multi-source fusion strategy proposed in this paper. MVSFNet, by effectively fusing IR and Raman spectral information, surpasses all single-modal models in overall performance. Even for the ResNet model, which already performs well in single-modal applications, fusion can further improve performance, providing a powerful solution for non-invasive and precise cardiovascular disease auxiliary diagnosis based on vibrational spectroscopy.

[0146] Table 4 Comparison of MVSFNet and unimodal classification results

[0147]

[0148] (4) Generalization

[0149] To verify the generalization ability of MVSFNet, this embodiment applies it to a binary classification diagnostic task for Alzheimer's disease (AD) and severe Asthenia Gravis (SJ) (Type C1). The MVSFNet model exhibits excellent performance, as shown in Table 5. This model achieves 97.37% accuracy, 97.59% precision, 97.37% recall, 98.15% specificity, and an AUC of 0.9832, making it the best performing model overall among all comparison models. (From the confusion matrix...) Figure 11 (a) It can be seen that MVSFNet has a very strong ability to distinguish between AD and severe AS (C1) samples, with true positives and true negatives dominating and very few misclassified samples. The loss curve during the training process is shown in Figure 1. Figure 11 As shown in (b), the model converges smoothly and rapidly, with close matching between the training and validation set losses, indicating no significant overfitting and demonstrating good generalization ability and stability. Compared to other fusion models, MVSFNet maintains high accuracy while also exhibiting an excellent AUC value. Notably, although Bilinear CNN Fusion and Decision-Level Fusion achieve an AUC of 1, their accuracy and recall are still lower than or on par with MVSFNet, indicating that MVSFNet achieves more balanced and reliable performance across multiple metrics. Tensor Fusion Network performs poorly on this task, with an accuracy of only 92.12% and an AUC of 0.8821, significantly lagging behind MVSFNet. In conclusion, MVSFNet demonstrates strong discriminative ability and robustness in the differential diagnosis of AD and severe AS, significantly outperforming traditional fusion methods and proving its significant potential for rapid identification of acute and critical illnesses in real-world clinical scenarios.

[0150] Table 5. Comparison of MVSFNet and other fusion models in binary classification of AD and severe AS (C1).

[0151]

[0152] This invention addresses the challenge of rapid, non-invasive differential diagnosis of aortic dissection (AD) and aortic stenosis (AS) in acute cardiovascular emergencies by successfully designing and validating a novel multi-source vibrational spectral fusion network, MVSFNet. The core innovation of this technical solution lies in the explicit separation and fusion of common and private features from infrared and Raman spectroscopy. Through multi-scale feature extraction, dynamic weight sharing, and a bidirectional attention mechanism, it fully exploits the complementary information and discriminative features between the two modal data. Experimental results show that MVSFNet significantly outperforms traditional tensor fusion, bilinear fusion, decision-level fusion, and various single-modal models in multiple tasks, including AD diagnosis, AS calcification grading, and differential diagnosis between AD and severe AS. All evaluation metrics reach leading levels. This fully demonstrates the absolute advantage of fusing bimodal spectral information over single-modality models, as well as the effectiveness and advancement of the MVSFNet fusion architecture. Systematic ablation studies further confirm that the three core modules—dynamic weight sharing, multi-scale perception, and bidirectional attention—are indispensable for improving the model's discriminative power, robustness, and generalization ability. In summary, MVSFNet overcomes many limitations of existing diagnostic methods, providing a novel, high-precision, non-invasive, objective, and rapid method for the auxiliary diagnosis of cardiovascular diseases. Its superior performance indicates significant clinical translational potential, and it is expected to provide a powerful decision support tool for the rapid screening and risk stratification of patients with acute chest pain in emergency rooms and primary healthcare institutions.

[0153] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Any simple modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention shall still fall within the scope of the technical solution of the present invention.

Claims

1. A cardiovascular disease classification system based on a multi-source vibrational spectral fusion network with shared-private features, characterized in that, include: The data collection module collects Raman and infrared spectral data of serum. The data preprocessing module preprocesses the Raman spectral data and infrared spectral data. The multi-scale private feature extractor module uses a deep convolutional neural network structure to extract multi-scale private features from the preprocessed Raman spectral data and infrared spectral data. The dynamic shared feature generation module adopts a two-level structure of dual-modal projection-dynamic gating fusion to extract the dynamic shared features of the preprocessed Raman spectral data and infrared spectral data. The bidirectional attention fusion module performs deep interaction and adaptive fusion on the aforementioned multi-scale private features; The classifier module concatenates the fused multi-scale private features with the dynamically shared features and performs classification.

