Aortic dissection risk prediction method

Abstract

The embodiment of the invention provides an aortic dissection risk prediction method, and belongs to the technical field of medical care informatics, and the method specifically comprises the steps: obtaining the Fourier transform infrared spectrum, Raman spectrum and clinical data of a serum sample of a to-be-detected object; preprocessing the Fourier transform infrared spectrum and the Raman spectrum, and standardizing clinical data; respectively extracting multi-scale features of the infrared spectrum and the Raman spectrum through a multi-scale feature extraction module; inputting the infrared and Raman spectrum features of the same scale into a cross-modal fusion module to obtain enhanced cross-modal features; performing information interaction through a multi-scale interaction module to obtain unified feature representation; and outputting a disease risk prediction result or a prognosis risk prediction result by using a classification and discrimination module based on the unified feature representation and the clinical data. According to the scheme, the stability and accuracy of disease risk prediction and prognosis risk prediction are improved.

Description

Technical Field

[0001] This invention relates to the field of healthcare informatics technology, and in particular to a method for predicting the risk of aortic dissection. Background Technology

[0002] Currently, the detection of aortic dissection mainly relies on imaging techniques such as computed tomography angiography, magnetic resonance imaging, and transesophageal ultrasound, while the diagnosis of myocardial infarction mainly relies on electrocardiography and the detection of biochemical markers such as cardiac troponin. All of these detection techniques have certain limitations in practical application, such as high dependence on equipment conditions, long examination times, high requirements for patient cooperation, and the lag in the dynamic changes of some biochemical markers.

[0003] It is evident that there is an urgent need for a method that can predict the risk and prognosis of aortic dissection with high efficiency, accuracy, and adaptability. Summary of the Invention

[0004] In view of this, embodiments of the present invention provide a method for predicting the risk of aortic dissection, which at least partially solves the problems of poor prediction efficiency, accuracy and adaptability in the prior art.

[0005] This invention provides a method for predicting the risk of aortic dissection, comprising: Step 1: Obtain the Fourier transform infrared spectrum and Raman spectrum of the serum sample of the subject to be tested, and obtain the clinical indicator data of the subject to be tested; Step 2: Preprocess the Fourier transform infrared spectrum and Raman spectrum respectively, and standardize the clinical indicator data. Step 3: Input the preprocessed Fourier transform infrared spectrum and Raman spectrum into the pre-trained multi-scale feature extraction module to extract the multi-scale feature set corresponding to the Fourier transform infrared spectrum. and the multi-scale feature set corresponding to Raman spectra. ,in, , These represent the high-resolution features of infrared and Raman spectra, respectively. , They represent medium-resolution features, , These represent low-resolution features; Step 4: Input the infrared spectral features and Raman spectral features of the same scale into the cross-modal feature fusion module for fusion to obtain the enhanced infrared spectral cross-modal feature sets. Raman spectroscopy cross-modal feature set ; Step 5: Input the enhanced infrared spectral cross-modal feature set and Raman spectral cross-modal feature set into the multi-scale interaction module to perform bidirectional information interaction from low resolution to high resolution and from high resolution to low resolution, and obtain the unified feature representation of infrared spectrum and Raman spectrum. Step 6: Based on the unified feature representation of infrared spectroscopy and Raman spectroscopy, and combined with the standardized clinical indicator data, output the disease risk prediction result or prognosis risk prediction result of the subject to be tested through the classification and discrimination module.

[0006] According to a specific implementation of an embodiment of the present invention, the step of preprocessing the Fourier transform infrared spectrum and the Raman spectrum respectively includes: Delete data features within a preset band in the Fourier transform infrared spectrum; The airPLS algorithm is used for baseline correction of the Raman spectrum, and a smoothing algorithm is used for noise reduction of the Raman spectrum.

[0007] According to a specific implementation of an embodiment of the present invention, the multi-scale feature extraction module is a three-branch parallel one-dimensional convolutional neural network; Each convolutional branch contains four one-dimensional convolutional layers, with a kernel size of 7. Different branches extract features at different scales by setting different convolutional strides and output channel numbers. The first branch has a convolution stride of 2 and an output channel of 8, which is used to extract high-resolution local detail features; The second branch has a convolution stride of 4 and 16 output channels, and is used to extract medium-resolution spectral peak structure features. The third branch has a convolution stride of 8 and 32 output channels, and is used to extract low-resolution global trend features.

