Raman spectrum myocardial infarction screening method, system, equipment and medium

Abstract

The invention discloses a Raman spectrum myocardial infarction screening method, a Raman spectrum myocardial infarction screening system, Raman spectrum myocardial infarction screening equipment and a medium, and belongs to the field of biomedical spectrum detection. The method comprises the following steps: collecting a trace peripheral whole blood sample, mixing with an SERS (Surface Enhanced Raman Scattering) reagent, dispensing, depositing and drying, and preparing a to-be-detected sample by utilizing a coffee ring effect; collecting the Raman spectrum to obtain original data; self-adaptive dual-channel interference elimination processing is carried out on original data to obtain a pure biomarker spectrum, and the processing fuses physical correction based on a standard spectrum library and data driving correction based on an adversarial auto-encoder; microcosmic, mesoscopic and macroscopic biomarker features are extracted from the pure spectrum and aggregated to generate a multi-dimensional feature vector; and finally, inputting the feature vector into a lightweight deep learning model to obtain a risk probability and finish grading evaluation. According to the method, rapid and high-precision myocardial infarction screening of trace whole blood without centrifugation is realized, and the accuracy and practicability of instant detection are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of biomedical spectral detection technology, specifically to a Raman spectroscopy method, system, device, and medium for screening myocardial infarction. Background Technology

[0002] Myocardial infarction (MI) is one of the leading causes of cardiovascular disease death worldwide, and early diagnosis is crucial to reducing mortality. Current clinical diagnosis primarily relies on immunological testing of cardiac troponin I / CTnT. While highly specific, this method has limitations such as a delayed detection window (detectable 3-4 hours after symptom onset), complex procedures (requiring centrifugation to separate serum / plasma), and expensive equipment, making it difficult to meet the point-of-care testing (POCT) needs of emergency and primary care settings.

[0003] Raman spectroscopy, with its advantages of label-free, rapid, and non-destructive detection, has received widespread attention in the field of biomedical testing in recent years. Its principle is based on the inelastic scattering of incident light by molecules, which provides unique molecular fingerprint information. Previous studies have shown that the Raman spectra of blood from patients with myocardial infarction exhibit specific wavenumber regions (e.g., 1000 cm⁻¹ for phenylalanine). - ¹, Tyrosine at 825 cm - ¹) Characteristic changes will be observed, especially the intensity ratio of phenylalanine to tyrosine (Phe-Tyr Ratio), which has been reported as a potential sensitive indicator of inflammatory state and tissue damage.

[0004] However, directly applying Raman spectroscopy to myocardial infarction screening in whole blood samples still faces key technical bottlenecks: (1) Complex matrix interference is difficult to eliminate effectively: Blood samples have complex components, containing high abundance of proteins (such as albumin and immunoglobulins), hemoglobin, lipids, etc. These substances generate strong background Raman signals, which severely mask the characteristic signals of low-concentration biomarkers. Traditional single-channel interference elimination methods, such as physical model methods that only use extended multiplicative signal correction (EMSC) or rely solely on general machine learning algorithms, are difficult to adaptively handle the complex and variable interference patterns of blood matrix in different individuals and under different physiological or pathological conditions, resulting in impure and inaccurate biomarker signal extraction.

[0005] (2) Biomarkers are limited in scope and provide incomplete diagnostic information: Existing studies often rely excessively on a single ratio or a limited number of spectral features (such as the Phe-Tyr ratio), neglecting the pathophysiological complexity of myocardial infarction as a systemic disease. A single biomarker is insufficient to effectively distinguish different subtypes of myocardial infarction (such as ST-segment elevation myocardial infarction (STEMI) and non-ST-segment elevation myocardial infarction (NSTEMI), assess disease severity, or monitor treatment response, thus limiting the comprehensiveness of diagnosis and its clinical applicability.

[0006] (3) The sample pretreatment process is cumbersome and does not meet the requirements of POCT: In order to reduce matrix interference, most existing solutions still require centrifugation of blood samples to obtain serum / plasma, or use complex surface-enhanced Raman scattering (SERS) substrates to enhance the signal. These additional steps not only increase the operation time, cost and technical threshold, but also introduce sample processing errors, which is contrary to the goal of simple, rapid and integrated testing pursued by primary healthcare, pre-hospital emergency care and home health monitoring.

[0007] Therefore, there is an urgent need for a new method that can quickly and accurately screen for myocardial infarction directly from trace amounts of peripheral whole blood. This method should be able to effectively overcome blood matrix interference, comprehensively utilize multi-dimensional pathological information, and be suitable for resource-constrained real-time testing scenarios. Summary of the Invention

[0008] To address the problems in existing technologies, such as the complexity and difficulty in adaptively eliminating blood matrix interference, the lack of comprehensive diagnostic information due to the single dimension of biomarkers, and the cumbersome sample pretreatment process that cannot meet the needs of immediate testing, this invention provides a Raman spectroscopy method, system, device, and medium for myocardial infarction screening, thereby solving the aforementioned technical deficiencies.

[0009] This invention proposes a Raman spectroscopy method for screening myocardial infarction, which includes the following steps: S1. Collect peripheral whole blood samples, mix the peripheral whole blood samples with surface-enhanced Raman scattering reagent, and then perform drop-coating deposition and drying on the substrate to prepare the test sample using the coffee ring effect; S2. Raman spectroscopy is performed on the central region of the sample to be tested to obtain the raw spectral data; S3. Adaptive dual-channel interference elimination processing is performed on the original spectral data to obtain pure biomarker spectra. The adaptive dual-channel interference elimination processing includes: correcting the original spectral data using a physical prior method to obtain a first corrected spectrum, correcting the original spectral data using a data-driven method to obtain a second corrected spectrum, and fusing the first corrected spectrum and the second corrected spectrum. S4. Extract multi-scale biomarker features from pure biomarker spectra, and aggregate the multi-scale biomarker features to generate multi-dimensional feature vectors. S5. Input the multidimensional feature vector into the lightweight deep learning model for processing to obtain the risk probability of myocardial infarction, and perform risk assessment and classification based on the risk probability.

[0010] Preferably, step S3 involves adaptive dual-channel interference cancellation processing of the original spectral data, including the following sub-steps: S31. The original spectral data is processed through the physical prior channel, wherein the extended multiplicative signal correction algorithm is used to remove known matrix interference from the original spectral data based on the standard interference spectral library to obtain the first corrected spectrum; S32. The original spectral data is processed through the data-driven channel, wherein a pre-trained adversarial autoencoder model is used to extract and remove interference patterns from the original spectral data to obtain the second corrected spectrum. S33. Based on the signal-to-noise ratio of the original spectral data, the first and second corrected spectra are dynamically fused to obtain the pure biomarker spectrum.

[0011] More preferably, in step S32, the training process of the pre-trained adversarial autoencoder model includes the following sub-steps: S321. Use the encoder to extract latent variables from the raw spectral data; S322. Reconstruct the interference spectrum based on the latent variables using a generator; S323. The discriminator forces the distribution of the latent variables extracted by the encoder to approximate the preset prior distribution. S324. Construct a total loss function based on the reconstruction loss function, the generator adversarial loss function, and the discriminator loss function, and train the adversarial autoencoder model by optimizing the total loss function.

