OSA grading method and system based on multidimensional marker instant signal denoising and physiological adversarial network

Abstract

The invention provides an OSA grading method and system based on multidimensional marker instant signal denoising and a physiological adversarial network, and relates to the technical field of medical instrument data processing and artificial intelligence auxiliary diagnosis. The method comprises the following steps: acquiring an original optical absorption signal of a multi-dimensional marker and a synchronous environment temperature and humidity sequence; inputting the data into a thermodynamic-optical compensation model guided by physical information to carry out bottom noise deduction and denoising to obtain high-fidelity marker quantitative data; physiologically decoupling the high-fidelity data and clinical baseline features, respectively mapping the high-fidelity data and clinical baseline features into a mechanical hypoxia feature group and an inflammatory barrier feature group, and inputting the mechanical hypoxia feature group and the inflammatory barrier feature group into a double-flow network; a pathological confrontation cross gating unit is used for simulating the human body confrontation process that negative pressure is sucked in the chest to break through an anti-reflux barrier, nonlinear game fusion is carried out, and a grading result of Obstructive Sleep Apnea (OSA) and the risk of concurrent chronic reflux cough are output. According to the method, the problem that the basic layer detection is easy to distort is solved, and the technical effects of high interpretability and accurate diagnosis of the OSA co-disease mechanism are realized.

Description

Technical Field

[0001] This invention relates to the fields of medical device data processing and artificial intelligence-assisted diagnosis technology, and in particular to an OSA classification method and system based on real-time signal denoising of multidimensional biomarkers and physiological adversarial networks. Background Technology

[0002] Obstructive sleep apnea (OSA) complicated by reflux chronic cough (GERC) is a complex anatomical and physiological syndrome with a high incidence and misdiagnosis rate in clinical practice. Its core pathophysiological mechanism is as follows: During sleep, OSA patients experience complete or partial collapse of the upper airway, leading to a strong inspiratory effort to overcome airway resistance, resulting in significant negative thoracic pressure within the pleural cavity. Simultaneously, systemic inflammatory response causes a decrease in lower esophageal sphincter (LES) tone and damage to the airway's protective barrier. When the "suction force" generated by the negative thoracic pressure exceeds the LES's anti-reflux barrier resistance, gastric contents reflux into the pharynx, stimulating the already hypersensitive airway and inducing a severe cough.

[0003] Currently, various AI-assisted diagnostic models (such as Multilayer Perceptron (MLP), Support Vector Machine (SVM), and conventional Graph Neural Networks) have been attempted to process multidimensional biochemical and clinical data related to OSA. However, these existing technologies have the following significant drawbacks: First, existing models generally adopt a "purely data-driven" black-box approach, only splicing statistical features and completely ignoring the antagonistic physical mechanisms between the "negative pressure suction force in the thoracic cavity" and the "esophageal sphincter anti-reflux barrier" in the human body. This modeling approach, which lacks medical interpretability, cannot explain the causes of reflux risk to clinicians, nor can it provide targeted treatment recommendations (such as adjusting CPAP treatment pressure or choosing anti-inflammatory drugs).

[0004] Second, in primary healthcare settings, point-of-care testing (POCT) devices used to detect markers of inflammation and hypoxia (such as anti-IL-6Ra antibodies and oxidized low-density lipoprotein ox-LDL) are susceptible to changes in ambient temperature and humidity, leading to severe baseline drift in the optical colloidal gold signal. Existing AI models typically input these small signals with significant background noise directly into the network, making them prone to overfitting and exhibiting poor generalization ability under conditions of small sample data in primary healthcare settings.

[0005] Therefore, there is an urgent need for an OSA grading and complication risk assessment method that can overcome environmental noise interference and has a clear pathological mechanism that can be interpreted. This application addresses these technical problems by proposing an OSA grading method and system based on multidimensional biomarker real-time signal denoising and a physiological adversarial network. Summary of the Invention

[0006] The purpose of this application is to provide an OSA classification method, system, electronic device, and computer-readable storage medium based on multidimensional biomarker real-time signal denoising and physiological adversarial network, which can achieve the technical effects of improving the medical interpretability of disease prediction, overcoming the background noise of primary POCT equipment, and improving the accuracy of diagnosis of OSA and reflux complications.