[0008] According to a specific implementation of an embodiment of the present invention, step 4 specifically includes: Step 4.1, for any scale Using the Raman spectral features at this scale as the query vector and the infrared spectral features at this scale as the key vector and value vector, the first attention calculation is performed. Step 4.2: Using the infrared spectral features at this scale as the query vector and the Raman spectral features at this scale as the key vector and value vector, perform the second attention calculation; Step 4.3: The results of the first attention calculation and the second attention calculation are weighted and fused with the original Raman spectral features and infrared spectral features respectively through an adaptive residual fusion structure and fusion weights. The fusion weights are... , These are learnable parameters; Step 4.4: Perform nonlinear enhancement processing on the weighted fused features to obtain the enhanced infrared spectral cross-modal features and Raman spectral cross-modal features at this scale; Step 4.5: Form an enhanced infrared spectral cross-modal feature set based on the enhanced infrared spectral cross-modal features and Raman spectral cross-modal features at each scale. Raman spectroscopy cross-modal feature set .

[0009] According to a specific implementation of an embodiment of the present invention, step 5 specifically includes: Step 5.1, the first stage is the low-resolution to high-resolution interaction: after dimensional alignment of the low-resolution features, cross-scale attention interaction is performed with the adjacent higher-resolution features, while the lowest resolution features remain unchanged. Step 5.2, the second stage is the interaction from high resolution to low resolution: after dimensional alignment of the high resolution features, they are fused with the adjacent lower resolution features, while the highest resolution features remain unchanged. Step 5.3: After the two-stage interaction, the cross-modal features of the Raman spectrum at the three scales are spliced ​​together according to the channel dimension to obtain a unified feature representation of the Raman spectrum. Similarly, the cross-modal features of the infrared spectrum at the three scales are spliced ​​together according to the channel dimension to obtain a unified feature representation of the infrared spectrum.

[0010] According to a specific implementation of an embodiment of the present invention, when performing disease risk prediction, step 6 specifically includes: the classification and discrimination module fuses the unified feature representation of infrared spectroscopy and the unified feature representation of Raman spectroscopy, inputs it into the multilayer perceptron classifier, and outputs the risk probability of the subject belonging to the aortic dissection, myocardial infarction or healthy control category as the disease risk prediction result.

[0011] According to a specific implementation of an embodiment of the present invention, when performing disease risk prediction, step 6 specifically includes: The standardized clinical indicator data is mapped into clinical feature vectors through an embedding layer; The clinical feature vector, the unified feature representation of infrared spectroscopy, and the unified feature representation of Raman spectroscopy are fused across modes to obtain fused multimodal features; The multimodal features are fused and input into a multilayer perceptron classifier, and the output prognostic risk value of the object to be tested is used as the prognostic risk prediction result.

[0012] The aortic dissection risk and prognosis prediction scheme in this embodiment of the invention includes: Step 1, acquiring the Fourier transform infrared spectrum and Raman spectrum of the serum sample of the subject to be tested, and acquiring the clinical indicator data of the subject to be tested; Step 2, preprocessing the Fourier transform infrared spectrum and Raman spectrum respectively, and standardizing the clinical indicator data; Step 3, inputting the preprocessed Fourier transform infrared spectrum and Raman spectrum into a pre-trained multi-scale feature extraction module to extract the multi-scale feature set corresponding to the Fourier transform infrared spectrum. and the multi-scale feature set corresponding to Raman spectra. ,in, , These represent the high-resolution features of infrared and Raman spectra, respectively. , They represent medium-resolution features, , These represent low-resolution features respectively; Step 4: Input the infrared spectral features and Raman spectral features of the same scale into the cross-modal feature fusion module for fusion, and obtain the enhanced infrared spectral cross-modal feature sets respectively. Raman spectroscopy cross-modal feature set Step 5: Input the enhanced infrared spectral cross-modal feature set and Raman spectral cross-modal feature set into the multi-scale interaction module to perform bidirectional information interaction from low resolution to high resolution and from high resolution to low resolution, and obtain the unified feature representation of infrared spectrum and Raman spectrum; Step 6: Based on the unified feature representation of infrared spectrum and Raman spectrum, combined with the standardized clinical indicator data, output the disease risk prediction result or prognostic risk prediction result of the subject to be tested through the classification and discrimination module.

[0013] The beneficial effects of the embodiments of the present invention are as follows: By constructing a multi-modal vibrational spectroscopy analysis framework with multi-module collaboration, the present invention achieves multi-scale feature extraction, cross-modal semantic alignment, and cross-scale structural information interaction between Raman spectroscopy and infrared spectroscopy, which can effectively improve the stability and accuracy of disease risk prediction and prognosis risk prediction. Attached Figure Description

[0014] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the embodiments will be 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.