[0012] More preferably, in step S324, the expression for the reconstruction loss function is: ; The discriminator loss function is expressed as follows: ; The expression for the generator adversarial loss function is: ; The expression for the total loss function is: ; in, To reconstruct the loss function; This represents the number of samples in the training batch. For the first One set of raw spectral data; To combat the reconstruction of the autoencoder model One interference spectrum; The discriminator loss function; For the first sampled from the standard Gaussian distribution N(0,I) A priori code; For the generator adversarial loss function, For discriminator networks; To counter the encoder of the autoencoder model; This is the total loss function; These are the weighting coefficients of the discriminator loss function; These are the weighting coefficients for the generator's adversarial loss function.

[0013] More preferably, in step S33, the first corrected spectrum and the second corrected spectrum are dynamically fused according to the signal-to-noise ratio of the original spectral data to obtain the pure biomarker spectrum, including the following sub-steps: S331. Calculate the signal-to-noise ratio of the original spectral data; S332. Based on the calculated signal-to-noise ratio, dynamically allocate the weight coefficients of the physical prior channel and the data-driven channel; S333. According to the weighting coefficients, the first correction spectrum and the second correction spectrum are weighted and fused.

[0014] Preferably, in step S4, multi-scale biomarker features are extracted from the pure biomarker spectrum, and the multi-scale biomarker features are aggregated to generate a multi-dimensional feature vector, including the following sub-steps: S41. Extract microscale molecular features from the pure biomarker spectrum. The microscale molecular features include calculating the ratio of the characteristic peak intensity of phenylalanine to that of tyrosine. S42. Extracting mesoscale cellular component characteristics from pure biomarker spectra; S43. Construct a pathophysiological graph network based on microscopic molecular features and mesoscopic cellular component features, and apply graph convolutional networks to aggregate information from the pathophysiological graph network to generate multidimensional feature vectors.

[0015] Preferably, step S5 involves inputting the multidimensional feature vector into a lightweight deep learning model for processing to obtain the risk probability of myocardial infarction, including the following sub-steps: S51. Input the multidimensional feature vector into the lightweight deep learning model, specifically a lightweight student model obtained through knowledge distillation training. S52. The lightweight student model adaptively selects an internal calculation path based on the signal-to-noise ratio of the multidimensional feature vector, wherein: when the signal-to-noise ratio is higher than the first threshold, a simplified calculation path is enabled; when the signal-to-noise ratio is between the first threshold and the second threshold, a standard calculation path is enabled; and when the signal-to-noise ratio is not higher than the second threshold, a data reprocessing mechanism is triggered. S53. The lightweight student model outputs the probability of myocardial infarction. S54. Calculate the confidence level based on the risk probability. The confidence level assessment includes: calculating the prediction entropy of the risk probability and determining whether the prediction entropy is greater than a preset threshold. If it is greater, mark the risk probability or the risk assessment level derived from it as a low confidence result and trigger the review process.

[0016] The present invention also proposes a Raman spectroscopy myocardial infarction screening system for implementing the method as described in any of the above, the system comprising: The sample collection and processing module is configured to collect peripheral whole blood samples, mix the peripheral whole blood samples with surface-enhanced Raman scattering reagents, perform drop-coating deposition and drying on a substrate, and prepare the test sample by utilizing the coffee ring effect; The spectral acquisition module is configured to acquire Raman spectra of the central region of the sample to be tested, thereby obtaining raw spectral data. The spectral processing and analysis module is configured to perform the following operations: The original spectral data is subjected to adaptive dual-channel interference elimination processing to obtain a pure biomarker spectrum. The adaptive dual-channel interference elimination processing includes: correcting the original spectral data using a physical prior method to obtain a first corrected spectrum, correcting the original spectral data using a data-driven method to obtain a second corrected spectrum, and fusing the first corrected spectrum and the second corrected spectrum. Multi-scale biomarker features are extracted from the pure biomarker spectrum, and the multi-scale biomarker features are aggregated to generate a multi-dimensional feature vector. The risk assessment and decision-making module is configured to input the multidimensional feature vector into a lightweight deep learning model for processing, obtain the risk probability of myocardial infarction, and perform risk assessment and classification based on the risk probability.

[0017] The present invention also proposes a terminal device, including a memory, a processor, and a computer program stored in the memory and capable of running on the processor, wherein the processor executes the computer program to implement the steps of any of the Raman spectroscopy myocardial infarction screening methods described above.

[0018] The present invention also proposes a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of any of the Raman spectroscopy myocardial infarction screening methods described above.

[0019] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) This method enables rapid processing of ultra-small whole blood samples without centrifugation, greatly improving the convenience and immediacy of testing. By employing a sample preparation method based on drop-coating deposition and the coffee ring effect, only a small amount of capillary whole blood is needed to complete the sample preparation, eliminating the need for complex centrifugation steps. This method utilizes the physical effects of high molecular weight proteins migrating to the edges and low molecular weight biomarkers accumulating in the center, achieving preliminary separation of interfering substances during the preparation stage. This lays the physical foundation for subsequent high-precision analysis, thereby significantly shortening the total time from sampling to obtaining screening results, truly meeting the stringent requirements for point-of-care testing in scenarios such as emergency departments, primary healthcare, and home monitoring.

[0020] (2) By employing a dual-channel adaptive interference elimination architecture combining physical and data methods, this invention effectively overcomes interference from complex blood matrix, significantly improving the signal-to-noise ratio and detection accuracy. This invention does not rely solely on traditional physical models or general machine learning algorithms. Instead, it combines an extended multiplicative signal correction (EMSC) physical prior channel based on a standard interference spectral library with a data-driven channel based on an adversarial autoencoder (AAE) for learning unknown interference patterns. The results from both channels are dynamically fused based on the signal-to-noise ratio of the original spectrum. This adaptive mechanism ensures optimal elimination results for both typical and unknown, variable interference patterns. This dual-channel method significantly improves the signal-to-noise ratio of characteristic peaks, thereby fundamentally solving the core challenge of direct whole-blood detection and providing pure biomarker spectral signals for high-sensitivity, high-specificity detection.

[0021] (3) A multi-scale biomarker feature network at the micro, meso, and macro scales was constructed, enabling a systematic and comprehensive diagnosis of the pathological state of myocardial infarction. This invention breaks through the dependence on a single ratio (such as Phe-Tyr), systematically extracting micro-scale molecular features (such as Phe-Tyr ratio and oxidative stress signals) and meso-scale cellular component features (such as cell-free DNA (cfDNA) signals and lipid metabolism signals). Furthermore, by constructing a pathophysiological graph network and applying graph convolutional networks (GCNs) for information aggregation, a macro-scale multi-dimensional feature vector reflecting the pathophysiological changes of the disease system was formed. This multi-scale feature fusion strategy not only significantly enhances the information richness and specificity of diagnosis, but also gives it the potential to distinguish different subtypes of myocardial infarction and assess the complexity of the disease, overcoming the shortcomings of existing technologies in terms of one-sided diagnostic information and limited clinical applicability.

[0022] (4) A lightweight and highly reliable intelligent inference engine for resource-constrained scenarios was designed, achieving fast edge computing while ensuring high accuracy. By employing knowledge distillation technology, the knowledge of the complex teacher model is transferred to the lightweight student model with a very small number of parameters, and a dynamic computation graph mechanism based on the input feature signal-to-noise ratio is introduced, enabling the model to adaptively select the computation path and maximize computational efficiency while ensuring high accuracy. This engine can quickly complete inference on the device side and exhibits excellent comprehensive diagnostic performance, including high sensitivity, high specificity, and high accuracy. In addition, the built-in confidence assessment and review triggering mechanism based on prediction entropy provides dual protection for the reliability of screening results, effectively reducing the risk of misdiagnosis and missed diagnosis.