2. The cardiovascular disease classification system according to claim 1, characterized in that, The multi-scale private feature extractor module includes a private feature extractor for Raman spectral data and a private feature extractor for infrared spectral data.

3. The cardiovascular disease classification system according to claim 2, characterized in that, The Raman spectral data private feature extractor is a parallel multi-scale one-dimensional convolutional neural network consisting of three parallel branches. The output feature maps of the branches are concatenated into feature vectors after global average pooling, and finally dimensionality reduction and fusion are performed through a fully connected layer to output the Raman spectral private feature vector. The private feature extractor for the infrared spectral data is a serial-parallel hybrid multi-scale convolutional network containing two main branches. The outputs of the two main branches are both subjected to global average pooling to obtain two feature vectors, which are then concatenated to obtain fused features. Finally, the feature vectors are projected onto the infrared spectral private feature vectors through a fully connected layer.

4. The cardiovascular disease classification system according to claim 3, characterized in that, In the private feature extractor of the Raman spectral data, the branch kernel sizes are 3, 5, and 7, respectively.

5. The cardiovascular disease classification system according to claim 3, characterized in that, In the private feature extractor of the infrared spectral data, the first main branch path first uses a convolutional layer with a kernel size of 5 to extract basic features, then reduces the sequence length through a max pooling layer; and then refines semantic features through a convolutional layer with a kernel size of 3. The second main branch first uses a large convolutional kernel with a kernel size of 15, then reduces the sequence length through an aggressive pooling layer; then it optimizes the coarse-grained features after pooling through a convolutional layer with a kernel size of 5.

6. The cardiovascular disease classification system according to claim 1, characterized in that, The dynamic shared feature generation module maps infrared and Raman features to a feature space of the same dimension through an independent linear projection layer, generates sample-level weights through a gating network, and performs weighted fusion of the dual-modal projection features to finally output dynamic shared features.

7. The cardiovascular disease classification system according to claim 1, characterized in that, In the aforementioned bidirectional attention fusion module, the infrared private features are used as queries and the Raman private features are used as keys. The fusion process is as follows: Attention employs a scaled dot product mechanism, supporting multi-head attention to capture multi-dimensional dependencies. For single-head attention, the attention score is... ,in It is a single-head dimension; Then, the scores are normalized to a probability distribution using Softmax, ensuring that the sum of the weights is 1. Applying attention weights to the value vector yields attention features that focus on key information. ; Concatenate the bidirectional attention features into The feature vectors are used to output gate weights through a two-layer network. ,in , ; Then, a weighted fusion process is performed.

8. The cardiovascular disease classification system according to claim 7, characterized in that, The bidirectional attention fusion module employs residual path and layer normalization processing. The process is as follows: the mean of the original private features is used as the residual term. After retaining the basic feature information, the result of the residual addition is subjected to layer normalization, and the formula is: ,in and These are the mean and variance of the feature, respectively. and For learnable scaling and offset parameters, To prevent the denominator from being 0.

9. A cardiovascular disease classification method based on a multi-source vibrational spectral fusion network with shared-private features, characterized in that, The cardiovascular disease classification system according to any one of claims 1-8 includes the following steps: (1) Collect Raman and infrared spectral data of serum and tears; (2) Preprocess the Raman spectral data and infrared spectral data; (3) Extract the multi-scale private features of the preprocessed Raman spectral data and infrared spectral data using a multi-scale private feature extractor; A two-level dynamic shared feature generation method based on dual-modal projection-dynamic gating fusion is used to extract the dynamic shared features of the preprocessed Raman spectral data and infrared spectral data. (4) After performing deep interaction and adaptive fusion on the multi-scale private features, they are concatenated with the dynamic shared features for classification.

10. The application of the cardiovascular disease classification system according to any one of claims 1-8, or the cardiovascular disease classification method according to claim 9, in a cardiovascular disease auxiliary diagnostic device.

Images