[0007] Firstly, this application provides an OSA classification method based on multidimensional biomarker instantaneous signal denoising and physiological adversarial networks, including: The original optical absorption signal sequence of multidimensional markers obtained by scanning multiple colloidal gold test strips with a grassroots instant detection device was acquired, and the temperature and humidity sequences of the detection environment were acquired simultaneously. The original optical absorption signal sequence, the temperature sequence, and the humidity sequence are input into a pre-trained physical information-guided thermodynamic-optical compensation model to perform noise floor subtraction and cross-modal denoising to obtain high-fidelity quantitative data of biomarkers. Based on the pathological and anatomical mechanism of obstructive sleep apnea (OSA) complicated with reflux chronic cough, the high-fidelity biomarker quantitative data and the patient's baseline clinical characteristics are physiologically decoupled and mapped to a mechanical hypoxia feature group characterizing intrathoracic negative pressure and an inflammatory barrier feature group characterizing tissue defense. The data are then input into a pre-constructed two-stream network to extract independent latent variable potential energy features. By using the pathological adversarial cross-gating unit built into the dual-flow network, the latent variable potential energy characteristics are subjected to nonlinear game fusion to simulate the physical adversarial mechanism of human pathology, and the severity classification results of obstructive sleep apnea (OSA) and the risk assessment probability of combined reflux chronic cough are output.

[0008] In the aforementioned implementation process, this method creatively abandons the traditional flat data mining model and introduces "bionic anatomical structure design" into the neural network topology for the first time. The front end uses physical equations to remove the signal noise of temperature and humidity in the grassroots environment, while the back end cleverly decouples biomarker and vital sign data into "mechanical hypoxia-driven flow" and "inflammatory barrier damage flow". The "pathological adversarial cross-gating unit" in the network directly simulates the human physical process of "negative pressure suction overcoming the defense threshold" at the mathematical level, thereby achieving technical effects of extremely high clinical interpretability and cross-sample generalization diagnostic accuracy.

[0009] Furthermore, the multidimensional biomarkers include at least an anti-IL-6Ra antibody representing immune imbalance and oxidized low-density lipoprotein (ox-LDL) representing lipid peroxidation damage; the steps of noise reduction and cross-modal denoising include: extracting theoretical drift characteristics based on the thermodynamic kinetic equation of colloidal gold particle aggregation, as shown in the formula: Obtain the baseline-removed residual signal: The baseline-removed residual signal As a query vector matrix, environmental disturbance features are mapped to key and value vector matrices, cross-modal self-attention calculation is performed, and the high-fidelity biomarker quantitative data is output.

[0010] In the above implementation process, since the colloidal gold antigen-antibody binding reaction of the basic micro-detection equipment is easily affected by the ambient temperature and humidity, the background optical drift can be calculated from the physical source through the preset thermodynamic kinetic equation, and the high-frequency noise pulse can be effectively suppressed through the cross-modal self-attention mechanism, thereby ensuring that the data input to the medical AI model is a high-fidelity quantitative result.

[0011] Furthermore, the step of physiologically decoupling the high-fidelity biomarker quantitative data from the patient's baseline clinical characteristics includes: combining the patient's body mass index (BMI), neck circumference parameters, and high-fidelity ox-LDL quantitative data to construct the mechanical hypoxia characteristic group. The Leicester Cough Questionnaire (LCQ) score, DeMeester reflux severity score, and high-fidelity anti-IL-6Ra antibody quantitative data of the patients were combined to construct the inflammatory barrier characteristic group. .

[0012] In the above process, the system performed physiological targeting decoupling based on human pathological pathways. Obesity (high BMI, large neck circumference) and ox-LDL represent the structures that generate "huge negative pressure in the thoracic cavity" and the driving force of hypoxia; while LCQ, DeMeester and anti-IL-6Ra antibodies represent "defense vulnerability" due to loss of lower esophageal sphincter tone and airway hypersensitivity.