[0023] (5) A complete and actionable clinical decision support closed loop has been formed, enhancing the clinical value and application depth of screening results. This invention is not limited to outputting a single risk probability, but further transforms it into a clear multi-level risk assessment grading, and provides clear clinical action recommendations for each level. At the same time, the system supports longitudinal comparison of key biomarker indicators, which can be used to monitor treatment response. Combined with the edge-cloud collaborative mechanism, this invention constructs a complete solution from rapid detection and intelligent analysis to clinical decision support and dynamic monitoring, significantly enhancing its practical value in real medical environments. Attached Figure Description

[0024] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments, taken with reference to the accompanying drawings: Figure 1 This is a flowchart of the Raman spectroscopy method for screening myocardial infarction; Figure 2 This is a flowchart of the adaptive dual-channel interference cancellation algorithm; Figure 3 This is a diagram of the architecture of a Raman spectroscopy myocardial infarction screening system; Figure 4 This is a schematic diagram of the structure of a computer system suitable for implementing the embodiments of the present invention. Detailed Implementation

[0025] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings.

[0026] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0027] Figure 1 A flowchart of the Raman spectroscopy method for screening myocardial infarction is shown, such as... Figure 1 As shown in the figure, an embodiment of the present invention provides a Raman spectroscopy method for screening myocardial infarction based on multi-scale feature fusion and adaptive interference cancellation, comprising the following steps: S1. Collect peripheral whole blood samples, mix the peripheral whole blood samples with a surface-enhanced Raman scattering reagent, and then perform drop-coating deposition on a substrate followed by drying. The sample to be tested is prepared using the coffee ring effect. The specific implementation method is as follows: First, a capillary whole blood sample is collected from the subject. This invention uses capillary whole blood collected directly from the fingertip, requiring a very small sample volume of only about 0.1 μL, thus avoiding the complex pretreatment of venous blood collection and centrifugation to separate serum / plasma required in traditional methods. Sample collection can be accomplished using a device integrated with a microfluidic chip, which utilizes capillary action to automatically and accurately quantify the required blood volume. To ensure the stability of the target biomarker signal, subsequent spectral measurements should be performed immediately after sample collection, and the entire process should be completed within 2 minutes to prevent spectral signal drift caused by metabolic activity after blood leaves the body.

[0028] After obtaining the peripheral whole blood sample, sample preparation was immediately performed to obtain the test sample. The preparation process employed a drop-coating deposition and drying method. Specifically, 0.1 μL of the collected whole blood sample was thoroughly mixed with a surface-enhanced Raman scattering reagent (e.g., 0.4 μL of a 4 mg / mL colloidal silver nanoparticle solution) to obtain a mixture. Subsequently, this mixture was precisely drop-coated onto the central region of a hydrophobically treated substrate (e.g., a silicon wafer) with a water contact angle greater than 90°. Then, the droplet was allowed to air dry naturally for approximately 60 ± 5 seconds under a constant temperature and humidity environment (e.g., 25 ± 1°C, 50 ± 5% relative humidity) to form a solid test film.

[0029] During this drying process, the evaporation rate at the droplet edge is faster than at the center, causing the internal liquid to flow outward, resulting in the coffee ring effect. This physical effect leads to differential deposition of components in the solution: high molecular weight proteins (such as albumin and immunoglobulin IgG) migrate towards the droplet edge and deposit, while low molecular weight target biomarkers (such as phenylalanine, tyrosine, and cell-free DNA) are significantly enriched in the droplet center. Therefore, the sample prepared through this step has its central region physically separated from the main matrix interferences and enriched with target analytes, creating ideal conditions for subsequent high signal-to-noise ratio Raman spectroscopy acquisition.

[0030] Continue to refer to Figure 1 The Raman spectroscopy method for screening myocardial infarction based on multi-scale feature fusion and adaptive interference elimination provided in this embodiment of the invention further includes the following steps: S2. Raman spectroscopy is performed on the central region of the sample to obtain raw spectral data. The specific implementation method is as follows: The sample prepared in step S1 was subjected to spectral acquisition using a Raman spectrometer. During acquisition, the focus of the excitation laser must be precisely aligned with the central region of the sample. This central region is where the target biomarker is enriched and macromolecular interferences are relatively few after the coffee ring effect treatment; acquisition at this location will maximize the acquisition of effective signals.

[0031] The specific configuration of the spectral acquisition device is as follows: A semiconductor laser with a wavelength of 785 nm was selected as the excitation source, and its output power was set to 100 mW. This wavelength was chosen to effectively avoid strong interference from common fluorescent substances in blood samples, while the 100 mW power ensured sufficient signal intensity without causing photodamage to biomolecules. The spectrometer covers a spectral range of 400 cm⁻¹. -1 Up to 1800 cm -1 This range encompasses the characteristic Raman peaks of key biomarkers such as phenylalanine, tyrosine, cell-free DNA, and various lipid metabolites, with a spectral resolution set to 6 cm⁻¹. -1 This ensures that adjacent characteristic spectral peaks can be distinguished. A high numerical aperture microscope objective focuses the laser onto the sample surface, forming a focused spot with a diameter of approximately 16 μm, precisely covering the central region to be measured.

[0032] The specific acquisition process is as follows: Nine independent measurement points are selected in the central region of each sample to be tested for scanning. At each point, the spectrometer's integration time is set to 3 seconds to accumulate a sufficient number of scattered photons and obtain a single spectrum with a high signal-to-noise ratio. To improve data reliability, each point needs to be acquired twice. Finally, the raw spectral data obtained through this step is mathematically represented as a two-dimensional array of wavenumber (λ) and intensity, denoted as... This serves as the input for all subsequent algorithmic processing.

[0033] Figure 2 The flowchart of the adaptive dual-channel interference cancellation algorithm is shown, as follows: Figure 1 and Figure 2 As shown in the figure, the Raman spectroscopy method for screening myocardial infarction based on multi-scale feature fusion and adaptive interference elimination provided by the embodiments of the present invention further includes the following steps: S3. Adaptive dual-channel interference elimination processing is performed on the original spectral data to obtain pure biomarker spectra. The adaptive dual-channel interference elimination processing includes: correcting the original spectral data using a physical prior method to obtain a first corrected spectrum, correcting the original spectral data using a data-driven method to obtain a second corrected spectrum, and fusing the first corrected spectrum and the second corrected spectrum.

[0034] Step S3 is the adaptive dual-channel interference elimination process, which aims to extract pure biomarker spectra reflecting disease pathological changes from the original spectral data containing complex matrix interference with high fidelity. This step is achieved through physical prior and data-driven dual-channel parallel processing, and adaptive fusion based on the sample's own signal quality. Specifically, it includes the following sub-steps: S31. Process the raw spectral data through the physical prior channel.

[0035] This sub-step, based on known biochemical knowledge, constructs a standard interference spectral library containing pure Raman spectra of common high-abundance interfering components in blood (such as albumin, immunoglobulin G, and hemoglobin). The Extended Multiplicative Signal Correction (EMC) algorithm is then applied to the raw spectral data. The process involves subtracting the linear combination contribution of these known interfering substances from the original spectrum and correcting for baseline drift to obtain the first corrected spectrum. (That is, the spectrum after processing by the physical prior channel). This process is represented by the following formula: ; in, This is the raw spectral data; For the first A reference interfering spectrum (such as albumin, IgG, hemoglobin); For the first Weighting coefficients for each reference spectrum; This is a polynomial baseline function used to correct background drift.