[0015] Figure 1 A flowchart illustrating a method for predicting the risk of aortic dissection provided in an embodiment of the present invention; Figure 2 A schematic diagram of the algorithm flow for aortic dissection risk prediction method provided in an embodiment of the present invention; Figure 3 A cross-modal feature fusion module diagram of a same-scale matching strategy provided in an embodiment of the present invention; Figure 4 This is a diagram of a two-stage, multi-scale interactive module provided in an embodiment of the present invention. Detailed Implementation

[0016] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0017] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, in the absence of conflict, the following embodiments and features in the embodiments can be combined with each other. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0018] It should be noted that various aspects of embodiments within the scope of the appended claims are described below. It will be apparent that the aspects described herein can be embodied in a wide variety of forms, and any particular structure and / or function described herein is merely illustrative. Based on this invention, those skilled in the art will understand that one aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number of aspects set forth herein can be used to implement the device and / or practice the method. Additionally, this device and / or method can be implemented using structures and / or functionalities other than one or more of the aspects set forth herein.

[0019] It should also be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. The illustrations only show the components related to the present invention and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0020] Furthermore, specific details are provided in the following description to facilitate a thorough understanding of the examples. However, those skilled in the art will understand that the described aspects can be practiced without these specific details.

[0021] This invention provides a method for predicting the risk of aortic dissection, which can be applied to the aortic dissection risk warning process in medical settings.

[0022] See Figure 1 This is a flowchart illustrating a method for predicting the risk of aortic dissection provided in an embodiment of the present invention. Figure 1 and Figure 2 As shown, the method mainly includes the following steps: Step 1: Obtain the Fourier transform infrared spectrum and Raman spectrum of the serum sample of the subject to be tested, and obtain the clinical indicator data of the subject to be tested; This invention provides a disease intelligent risk identification method based on multimodal vibrational spectroscopy, aiming to solve the problems of insufficient diagnostic capability of single spectral modes, low utilization rate of multi-scale resolution features, and insufficient cross-modal data fusion in existing technologies. The method, multi-scale feature extraction module, cross-modal feature fusion module, multi-scale interaction module, and classification and discrimination module of this invention work synergistically to achieve deep integrated expression of Raman and infrared spectroscopy and high-precision disease risk identification.

[0023] In practice, blood samples can be obtained from the subjects to be tested. Immediately after collection, all blood samples are centrifuged at 4 ℃ and 4000 r / min to separate the serum layer, which is then transferred to sterile EP tubes. The samples were stored in a refrigerator for FTIR and Raman spectroscopy acquisition. Before spectral measurements, all serum samples were thawed at room temperature (22–25 °C) for 15 min to ensure stable sample temperature and homogeneous composition, reducing spectral deviations caused by phase changes. FTIR spectra were then acquired using a Fourier transform infrared spectrometer, and Raman spectra were obtained using a high-resolution confocal Raman spectrometer. A multimodal dataset was constructed by combining this data with corresponding clinical indicators.

[0024] Step 2: Preprocess the Fourier transform infrared spectrum and Raman spectrum respectively, and standardize the clinical indicator data. In practice, data features within a preset band in the Fourier transform infrared spectrum can be deleted. At the same time, the airPLS algorithm is used to perform baseline correction on the Raman spectrum, and a smoothing algorithm is used to denoise the Raman spectrum.

[0025] Step 3: Input the preprocessed Fourier transform infrared spectrum and Raman spectrum into the pre-trained multi-scale feature extraction module to extract the multi-scale feature set corresponding to the Fourier transform infrared spectrum. and the multi-scale feature set corresponding to Raman spectra. ,in, , These represent the high-resolution features of infrared and Raman spectra, respectively. , They represent medium-resolution features, , These represent low-resolution features; In practical implementation, for multi-scale feature modeling, this invention proposes a three-branch parallel structure one-dimensional convolutional feature extraction network. This module uses an L2S (Large-to-Small) frequency attention path to abstract large-scale structures (molecular overall skeleton vibrations) and pass semantic guidance to small-scale local peaks; an S2L (Small-to-Large) frequency attention path to inversely enhance global structural information using local peak features (such as functional group fingerprint regions); bidirectional multi-head cross-attention for cross-scale feature flow; and a parallel structure of convolution and linear mapping for scale alignment and frequency weighting. This enables frequency segmentation modeling of FTIR infrared and Raman spectral features and allows for complementary enhancement of spectral features at different scales.