[0013] Furthermore, the step of extracting independent latent variable potential features from the input to the pre-constructed two-stream network includes: The input is fed into the self-attention encoder of the hypoxia-driven flow module, and the output is a thoracic aspiration potential energy feature vector. ;Will The input is fed into the multilayer perceptron feature extraction layer of the inflammatory barrier flow module, and the output is a feature vector of vulnerability to backflow barrier. .

[0014] In the above implementation process, a dual-flow path with mutual isolation is used to extract features, which avoids the confusion of physical meaning caused by early feature fusion, so that the latent variable vector learned by the model strictly corresponds to the two independent medical mechanics concepts of "pump suction" and "valve relaxation".

[0015] Furthermore, in the pathological adversarial cross-gating unit, the nonlinear game-theoretic fusion logic formula simulating the "negative pressure suction effect in the pleural cavity" breaking through the "esophageal sphincter anti-reflux barrier" is expressed as: the probability of breaking through with combined reflux complications: .

[0016] In the above implementation process, this is the most core and original computational unit of the algorithm of this invention. The formula cleverly uses... Represents the simulated negative pressure suction force, using This represents absolute defensive tension. The difference within the parentheses is positive if and only if the suction force overcomes the resistance force, and the Sigmoid function instantly saturates to output a high probability of backflow risk. This completely aligns the internal activation logic of AI with the pathological anatomical mechanisms of the human body.

[0017] Furthermore, the step of outputting the severity grading results of obstructive sleep apnea (OSA) includes: constructing a game-theoretic feature fusion representation: ; The data is then input into a multilayer perceptron classifier, which outputs the classification probability distribution of the patient's mild, moderate, or severe OSA through the Softmax function.

[0018] In the above implementation process, the vicious cycle mechanism between mechanical obstruction and inflammation is metaphorically represented by cooperative multiplication, so as not only to predict the risk of complications, but also to accurately classify the severity level of OSA itself (AHI classification).

[0019] Secondly, this application provides an OSA grading system based on multidimensional biomarker instantaneous signal denoising and a physiological adversarial network, comprising: a signal base layer acquisition module, used to acquire the original optical absorption signal sequence of the multidimensional biomarker obtained by scanning multiple colloidal gold test strips with a base layer instantaneous detection device, and simultaneously acquire the temperature and humidity sequences of the detection environment; a physical compensation denoising module, used to input the original optical absorption signal sequence, the temperature sequence, and the humidity sequence into a pre-trained physical information-guided thermodynamic-optical compensation model to perform background noise reduction and cross-modal denoising to obtain high-fidelity biomarker quantitative data; pathogenesis analysis The dual-flow module is used to physiologically decouple the high-fidelity biomarker quantitative data and the patient's baseline clinical characteristics, mapping them into a mechanical hypoxia feature group characterizing intrathoracic negative pressure and an inflammatory barrier feature group characterizing tissue defense, and inputting them into a pre-constructed dual-flow network to extract independent latent variable potential energy features; the adversarial game decision module is used to perform nonlinear game fusion of the latent variable potential energy features through the pathological adversarial cross-gating unit built into the dual-flow network to simulate the physical adversarial mechanism of human pathology, and output the severity classification result of obstructive sleep apnea (OSA) and the risk assessment probability of comorbid chronic cough.

[0020] Thirdly, this application provides an electronic device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the steps of the method described in any of the first aspects.

[0021] Fourthly, this application provides a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the method described in any of the first aspects.

[0022] Fifthly, this application provides a computer program product that, when run on a computer, causes the computer to perform the method described in any of the first aspects.

[0023] The technical effects and advantages of this invention are as follows: (1) An adversarial feature fusion method based on pathological anatomy mechanism was introduced into the neural network structure, encoding the physical adversarial relationship between "thoracic negative pressure suction force" and "esophageal sphincter anti-reflux barrier" as a computable game unit. This design enables the model to provide intermediate feature explanations corresponding to the pathophysiological mechanism when outputting OSA grading and reflux risk assessment results, which helps clinicians understand the basis of the model's decision-making.