[0036] S32. Process the raw spectral data through the data-driven channel.

[0037] This sub-step utilizes a pre-trained Adversarial Autoencoder (AAE) model to learn and eliminate unknown or complex perturbation patterns. The training and implementation of the AAE model are detailed below: 1. Model Architecture: The adversarial autoencoder model consists of an encoder, a generator (decoder), and a discriminator. The encoder is a three-layer fully connected neural network with 512, 256, and 128 neurons respectively, ultimately mapping the input spectrum into a 64-dimensional latent variable vector. This is used to characterize the learned perturbation pattern features. The generator is also a three-layer fully connected neural network with 128, 256, and 512 neurons respectively, responsible for representing the perturbation pattern features learned from the latent variables. Reconstruct the predicted interference spectrum Discriminator Network Used to distinguish the latent variables generated by the encoder from the encoding sampled from the standard Gaussian prior distribution N(0,I).

[0038] 2. Training Dataset Configuration: The model was pre-trained using a Raman spectroscopy dataset containing 100 subjects, 50 with myocardial infarction and 50 without. The data was divided into a training set (70 cases), a validation set (15 cases), and a test set (15 cases) in a 7:1.5:1.5 ratio. The pre-training strategy was to use only the 70 unlabeled spectral data from the training set for unsupervised pre-training, aiming to learn general perturbation patterns in the blood matrix without relying on disease label information.

[0039] 3. Training Process and Key Hyperparameters: The model was trained using the Adam optimizer with a learning rate set to 1×10⁻⁶. -4 (β1=0.9, β2=0.999). The total training duration is 150 epochs, with an early stopping mechanism: training is terminated if the validation set loss does not decrease for 20 consecutive epochs. The batch size is set to 16, and weight decay (coefficient 5×10) is used. -5 Regularization is performed using Dropout (ratio 0.3) to prevent overfitting.

[0040] 4. Loss Function and Training Objective: The training of a pre-trained adversarial autoencoder model is achieved by optimizing a joint total loss function. Completed, this function is the reconstruction loss function. Discriminator loss function Adversarial loss function against generator Weighted sum: ; The expression for the reconstruction loss function is as follows: ; The discriminator loss function is expressed as follows: ; The expression for the generator adversarial loss function is: ; in, This is the total loss function; These are the weighting coefficients of the discriminator loss function; The weight coefficients of the generator adversarial loss function; preferably, =0.5, =0.2, determined through validation set optimization; The reconstruction loss function measures the difference between the original spectrum and the reconstructed spectrum. This represents the number of samples in the training batch. For the first One set of raw spectral data; To combat the reconstruction of the autoencoder model One interference spectrum; This is the discriminator loss function, used to train the discriminator to distinguish between real and generated codes; For the first sampled from the standard Gaussian distribution N(0,I) A priori code; The generator adversarial loss function is used to make the encoder output approximate the prior distribution; For discriminator networks; To counter the encoder of the autoencoder model.

[0041] After the adversarial autoencoder model is trained, the raw spectral data to be processed is... Input the pre-trained adversarial autoencoder model. The encoder of the adversarial autoencoder model extracts its interference features, and the generator reconstructs the corresponding interference spectrum. From raw spectral data Subtract this reconstructed interference spectrum That is, the second corrected spectrum is obtained after processing by the data-driven channel. : ; S33. Based on the signal-to-noise ratio of the original spectral data, dynamically fuse the first corrected spectrum and the second corrected spectrum.

[0042] This sub-step first calculates the signal-to-noise ratio of the original spectral data. (Signal-to-Noise Ratio) serves as the basis for fusion. One specific calculation method is as follows: ; In the formula, The original spectral data are in the range of 900-1100 cm⁻¹ -1 Average strength of the interval (signal region); The original spectrum is in the range of 200-400 cm⁻¹ -1 Standard deviation of the interval (noise region).

[0043] Then, based on the calculated signal-to-noise ratio (SNR), the weight coefficients of the physical prior channel and the data-driven channel are dynamically allocated. A weighting function is designed such that, under high SNR conditions, the results of the physical prior channel are relied upon more, while under low SNR conditions, the results of the data-driven channel are relied upon more.

[0044] Finally, the first and second corrected spectra are weighted and fused according to the weighting coefficients to obtain the final pure biomarker spectrum. : ; In the formula, The weight values ​​are dynamically calculated based on the signal-to-noise ratio. Through this adaptive fusion mechanism, this method can achieve optimal interference cancellation for input spectra of varying quality.

[0045] It should be understood that the aforementioned adaptive dual-channel interference cancellation processing has multiple feasible implementation methods. In some variations, the physical prior method is not limited to the extended multiplicative signal correction algorithm; other signal processing algorithms based on prior models, such as wavelet transform and morphological filtering, can also be used to adapt to known matrix interferences with different characteristics. In the data-driven method, the adversarial autoencoder model can be replaced by other generative models, such as variational autoencoders, to learn and remove interference patterns. The fusion mechanism of the first and second corrected spectra is not limited to dynamic weighting based on the signal-to-noise ratio; for example, an attention mechanism can be used for weighted fusion. These variations are all based on the core concept of obtaining pure biomarker spectra by combining physical prior methods with data-driven methods.

[0046] Continue to refer to Figure 1 The Raman spectroscopy method for screening myocardial infarction based on multi-scale feature fusion and adaptive interference elimination provided in this embodiment of the invention further includes the following steps: S4. Extract multi-scale biomarker features from the pure biomarker spectrum, and aggregate the multi-scale biomarker features to generate a multi-dimensional feature vector.

[0047] Step S4 is the multi-scale biomarker feature extraction and aggregation step, the purpose of which is to obtain the pure biomarker spectrum from step S3. In this process, features at different pathophysiological scales are systematically extracted and fused into a comprehensive multidimensional feature vector, providing a high-dimensional discriminative basis for the final risk assessment. Specifically, this includes the following sub-steps: S41. Extract microscale molecular features from pure biomarker spectra.

[0048] This sub-step extracts microscale molecular features reflecting changes at the molecular level from the pure biomarker spectrum. One of the most crucial features is the calculation of the ratio of the characteristic peak intensity of phenylalanine to that of tyrosine (Phe-TyrRatio), a sensitive indicator of inflammation and tissue damage. The specific calculation formula is as follows: ; in, The phenylalanine-tyrosine ratio is a microscale characteristic. 1000±5 cm -1 Raman intensity at wavenumber (characteristic peak of phenylalanine); 825±3 cm -1 Raman intensity at wavenumber (tyrosine characteristic peak).

[0049] In addition, other microscopic molecular features can be extracted, such as the characteristic peak intensity I of glutathione (1404±3 cm⁻¹), which reflects oxidative stress. - ¹).

[0050] S42. Extract mesoscale cellular component characteristics from pure biomarker spectra.

[0051] This sub-step extracts mesoscale cellular component features reflecting changes in cellular component levels from the purified biomarker spectrum. This includes cell-free DNA (cfDNA) signals associated with cell damage and death. cfDNA can be characterized by its intensity at specific wavenumbers, for example: ; in, Mesoscale characteristics, cell-free DNA signals; 1318±2 cm -1 Raman intensity at wavenumber (adenine characteristic peak); 1337±2 cm -1 Raman intensity at wavenumber (guanine characteristic peak).