[0026] The input to this convolutional feature extraction network is a spectral sequence that has undergone baseline correction, denoising, and normalization. This network comprises three convolutional branches with different strides and channel widths, each consisting of four convolutional layers with a uniform kernel size of 7. It achieves multi-scale decoupling in frequency space through varying strides (S) and channel counts (C): the first branch, with a stride of S=2 and channel count C=8, captures high-resolution local details; the second branch, with S=4 and C=16, extracts mid-resolution spectral peak structure variations; and the third branch, with S=8 and C=32, models low-resolution global trends and slow-changing patterns. Each convolutional branch consists of four convolutional layers, followed by batch normalization and non-linear activation functions to improve network training stability and feature generalization ability. This structure was applied to Raman and infrared spectroscopy, yielding feature sets at three different resolutions. and This provides a semantic expression with scale-differentiation for subsequent fusion.

[0027] Alternatively, other multi-scale modeling methods include: 1. Multi-scale spectral decomposition based on wavelet transform: Continuous wavelet transform (CWT) or discrete wavelet transform (DWT) can be used to decompose spectral signals into different frequency bands, and then the features of each frequency band can be fused. However, this method relies on manually designing wavelet bases and is difficult to adaptively learn the frequency structure like convolutional networks.

[0028] 2. Pyramid pooling: Multi-scale representations can be generated through different pooling windows and then fused. However, pooling lacks the learnability of convolution and is not suitable for capturing complex vibration peak shapes.

[0029] Step 4: Input the infrared spectral features and Raman spectral features of the same scale into the cross-modal feature fusion module for fusion to obtain the enhanced infrared spectral cross-modal feature sets. Raman spectroscopy cross-modal feature set ; In specific implementation, regarding cross-modal fusion, this invention constructs a cross-modal feature fusion module (Cross-modal Attention Fusion, CAF) based on a same-scale matching strategy, such as... Figure 3 As shown, this module achieves cross-modal information complementarity by pairing two spectral modes at the same scale. Specifically, the query vector is generated from FTIR or Raman features, and the key and value come from another mode. The inter-modal dependency strength is controlled by an adjustable residual fusion coefficient. The fused features not only serve as input to subsequent classification or multi-scale interaction modules, but also enhance the original modal features through a feature feedback mechanism. That is, the fused features are fed back to the original modal channels in a certain proportion, and the expressive power of the original features is improved through weighted superposition, normalization, or nonlinear mapping. This ensures that the cross-modal complementary information not only affects the output features, but also strengthens the key frequency band signals within the original mode. Subsequently, the fused features are redistributed through MLP enhancement blocks and frequency attention weights to achieve nonlinear feature recombination between spectra and supplementation of molecular information.

[0030] This module establishes separate fusion paths for high, medium, and low scales. For features at any scale i, this invention first constructs a bidirectional frequency-guided attention mechanism. By using Raman features as the query vector and infrared features as the key vector, and infrared features as the query vector and Raman features as the key vector, cross-modal attention operations are performed to achieve information complementarity and frequency band correspondence between the two modalities at the same scale. Subsequently, this invention designs an adaptive residual fusion structure, using weighted coefficients... , By controlling the injection ratio of cross-modal information to original information, the stability of the fusion result is adjusted. Furthermore, this module performs a secondary mapping on the fusion result using an MLP nonlinear enhancement structure based on layer normalization and a two-layer feedforward network, thereby obtaining the enhanced cross-modal feature set. and This effectively improves the semantic consistency and feature recognition capabilities between modalities.

[0031] Alternatively, the following methods can also be used to fuse information at the same scale across modalities: 1. The simple splicing or weighted summation method of this invention: The most basic method is to directly splice or linearly weight the spectral features of Raman and FTIR infrared spectra of this invention. However, it cannot achieve fine-grained correspondence between frequency bands, nor can it obtain bidirectional information guidance.

[0032] 2. This invention is based on co-attention fusion: it can draw on the co-attention mechanism in image-text retrieval to simultaneously calculate attention for both modalities. However, unlike this invention, such methods do not include the adaptive control of the residual balance factor, which is prone to generating modal bias in high-noise spectra.

[0033] 3. This invention uses a cross-modal Transformer-based token exchange mechanism: the Raman and FTIR infrared spectra of this invention are treated as token sequences and interacted through a Cross-Transformer. However, this method has high computational costs, and one-dimensional spectra are not suitable for excessive discretization.