[0052] Simultaneously, lipid characteristics related to cell membrane metabolism and energy metabolism can be extracted, such as the characteristic peak intensity of phospholipids. .

[0053] S43. Construct a pathophysiological graph network based on microscopic molecular features and mesoscopic cellular component features, and apply graph convolutional networks to aggregate information from the pathophysiological graph network to generate multidimensional feature vectors.

[0054] This sub-step first constructs a pathophysiological map network based on the extracted microscopic molecular features and mesoscopic cellular component features. In this network, each extracted feature (e.g., phenylalanine-tyrosine ratio, glutathione intensity, two dimensions of the cfDNA signal vector, phospholipid intensity, etc.) serves as a node. The connections (adjacency relationships) between nodes and their weights can be obtained based on known biochemical correlations, statistical correlations (such as Pearson correlation coefficients), or through data learning, to encode the pathophysiological associations between different biomarkers.

[0055] Subsequently, a Graph Convolutional Network (GCN) is applied to this pathophysiological graph network for information aggregation and high-order feature learning. GCN updates node representations by aggregating information from each node and its neighboring nodes; its forward propagation formula for one layer can be expressed as: ; in, For the first +1 layer node feature representation; It is the ReLU activation function; for The degree matrix (for angle matrix); Let A+I be the adjacency matrix with self-loops (A is the original adjacency matrix, and I is the identity matrix). For the first Layer node feature representation; For the first The learnable weight matrix of the layer.

[0056] By stacking multiple layers of GCN, the model is able to capture the complex relationships between multi-hop neighbors in the graph, thereby learning high-level node representations that incorporate multi-scale, structured pathological information.

[0057] Finally, all node features (or graph representations after global pooling) processed by the last GCN layer are concatenated or aggregated to form a multidimensional feature vector for final classification. This feature vector integrates systematic information from molecules to cellular components and their interaction networks, and its dimensions can be, for example, 128, comprehensively representing the pathophysiological state associated with myocardial infarction.

[0058] It should be understood that the methods for extracting and aggregating multi-scale biomarker features can be expanded. When extracting microscale molecular features, in addition to the ratio of phenylalanine to tyrosine peak intensities, other features such as tryptophan (1550 cm⁻¹) can also be included. -1 ), phenylalanine dimer (1030 cm -1 The characteristic peak intensity, etc., can be used to extract features of mesoscale cellular components. Raman features related to small extracellular vesicles can be further introduced when extracting features of mesoscale cellular components. When constructing a pathophysiological graph network based on microscale molecular features and mesoscale cellular component features and performing information aggregation, the graph convolutional network can be replaced by other graph neural network models, such as graph attention networks or Transformer architectures, to generate multidimensional feature vectors.

[0059] Continue to refer to Figure 1 The Raman spectroscopy method for screening myocardial infarction based on multi-scale feature fusion and adaptive interference elimination provided in this embodiment of the invention further includes the following steps: S5. Input the multidimensional feature vector into the lightweight deep learning model for processing to obtain the risk probability of myocardial infarction, and perform risk assessment and classification based on the risk probability.

[0060] Step S5 is the risk assessment and decision-making step. Its purpose is to transform the multidimensional feature vector generated in step S4 into an intuitive, clinically applicable probability of myocardial infarction risk, and to classify the risk accordingly. Specifically, it includes the following sub-steps: S51. Input the multidimensional feature vector into the lightweight deep learning model.

[0061] This sub-step inputs the multidimensional feature vector into a lightweight deep learning model for processing. Specifically, this lightweight deep learning model is a lightweight student model trained through knowledge distillation. In this process, a structurally complex and high-performance teacher model (e.g., a 3D convolutional neural network, 3D-CNN, with an input dimension of 1400×1) is first trained to achieve high accuracy on the training set. Then, the soft labels output by this teacher model (i.e., probability distributions smoothed by temperature parameters) and the true labels are used to jointly guide the training of a structurally streamlined lightweight student model (e.g., a variant based on the MobileNetV3 architecture with less than 1MB of parameters), specifically optimized for edge devices. The total loss function of knowledge distillation... Combined with cross-entropy loss function and knowledge distillation loss function (Usually KL divergence), expressed as: ; in, The weights can be dynamic or fixed (e.g., 0.3). Through this process, the student model can learn the implicit knowledge contained in the teacher model, thus maintaining diagnostic performance close to that of the teacher model while significantly reducing model size and computational complexity.

[0062] S52, The lightweight student model adaptively selects the internal computation path based on the signal-to-noise ratio of the input features.

[0063] In this sub-step, the lightweight student model possesses a dynamic computation graph mechanism, which adaptively selects internal computation paths based on the quality of the input multidimensional feature vectors (represented by the estimated signal-to-noise ratio) to maximize inference efficiency while ensuring accuracy. The specific strategy is as follows: When the estimated signal-to-noise ratio is higher than a first threshold (e.g., 15 dB), the model enables a simplified computation path, which may use only a few computational units such as two convolutional layers to achieve the fastest inference speed.

[0064] When the estimated signal-to-noise ratio is between the first threshold and the second threshold (for example, 8 dB < SNR ≤ 15 dB), the model enables the standard calculation path and processes it using standard structures such as a complete 4-layer convolution to balance speed and accuracy.

[0065] When the estimated signal-to-noise ratio is not higher than the second threshold (for example, SNR ≤ 8 dB), the model determines that the input data quality is poor and triggers the data reprocessing mechanism. This mechanism may include prompting the operator to resample, invoking a more complex cloud model for verification, or adopting an enhanced data preprocessing process on the device side.

[0066] S53. Output the risk probability of myocardial infarction by the lightweight student model.

[0067] After being processed by the selected calculation path, the lightweight student model finally outputs a scalar value between 0 and 1, which is the risk probability of the myocardial infarction. This probability value directly reflects the model confidence of the input sample belonging to the myocardial infarction category.

[0068] S54. Conduct confidence evaluation and risk assessment grading based on the risk probability.

[0069] This sub-step first conducts a confidence evaluation on the risk probability output by the model. The confidence evaluation is achieved by calculating the prediction entropy H of the risk probability distribution: ; where H is the prediction entropy, used to evaluate the confidence; is the probability of the th class predicted by the model.

[0070] Then, it is judged whether the prediction entropy H is greater than a preset confidence threshold (for example, H > 0.7). If it is greater than this threshold, it indicates that the model has a high uncertainty about the current prediction. Then, mark the current risk probability or the risk assessment level derived from it as a low-confidence result and automatically trigger the review process. The review process may include prompting a clinical doctor to review, suggesting a retest in the short term, or uploading the original data to the cloud for more complex analysis.

[0071] Finally, a clear risk assessment grading is conducted based on the obtained risk probability to provide direct action suggestions for clinical practice: High risk (probability ≥ 0.85): It indicates that the characteristics of acute myocardial injury are significant, and it is recommended to immediately refer to the emergency department or the cardiology department for diagnosis and treatment.

[0072] Medium risk (0.35 ≤ probability < 0.85): It indicates that there are abnormalities such as inflammation or metabolic disorders, and it is recommended to retest within 24 - 48 hours or further evaluate in combination with other clinical indicators.

[0073] Low risk (probability <0.35): Indicates no significant pathological features of myocardial infarction and can be included in routine health monitoring.

[0074] In addition, the system can also support treatment monitoring functions, such as longitudinally comparing changes in key biomarkers in the same patient. To assess treatment response. Where ΔR is the change in the Phe-Tyr ratio; R current R is the currently measured Phe-Tyr ratio. baseline The baseline (before treatment) Phe-Tyr ratio.

[0075] like If a certain threshold is exceeded (e.g., >0.5), it indicates that the treatment may be effective, and the patient's risk level can be updated accordingly. The entire system adopts an edge-cloud collaborative mechanism, with the device completing basic diagnosis with low latency (e.g., <2 seconds), while the cloud handles complex cases, performs model iteration updates, and integrates multi-center data.

[0076] The Raman spectroscopy screening method, system, device, and medium provided by this invention, based on core principles (including sample preparation, adaptive dual-channel interference elimination processing, multi-scale biomarker feature extraction and aggregation, and risk assessment based on a lightweight deep learning model), can be extended to the detection or monitoring of other cardiovascular diseases. For example, by adjusting the multi-scale biomarker features of interest, the method can be used to assess the stability of atherosclerotic plaques, assist in the diagnosis and monitoring of heart failure, or evaluate the efficacy of cardiovascular drugs. These applications are all based on the core technological concept of extracting pure spectral signals from complex biological fluids and performing intelligent multi-scale feature fusion analysis.

[0077] This invention effectively solves the challenge of complex blood sample matrices through a pioneering "physical-data" dual-channel adaptive interference elimination architecture, enabling the direct extraction of high-quality spectral information from whole blood without relying on complex surface-enhanced Raman scattering (SERS) substrates. Simultaneously, the constructed "micro-meso-macro" multi-scale biomarker network systematically integrates various pathophysiological features such as the phenylalanine-tyrosine ratio (Phe-Tyr Ratio), cell-free DNA (cfDNA), and glutathione, significantly enhancing the information dimension and comprehensiveness of diagnosis. Furthermore, a lightweight inference engine designed specifically for point-of-care testing (POCT) scenarios successfully achieves high-precision, low-latency (<2 seconds) real-time diagnosis on resource-constrained edge devices through knowledge distillation and dynamic computational graph technology. Finally, a closed-loop clinical decision support system incorporating risk stratification, treatment monitoring, and edge-cloud collaborative mechanisms has been established, capable of meeting diverse application scenarios from emergency medicine and primary healthcare to home health monitoring.

[0078] The performance indicators of this invention were obtained through the following experimental design and verification methods: 1. Experimental Sample Configuration To evaluate the performance of this method, a total of 100 subjects were included. The experimental group consisted of 50 patients diagnosed with myocardial infarction by cardiac troponin (cTnI or cTnT) testing, including 28 cases of ST-segment elevation myocardial infarction and 22 cases of non-ST-segment elevation myocardial infarction. The control group consisted of 50 non-myocardial infarction subjects, including 30 healthy volunteers and 20 patients with other cardiovascular diseases (such as stable angina and hypertension), whose cTnI / cTnT levels were below the clinical diagnostic threshold. The total sample size of 100 subjects, with a 1:1 ratio of myocardial infarction to non-myocardial infarction participants, constituted a balanced dataset.

[0079] 2. Gold Standard and Verification Process High-sensitivity cardiac troponin I (cTnI) testing was used as the gold standard for diagnosis, with specific thresholds set according to the European Society of Cardiology Fourth Edition (ECC) definition of myocardial infarction (cTnI ≥ 0.04 ng / mL or cTnT ≥ 14 ng / L). During the validation process, venous blood samples were collected from all subjects simultaneously with Raman spectroscopy acquisition (within ± 30 minutes) for cTnI / cTnT immunoassay. Diagnosis was confirmed by two independent cardiologists based on cTnI / cTnT test results and clinical symptoms. To ensure objectivity, a blinded design was employed, meaning the Raman spectroscopy data analysts were completely unaware of the subjects' cTnI / cTnT test results and clinical diagnostic conclusions.

[0080] 3. Model Performance Validation Methods The robustness of the screening model constructed using this method was evaluated using 5-fold cross-validation. Specifically, 100 samples were randomly divided into 5 subsets, each containing 10 cases of myocardial infarction and 10 cases of non-myocardial infarction. In each round of validation, 4 subsets (80 cases in total) were selected sequentially for model training, and the remaining subset (20 cases) was used for testing. This process was repeated 5 times to ensure that each sample was tested once. The results of the 5-fold cross-validation were accumulated, and a confusion matrix was obtained. Based on this matrix, the overall diagnostic performance index was calculated. The accumulated results are shown in Table 1. Table 1. Cumulative results of five-fold cross-validation

[0081] In the table, TP represents a true positive, the number of correctly identified myocardial infarction patients; FN represents a false negative, the number of missed myocardial infarction patients; TN represents a true negative, the number of correctly excluded non-myocardial infarction subjects; and FP represents a false positive, the number of misdiagnosed non-myocardial infarction subjects.

[0082] The overall diagnostic performance is calculated as follows: Sensitivity = TP / (TP+FN) = 49 / (49+1) = 98.0%; Specificity = TN / (TN+FP) = 49 / (49+1) = 98.0%; Accuracy = (TP+TN) / (TP+TN+FP+FN) = (49+49) / 100 = 98.0%; Positive predictive value = TP / (TP+FP) = 49 / (49+1) = 98.0%; Negative predictive value = TN / (TN+FN) = 49 / (49+1) = 98.0%; To assess the statistical reliability of the performance indicators, the Wilson scoring method was used to calculate 95% confidence intervals. Based on 50 patients with myocardial infarction, 49 were correctly identified, with a sensitivity 95% CI of [89.4%, 99.7%]; based on 50 non-myocardial infarction subjects, 49 were correctly excluded, with a specificity 95% CI of [89.4%, 99.7%]. The Wilson scoring method is calculated as follows: ; in, This represents the lower / upper limit of the confidence interval. This is the lower limit of the confidence interval; This represents the upper limit of the confidence interval; To observe the proportion, the proportion of positive (or negative) samples in the sample (49 / 50=0.98 in this example), n is the sample size used for confidence interval calculation (50 in this example), and z is the standard normal distribution quantile, z=1.96 at 95% CI.

[0083] 4. Verification of interference cancellation effect To quantitatively evaluate the effect of the adaptive dual-channel interference elimination in the core step S3 of this invention, 10 additional samples from hyperlipidemic patients (whose blood matrix interference was particularly severe) were selected from the 100 samples for a comparative experiment. Four methods were used for processing: unprocessed raw spectra, EMSC processing only, AAE processing only, and the dual-channel adaptive processing proposed in this invention. The characteristic peak of phenylalanine (1000 cm⁻¹) was used as the reference. - ¹) signal-to-noise ratio ( =Peak height intensity / baseline noise standard deviation) is used as the evaluation index, and the results are compared in Table 2: Table 2 Comparison of signal-to-noise ratio of characteristic peaks under different spectral processing methods

[0084] Experimental results show that the dual-channel adaptive interference cancellation method of the present invention can significantly improve the characteristic peak signal-to-noise ratio from the baseline of 5.2 to 19.8, an improvement of 3.8 times, and its effect is significantly better than any single-channel processing method.