[0034] Alternative methods for feature reflow after fusion can include: 1. Gated cyclic reflux of the present invention: but different from the present invention and Compared to the dual-path residual mechanism, the control precision is lower.

[0035] 2. The fusion result of this invention is used as an additional input condition: This method can input the fused features into subsequent networks (such as FiLM and AdaIN of this invention). However, this method does not replace the backbone features and cannot achieve "structural-level fusion".

[0036] 3. The fusion results of this invention are used for dynamic convolution weight adjustment: the fused features can be used to generate dynamic convolution kernels, thereby changing the subsequent feature extraction method. However, this is complex to implement and requires higher training stability.

[0037] Step 5: Input the enhanced infrared spectral cross-modal feature set and Raman spectral cross-modal feature set into the multi-scale interaction module to perform bidirectional information interaction from low resolution to high resolution and from high resolution to low resolution, and obtain the unified feature representation of infrared spectrum and Raman spectrum. In specific implementation, such as Figure 4 As shown, in terms of multi-scale fusion, this invention proposes a two-stage multi-scale interaction module to model the information flow relationship between different resolutions, thereby improving the comprehensiveness of feature representation. The first stage is low-to-high (L2H) interaction, which involves progressively passing abstract semantics from low resolution to high resolution. Specifically, the low-resolution features at scale i+1 are used to model the information flow relationship between different resolutions, thereby improving the comprehensiveness of feature representation. The first stage is a low-to-high interaction process, which involves progressively passing abstract semantics from low resolution to high resolution. In this process, the low-resolution features are dimensionally aligned and then interact with the high-resolution features across scales, with the lowest layer of the low-resolution features remaining unchanged as the information starting point. The second stage is a high-to-low interaction process, which involves injecting structural detail information from high resolution to low resolution. The high-resolution features are fused with the low-resolution features through alignment operations, while the highest resolution feature remains unchanged at the baseline layer. Finally, this invention concatenates the processed Raman features of the three scales according to the channel dimension to obtain a unified feature representation, and obtains an infrared feature representation in the same way. Both are used as input to the classification module to achieve risk prediction of disease categories or patient prognostic status.

[0038] Alternatively, alternative solutions to cross-scale interaction methods may include: 1. U-Net-style upsampling and skip-connection structure: Upsampling is performed using transposed convolution or interpolation, and then concatenated with high-resolution features. However, this structure focuses more on spatial reconstruction than on the semantic association of spectral frequency structures.

[0039] 2. Unidirectional multi-scale fusion: Similar to feature pyramid networks, it transmits information unidirectionally from high to low or from low to high. However, it lacks the bidirectional cross-scale information compensation capability emphasized in this invention.

[0040] Step 6: Based on the unified feature representation of infrared spectroscopy and Raman spectroscopy, and combined with the standardized clinical indicator data, output the disease risk prediction result or prognosis risk prediction result of the subject to be tested through the classification and discrimination module.

[0041] In its implementation, this invention employs a multilayer perceptron (MLP) structure as the core discriminator in the classification module. It refines and enhances the discrimination of the input fused features through layer-by-layer nonlinear mapping. Each perceptron layer consists of fully connected neurons, batch normalization, and activation functions to ensure stable training and prevent overfitting. For three-category disease risk identification tasks (such as aortic dissection, myocardial infarction, and healthy controls), the final output layer uses a softmax activation function to map the fused features to predicted probabilities for each category. The system can determine the category based on the maximum probability value and simultaneously provide the confidence score of each sample in each category, offering a reference for clinical decision-making.

[0042] In patient prognosis prediction, this invention constructs a binary classification prediction system for survival status based on a multimodal deep learning model for patients with aortic dissection, to assist in clinical risk assessment and individualized treatment plan formulation. Specifically, the model introduces the patient's clinical information (including demographic characteristics, medical history, blood and biochemical indicators) as a third modality, and performs joint modeling with Raman and infrared spectral features in a fusion layer, thereby achieving deep integration of molecular-level information and individual pathophysiological information. The model first performs multi-scale feature extraction for each of the three modalities. For Raman and infrared spectroscopy, a three-branch parallel convolutional network is used to progressively extract multi-scale resolution features from high-resolution local details to low-resolution global trends; for the clinical indicator modality, standardization and vectorization are performed according to feature type and data distribution, and then mapped to a feature space matching the spectral features through convolutional or fully connected networks to achieve alignment between modalities. In terms of cross-modal fusion, this invention adopts a phased strategy: The first phase involves pairwise fusion of Raman and infrared spectra, and infrared and clinical indicator modalities, respectively, to fully explore the spectral complementarity between spectral modalities and the representational correlation between spectra and clinical indicators, achieving preliminary modal feature synergy. The second phase fusion module further integrates the aforementioned preliminary fusion features, establishing deep semantic synergy relationships between multimodal features, and improving the semantic consistency and expressive integrity of the fused features through feature weighting and nonlinear enhancement strategies. Subsequently, the fused high-dimensional features are input into a multi-scale interaction module. This module, through low-to-high and high-to-low interaction mechanisms, achieves contextual information completion and local detail enhancement at different resolutions, thereby obtaining a final multimodal feature representation with comprehensive discriminative capabilities. Finally, the multimodal fusion features are input into a multilayer perceptron classifier to complete a binary classification prediction of the survival status of aortic dissection patients, outputting the risk prediction result for survival or death and the corresponding confidence level.

[0043] During the model training phase, five-fold cross-validation, cross-entropy loss, AdamW optimizer, learning rate of 0.001, 200 training epochs, and batch size of 256 were used to ensure model stability and generalization ability.

[0044] The aortic dissection risk and prognostic risk prediction method provided in this embodiment, by constructing a multi-module collaborative multimodal vibrational spectroscopy analysis framework, achieves multi-scale feature extraction, cross-modal semantic alignment, and cross-scale structural information interaction between Raman and infrared spectroscopy, effectively improving the stability and accuracy of disease risk identification. Specifically, this application has the following features and technical effects: This application first introduces a three-branch parallel convolutional structure in spectral feature modeling. Each branch uses a one-dimensional convolutional kernel with a stride S={2,4,8} and a channel number C={8,16,32} (the kernel size is fixed at 7). This allows for the simultaneous extraction of local high-resolution features, mesoscale structural changes, and large-scale global trend information within the same network. Compared to traditional spectral feature extraction methods with a single convolutional scale or fixed sampling stride, this structure avoids the problems of high-frequency information loss or insufficient characterization of low-frequency trends, enabling the model to have higher sensitivity and expression range in disease scenarios with subtle differences in molecular vibrational features.

[0045] Secondly, the cross-modal feature fusion module (CFB) proposed in this application introduces a bidirectional frequency-guided attention mechanism between features at the same scale, making the Raman mode and infrared mode the query and key of each other, respectively. Explicit modeling of the correspondence and complementary information between frequency bands is achieved through attention calculation. Compared with traditional simple splicing or weighted summation modal fusion methods, this application can automatically identify corresponding regions of frequency distribution in two spectral modes at each scale, thereby accurately capturing the response differences of the two types of spectra in different molecular groups and improving the alignment quality of cross-modal information. Furthermore, the adaptive residual fusion structure (… , Weighted fusion and MLP nonlinear enhancement further ensure that the cross-modal fusion results maintain the stability of the original distribution while introducing external information, and reduce the performance degradation caused by cross-modal noise interference.

[0046] Furthermore, the dual-stage, multi-scale interaction module designed in this application enables bidirectional information flow between different scales. The low-to-high (L2H) path aligns channel dimensions through linear transformation, allowing low-resolution abstract semantics to supplement the high-resolution feature layer; the high-to-low (H2L) path aligns features through max pooling, allowing high-resolution local details to compensate for the overall trend of low resolution. This bidirectional interaction mechanism breaks through the structural limitations of traditional unidirectional information propagation relying solely on bottom-up or top-down approaches, enabling the model to achieve cross-scale structural feature enhancement while maintaining multi-scale spatial consistency, thereby significantly improving the joint representation capability of spectral signals at different frequency levels.

[0047] Through the synergistic effect of multi-scale convolution, cross-modal attention, and two-stage interactive fusion, detailed modeling of spectral signals can be achieved while maintaining controllable network parameters, demonstrating higher discriminative ability and stability in disease classification tasks. Furthermore, the model structures used in this application are all based on one-dimensional convolution and attention mechanisms, applicable to spectral data of different sampling rates and lengths, exhibiting good adaptability and universality. This facilitates deployment in rapid clinical detection scenarios, enabling rapid and non-invasive risk prediction for acute diseases such as aortic dissection and myocardial infarction.