[0085] In summary, the Raman spectroscopy method for myocardial infarction screening provided by this invention requires only 0.1 μL of peripheral whole blood and no complex pretreatment. The average detection time from sample collection to result output is 4.2 ± 0.8 minutes. In cross-validation, it demonstrates excellent diagnostic performance with a sensitivity of 98.0% and a specificity of 98.0%. Furthermore, the unique algorithm architecture effectively improves signal quality, providing a complete and reliable technical solution for rapid, non-invasive, and accurate screening of myocardial infarction.

[0086] Further reference Figure 3 As an implementation of the above method, this invention also proposes an embodiment of a Raman spectroscopy myocardial infarction screening system. This system is based on a hierarchical, process-oriented architecture design, integrating the complete screening process into one unit, and can be applied to various electronic devices. For example... Figure 3 As shown, the overall system architecture is divided into a user interaction layer, a sample processing layer, a spectral acquisition layer, an algorithm processing layer, and an output decision layer from top to bottom. These layers work collaboratively to achieve full automation from sample to decision. This Raman spectroscopy myocardial infarction screening system specifically includes the following modules: The sample acquisition and processing module corresponds to the user interaction layer and sample processing layer in the system architecture. This module is configured to collect capillary whole blood samples and perform sample preparation. Specifically, it automatically and quantitatively collects approximately 0.1 μL of capillary whole blood from the fingertip using a device integrating a microfluidic chip, without centrifugation. The collected blood sample is then mixed with a surface-enhanced Raman scattering reagent (SERS) and drop-coated onto the central region of a hydrophobic silicon substrate, where it is dried under constant temperature and humidity conditions for approximately 60 seconds. This process utilizes the coffee ring effect to achieve physical interference separation: high-molecular-weight proteins migrate to the edges, while target low-molecular-weight biomarkers such as phenylalanine, tyrosine, and cell-free DNA (cfDNA) are enriched in the central region, resulting in a high-quality sample and laying the physical foundation for subsequent analysis.

[0087] The spectral acquisition module corresponds to the spectral acquisition layer in the system architecture. This module is configured to acquire Raman spectra of the central region of the sample to be tested. Its core components include a semiconductor laser with a wavelength of 785 nm and a power of 100 mW, and a laser covering a range of 400-1800 cm⁻¹. -¹A spectrometer with a wide spectral range. During acquisition, a high numerical aperture objective lens focuses the laser into a spot with a diameter of approximately 16 μm, precisely aligning it with the central region of the sample to be tested (i.e., the biomarker enrichment area). On each sample, the system automatically performs a 9-point scan, with an integration time of 3 seconds for each point, to ensure the acquisition of raw spectral data with a high signal-to-noise ratio, providing reliable input for algorithm processing.

[0088] The spectral processing and analysis module, corresponding to the algorithm processing layer in the system architecture, is the intelligent core of this system. This module is configured to execute a series of coherent algorithmic operations to transform the raw spectrum into high-dimensional features usable for decision-making. First, adaptive dual-channel interference cancellation processing is performed: a first corrected spectrum is obtained through a physical prior channel (using the EMSC algorithm combined with a standard interference spectral library), and a second corrected spectrum is obtained through a data-driven channel (using a pre-trained adversarial autoencoder (AAE) model to learn common interference patterns); then, the two channel results are dynamically fused based on the signal-to-noise ratio (SNR) of the original spectrum to obtain the pure biomarker spectrum, a process that can improve the signal-to-noise ratio by 3.8 times. Second, multi-scale feature extraction is performed from the pure spectrum: microscale molecular features (such as the Phe-Tyr ratio) and mesoscale cellular component features (such as cfDNA signal) are extracted, and a pathophysiological graph network is constructed. A graph convolutional network is applied for information aggregation, ultimately generating a 128-dimensional multi-dimensional feature vector that comprehensively represents the pathological state. Finally, the module integrates a lightweight inference engine (based on the MobileNetV3 architecture). This engine is trained using knowledge distillation technology and can perform fast and accurate calculations on the input multidimensional feature vectors. The latency for completing inference on the device is usually less than 2 seconds, and the accuracy is high.

[0089] The risk assessment and decision-making module corresponds to the output decision layer in the system architecture. This module receives the multidimensional feature vector generated by the spectral processing and analysis module, inputs it into the lightweight inference engine for processing, and ultimately outputs the risk probability of myocardial infarction. Based on this probability, the module performs a clear and actionable risk assessment classification (high risk, medium risk, low risk) and generates corresponding treatment recommendations and monitoring plans. For example, for a medium-risk result, the system will recommend retesting within 24-48 hours. In addition, this module supports treatment monitoring functions, assessing efficacy by longitudinally comparing changes in key biomarkers (such as the Phe-Tyr ratio). The entire system adopts an edge-cloud collaborative mechanism: the device end (edge) is responsible for low-latency basic diagnosis and real-time decision support; the cloud handles complex cases, performs model iteration updates, and integrates multi-center data, forming a complete closed-loop clinical decision support system.

[0090] In summary, through the close integration and collaborative work of the aforementioned modules, this system achieves a complete workflow from the collection of a 0.1μL capillary whole blood sample to the generation of a risk assessment report. The entire process takes less than 5 minutes, meeting the needs of real-time clinical diagnosis and providing a complete system solution integrating hardware, algorithms, and decision support for rapid, non-invasive, and accurate screening of myocardial infarction.

[0091] The present invention also proposes a terminal device, including a memory, a processor, and a computer program stored in the memory and capable of running on the processor, wherein the processor executes the computer program to implement the steps of any of the Raman spectroscopy myocardial infarction screening methods described above.

[0092] The present invention also proposes a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of any of the Raman spectroscopy myocardial infarction screening methods described above.

[0093] The following is for reference. Figure 4 It shows a schematic diagram of the structure of a computer system 400 suitable for implementing terminal devices or servers in the embodiments of this application. Figure 4 The terminal device or server shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0094] like Figure 4 As shown, the computer system 400 includes a central processing unit (CPU) 401, which can perform various appropriate actions and processes based on programs stored in read-only memory (ROM) 402 or programs loaded from storage section 408 into random access memory (RAM) 403. The RAM 403 also stores various programs and data required for the operation of the computer system 400. The CPU 401, ROM 402, and RAM 403 are interconnected via a bus 404. An input / output (I / O) interface 405 is also connected to the bus 404.

[0095] The following components are connected to I / O interface 405: input section 406 including keyboard, mouse, etc.; output section 407 including liquid crystal display (LCD) and speakers, etc.; storage section 408 including hard disk, etc.; and communication section 409 including network interface card such as LAN card, modem, etc. Communication section 409 performs communication processing via a network such as the Internet. Drive 410 is also connected to I / O interface 405 as needed. Removable media 411, such as disk, optical disk, magneto-optical disk, semiconductor memory, etc., are installed on drive 410 as needed so that computer programs read from them can be installed into storage section 408 as needed.

[0096] Specifically, according to embodiments of this disclosure, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this disclosure include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication section 409, and / or installed from removable medium 411. When the computer program is executed by central processing unit (CPU) 401, it performs the functions defined in the methods of this application. It should be noted that the computer-readable medium described in this application can be a computer-readable signal medium or a computer-readable medium or any combination thereof. The computer-readable medium can be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this application, a computer-readable medium can be any tangible medium containing or storing a program that can be used by or in connection with an instruction execution system, apparatus, or device. In this application, a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium can also be any computer-readable medium other than a computer-readable medium that can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. Program code contained on a computer-readable medium can be transmitted using any suitable medium, including but not limited to: wireless, wire, optical fiber, RF, etc., or any suitable combination thereof.