[0048] The method of the present invention will be further described below with reference to a specific embodiment. In this embodiment, serum from aortic dissection patients in clinical practice is used as the spectral acquisition material, and a multimodal dataset is constructed by combining the corresponding clinical information. All blood samples were centrifuged immediately after collection at 4 ℃ and 4000 r / min to separate the serum layer, and then transferred to sterile EP tubes and stored at -80 ℃ for FTIR and Raman spectral acquisition. Before spectral measurement, all serum samples were allowed to thaw at room temperature (22–25 ℃) for 15 min to ensure sample temperature stability and component homogeneity, reducing spectral deviations caused by phase changes.

[0049] II. FTIR Infrared Spectral Acquisition Implementation Process In this embodiment, a Fourier transform infrared spectrometer (model FTIR-850, Guangdong Scientific) was used to acquire FTIR infrared spectra. The specific steps are as follows.

[0050] 1. Spectral acquisition parameter settings: wavenumber range 700–4000 cm⁻¹ -1 Resolution 4 cm -1 ;Scanned 32 times cumulatively.

[0051] 2. Sample loading method: Add 5 µL of the thawed serum sample to the lens of the spectrometer (ZnSe or...). The sample is placed in the central region of the substrate and then placed in the sample chamber of the spectrometer and left to stand for 8 minutes to allow the sample to dry completely and avoid interference from water peak absorption.

[0052] 3. Measurement Process and Air Background Removal: Although the spectral acquisition is conducted in a closed space, it is still affected by the absorption of CO2 in the air. To reduce background interference, this invention removes the 2277.54–2408.68 cm⁻¹ range during the processing stage. -1 The characteristic range of this band includes CO2 in the range of ~2350 cm⁻¹. -1 The characteristic absorption peak.

[0053] 4. Repeated measurements and average spectrum calculation: Three sets of spectra were collected from different locations for each sample. The average value was taken to reduce the error caused by spatial inhomogeneity and obtain a highly stable final FTIR infrared spectrum.

[0054] III. Raman Spectroscopy Acquisition Procedure In this embodiment, a high-resolution confocal Raman spectrometer (LabRAM HR Evolution, gora Ramanspectroscopy, ideaoptics) was used to acquire serum Raman spectra. The acquisition process is as follows.

[0055] 1. Instrument parameter settings: spectral range 500–2000 cm⁻¹ -1 Excitation wavelength 785 nm; laser power 160 mW; integration time 15 s; continuous acquisition mode; 10× confocal microscope objective (NA≈0.25).

[0056] 2. Sample preparation and measurement: After thawing the serum sample, a small amount is directly dropped onto the sample slide, and the laser beam is focused on the sample surface through the objective lens to ensure stable Raman signal acquisition intensity.

[0057] 3. Repeated acquisition: For each sample, the spectrum is acquired 3 times at different locations to reduce location-dependent errors, and the three spectra are averaged to obtain the final Raman spectrum.

[0058] 4. Spectral preprocessing: airPLS baseline correction removes fluorescence background; Smooth algorithm smooths and reduces random noise.

[0059] IV. Clinical Data Collection Implementation Process All clinical information comes from the hospital's electronic medical records and LIS / HIS systems, including demographic information, past medical history, and biochemical and hematological indicators. The collected patient data undergoes cleaning to remove records containing severely missing or outlier values. Feature cleaning and filtering cover three dimensions: demographic characteristics, underlying medical history, and biochemical and hematological indicators.

[0060] V. Implementation Process of the Invention Method After acquiring FTIR infrared spectroscopy, Raman spectroscopy, and clinical data, these three types of data are input into the multimodal deep learning network proposed in this invention, including a multi-scale frequency modeling module, a cross-modal attention fusion module, and a clinical feature embedding module. This process ultimately achieves risk prediction and prognostic risk prediction for aortic dissection.

[0061] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof.

[0062] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for predicting the risk of aortic dissection, characterized in that, include: Step 1: Obtain the Fourier transform infrared spectrum and Raman spectrum of the serum sample of the subject to be tested, and obtain the clinical indicator data of the subject to be tested; Step 2: Preprocess the Fourier transform infrared spectrum and Raman spectrum respectively, and standardize the clinical indicator data. Step 3: Input the preprocessed Fourier transform infrared spectrum and Raman spectrum into the pre-trained multi-scale feature extraction module to extract the multi-scale feature set corresponding to the Fourier transform infrared spectrum. and the multi-scale feature set corresponding to Raman spectra. ,in, , These represent the high-resolution features of infrared and Raman spectra, respectively. , They represent medium-resolution features, , These represent low-resolution features; Step 4: Input the infrared spectral features and Raman spectral features of the same scale into the cross-modal feature fusion module for fusion to obtain the enhanced infrared spectral cross-modal feature sets. Raman spectroscopy cross-modal feature set ; Step 5: Input the enhanced infrared spectral cross-modal feature set and Raman spectral cross-modal feature set into the multi-scale interaction module to perform bidirectional information interaction from low resolution to high resolution and from high resolution to low resolution, and obtain the unified feature representation of infrared spectrum and Raman spectrum. Step 6: Based on the unified feature representation of infrared spectroscopy and Raman spectroscopy, and combined with the standardized clinical indicator data, output the disease risk prediction result or prognosis risk prediction result of the subject to be tested through the classification and discrimination module.

2. The method according to claim 1, characterized in that, The steps of preprocessing the Fourier transform infrared spectrum and the Raman spectrum respectively include: Delete data features within a preset band in the Fourier transform infrared spectrum; The airPLS algorithm is used for baseline correction of the Raman spectrum, and a smoothing algorithm is used for noise reduction of the Raman spectrum.

3. The method according to claim 1, characterized in that, The multi-scale feature extraction module is a three-branch parallel one-dimensional convolutional neural network; Each convolutional branch contains four one-dimensional convolutional layers, with a kernel size of 7. Different branches extract features at different scales by setting different convolutional strides and output channel numbers. The first branch has a convolution stride of 2 and an output channel of 8, which is used to extract high-resolution local detail features; The second branch has a convolution stride of 4 and 16 output channels, and is used to extract medium-resolution spectral peak structure features. The third branch has a convolution stride of 8 and 32 output channels, and is used to extract low-resolution global trend features.

4. The method according to claim 1, characterized in that, Step 4 specifically includes: Step 4.1, for any scale Using the Raman spectral features at this scale as the query vector and the infrared spectral features at this scale as the key vector and value vector, the first attention calculation is performed. Step 4.2: Using the infrared spectral features at this scale as the query vector and the Raman spectral features at this scale as the key vector and value vector, perform the second attention calculation; Step 4.3: The results of the first attention calculation and the second attention calculation are weighted and fused with the original Raman spectral features and infrared spectral features respectively through an adaptive residual fusion structure and fusion weights. The fusion weights are... , These are learnable parameters; Step 4.4: Perform nonlinear enhancement processing on the weighted fused features to obtain the enhanced infrared spectral cross-modal features and Raman spectral cross-modal features at this scale; Step 4.5: Form an enhanced infrared spectral cross-modal feature set based on the enhanced infrared spectral cross-modal features and Raman spectral cross-modal features at each scale. Raman spectroscopy cross-modal feature set .

5. The method according to claim 1, characterized in that, Step 5 specifically includes: Step 5.1, the first stage is the low-resolution to high-resolution interaction: after dimensional alignment of the low-resolution features, cross-scale attention interaction is performed with the adjacent higher-resolution features, while the lowest resolution features remain unchanged. Step 5.2, the second stage is the interaction from high resolution to low resolution: after dimensional alignment of the high resolution features, they are fused with the adjacent lower resolution features, while the highest resolution features remain unchanged. Step 5.3: After the two-stage interaction, the cross-modal features of the Raman spectrum at the three scales are spliced ​​together according to the channel dimension to obtain a unified feature representation of the Raman spectrum. Similarly, the cross-modal features of the infrared spectrum at the three scales are spliced ​​together according to the channel dimension to obtain a unified feature representation of the infrared spectrum.

6. The method according to claim 1, characterized in that, When performing disease risk prediction, step 6 specifically includes: The classification and discrimination module fuses the unified feature representations of infrared spectroscopy and Raman spectroscopy, inputs them into the multilayer perceptron classifier, and outputs the risk probability of the subject belonging to the category of aortic dissection, myocardial infarction or healthy control as the disease risk prediction result.

7. The method according to claim 1, characterized in that, When performing disease risk prediction, step 6 specifically includes: The standardized clinical indicator data is mapped into clinical feature vectors through an embedding layer; The clinical feature vector, the unified feature representation of infrared spectroscopy, and the unified feature representation of Raman spectroscopy are fused across modes to obtain fused multimodal features; The multimodal features are fused and input into a multilayer perceptron classifier, and the output prognostic risk value of the object to be tested is used as the prognostic risk prediction result.

Images