[0097] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as C or similar languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0098] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0099] The above description is merely a preferred embodiment of this application and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of the invention involved in this application is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-described inventive concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this application.

Claims

1. A Raman spectroscopy method for screening myocardial infarction, characterized in that, Includes the following steps: S1. Collect peripheral whole blood samples, mix the peripheral whole blood samples with surface-enhanced Raman scattering reagent, and then perform drop-coating deposition and drying on the substrate to prepare the test sample using the coffee ring effect; S2. Raman spectroscopy is performed on the central region of the sample to be tested to obtain raw spectral data; S3. Perform adaptive dual-channel interference elimination processing on the original spectral data to obtain a pure biomarker spectrum. The adaptive dual-channel interference elimination processing includes: correcting the original spectral data using a physical prior method to obtain a first corrected spectrum, correcting the original spectral data using a data-driven method to obtain a second corrected spectrum, and fusing the first corrected spectrum and the second corrected spectrum. S4. Extract multi-scale biomarker features from the pure biomarker spectrum, and aggregate the multi-scale biomarker features to generate a multi-dimensional feature vector. S5. Input the multidimensional feature vector into a lightweight deep learning model for processing to obtain the risk probability of myocardial infarction, and perform risk assessment and classification based on the risk probability.

2. The Raman spectroscopy method for screening myocardial infarction according to claim 1, characterized in that, Step S3 involves adaptive dual-channel interference cancellation processing of the original spectral data, including the following sub-steps: S31. The original spectral data is processed through a physical prior channel, wherein an extended multiplicative signal correction algorithm is used to remove known matrix interference from the original spectral data based on a standard interference spectral library to obtain the first corrected spectrum; S32. The original spectral data is processed through a data-driven channel, wherein a pre-trained adversarial autoencoder model is used to extract and remove interference patterns from the original spectral data to obtain the second corrected spectrum. S33. Based on the signal-to-noise ratio of the original spectral data, dynamically fuse the first corrected spectrum and the second corrected spectrum to obtain the pure biomarker spectrum.

3. The Raman spectroscopy method for screening myocardial infarction according to claim 2, characterized in that, In step S32, the training process of the pre-trained adversarial autoencoder model includes the following sub-steps: S321. Extract latent variables from the original spectral data using an encoder; S322. Reconstruct the interference spectrum based on the latent variables using a generator; S323. The discriminator forces the distribution of the latent variables extracted by the encoder to approximate the preset prior distribution. S324. Construct a total loss function based on the reconstruction loss function, the generator adversarial loss function, and the discriminator loss function, and train the adversarial autoencoder model by optimizing the total loss function.

4. The Raman spectroscopy method for screening myocardial infarction according to claim 3, characterized in that, In step S324, the expression for the reconstruction loss function is: ； The expression for the discriminator loss function is: ； The expression for the generator adversarial loss function is: ； The expression for the total loss function is: ； in, To reconstruct the loss function; This represents the number of samples in the training batch. For the first One set of raw spectral data; The first reconstructed model of the adversarial autoencoder One interference spectrum; The discriminator loss function; For the first sampled from the standard Gaussian distribution N(0,I) A priori code; For the generator adversarial loss function, For discriminator networks; The encoder for the adversarial autoencoder model; This is the total loss function; These are the weighting coefficients of the discriminator loss function; These are the weighting coefficients for the generator's adversarial loss function.

5. The Raman spectroscopy method for screening myocardial infarction according to claim 2, characterized in that, In step S33, based on the signal-to-noise ratio of the original spectral data, the first corrected spectrum and the second corrected spectrum are dynamically fused to obtain the pure biomarker spectrum, including the following sub-steps: S331. Calculate the signal-to-noise ratio of the original spectral data; S332. Based on the calculated signal-to-noise ratio value, dynamically allocate the weight coefficients of the physical prior channel and the data-driven channel; S333. The first corrected spectrum and the second corrected spectrum are weighted and fused according to the weighting coefficients.

6. The Raman spectroscopy method for screening myocardial infarction according to claim 1, characterized in that, In step S4, multi-scale biomarker features are extracted from the purified biomarker spectrum, and the multi-scale biomarker features are aggregated to generate a multi-dimensional feature vector, including the following sub-steps: S41. Extract microscale molecular features from the pure biomarker spectrum, wherein the microscale molecular features include calculating the ratio of the characteristic peak intensity of phenylalanine to the characteristic peak intensity of tyrosine. S42. Extract mesoscale cellular component characteristics from the purified biomarker spectrum; S43. Construct a pathophysiological graph network based on the microscopic molecular features and the mesoscopic cellular component features, and apply a graph convolutional network to aggregate information from the pathophysiological graph network to generate the multidimensional feature vector.

7. The Raman spectroscopy method for screening myocardial infarction according to claim 1, characterized in that, Step S5 involves inputting the multidimensional feature vector into a lightweight deep learning model for processing to obtain the risk probability of myocardial infarction, including the following sub-steps: S51. Input the multidimensional feature vector into the lightweight deep learning model, specifically a lightweight student model obtained through knowledge distillation training. S52. The lightweight student model adaptively selects an internal calculation path based on the signal-to-noise ratio of the multidimensional feature vector, wherein: when the signal-to-noise ratio is higher than the first threshold, a simplified calculation path is enabled; when the signal-to-noise ratio is between the first threshold and the second threshold, a standard calculation path is enabled; and when the signal-to-noise ratio is not higher than the second threshold, a data reprocessing mechanism is triggered. S53. The lightweight student model outputs the probability of myocardial infarction. S54. Calculate the confidence level based on the risk probability. The confidence level assessment includes: calculating the prediction entropy of the risk probability and determining whether the prediction entropy is greater than a preset threshold. If it is greater, mark the risk probability or the risk assessment level derived from it as a low confidence result and trigger the review process.

8. A Raman spectroscopy myocardial infarction screening system for implementing the method as described in any one of claims 1 to 7, characterized in that, The system includes: The sample collection and processing module is configured to collect peripheral whole blood samples, mix the peripheral whole blood samples with surface-enhanced Raman scattering reagents, perform drop-coating deposition and drying on a substrate, and prepare the test sample by utilizing the coffee ring effect; The spectral acquisition module is configured to acquire Raman spectra of the central region of the sample to be tested, thereby obtaining raw spectral data. The spectral processing and analysis module is configured to perform the following operations: The original spectral data is subjected to adaptive dual-channel interference elimination processing to obtain a pure biomarker spectrum. The adaptive dual-channel interference elimination processing includes: correcting the original spectral data using a physical prior method to obtain a first corrected spectrum, correcting the original spectral data using a data-driven method to obtain a second corrected spectrum, and fusing the first corrected spectrum and the second corrected spectrum. Multi-scale biomarker features are extracted from the pure biomarker spectrum, and the multi-scale biomarker features are aggregated to generate a multi-dimensional feature vector. The risk assessment and decision-making module is configured to input the multidimensional feature vector into a lightweight deep learning model for processing, obtain the risk probability of myocardial infarction, and perform risk assessment and classification based on the risk probability.

9. A terminal device, comprising a memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the Raman spectroscopy myocardial infarction screening method as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the Raman spectroscopy myocardial infarction screening method as described in any one of claims 1 to 7.

Images