[0024] (2) To address the issue that optical signals in grassroots real-time detection equipment are easily affected by ambient temperature and humidity, a denoising method combining thermodynamic kinetic equations and cross-modal self-attention is proposed. This method subtracts theoretical background drift through a physical model and compensates for it using environmental disturbance characteristics. It can improve the signal-to-noise ratio of quantitative biomarker data without additional hardware modifications and reduce the interference of environmental factors on subsequent diagnostic models.

[0025] (3) By decoupling multidimensional biomarkers from clinical features into mechanical hypoxia feature group and inflammatory barrier feature group, and using a two-stream network to extract latent variables separately, the confusion of physical meaning caused by early feature fusion is avoided. This design makes the model learn the two different pathological pathways relatively independently, which helps to improve the model's generalization ability in primary care data with limited sample size.

[0026] (4) It simultaneously outputs the severity level of OSA and the probability of combined chronic cough with reflux, providing two dimensions of auxiliary diagnostic information for clinical use without the need for additional testing items or equipment. Attached Figure Description

[0027] Figure 1 A flowchart illustrating an OSA classification method based on real-time signal denoising of multidimensional markers and physiological adversarial networks provided in this application embodiment; Figure 2 A detailed flowchart illustrating another OSA hierarchical method provided in this application embodiment; Figure 3 This is a structural block diagram of the OSA hierarchical system provided in an embodiment of this application. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be described below with reference to the accompanying drawings. It should be noted that similar reference numerals and letters in the following drawings indicate similar items; therefore, once an item is defined in one drawing, it does not need to be further defined and explained in subsequent drawings. Furthermore, in the description of this application, the terms "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0029] Generally, while current conventional AI-assisted medical diagnostic systems perform well with standardized data from large tertiary hospitals, they often lack a deep understanding of the underlying causal relationships in complex pathophysiological mechanisms, particularly when explaining the specific causes of reflux cough, a comorbidity of OSA. Furthermore, traditional methods often overlook signal noise issues in data acquisition from primary care devices. This application aims to address these problems through biomimetic anatomical network topology.

[0030] Example 1: OSA Classification Method Please see Figure 1 , Figure 1 This is a flowchart illustrating an OSA classification method based on multidimensional biomarker instantaneous signal denoising and physiological adversarial networks, provided in an embodiment of this application. The method includes steps S100 to S400.

[0031] Step S100: Acquire the original optical absorption signal of the multidimensional marker and the synchronous temperature and humidity sequence. For example, in primary healthcare institutions, a portable POCT device scans a colloidal gold test strip after a patient's serum has been added, obtaining a trace amount of light signal containing anti-IL-6Ra antibody and ox-LDL. Simultaneously, the ambient temperature T(t) and humidity H(t) inside the device are recorded.

[0032] Step S200: Denoising using a thermodynamic-optical compensation model guided by physical information. This step specifically includes: extracting theoretical drift characteristics based on the thermodynamic kinetic equations of colloidal gold particle aggregation, the formula being: Obtain the baseline-removed residual signal: ; in Indicates time The theoretical background of optical drift, and They represent time respectively Temperature and humidity, This represents the apparent activation energy of a combination reaction. Let be the ideal gas constant. , , All are physical constant matrices. This indicates the instrument's inherent dark current noise; Subsequently, the baseline-de-baseline residual signal As the query vector matrix, environmental perturbation features are mapped to key and value vector matrices, and cross-modal self-attention computation is performed to output high-fidelity biomarker quantitative data. This step subtracts theoretical background drift through a physical model and compensates for it using environmental perturbation features, thereby obtaining pure biochemical quantitative data and preventing subsequent networks from mislearning temperature and humidity fluctuations as disease features.

[0033] Step S300: Decouple the data from physiology and input it into a two-stream network based on anatomical and physiological constraints to extract features. This step differs from the traditional approach of processing all features together. Instead, it is based on the pathological and anatomical mechanisms of OSA complicated by reflux chronic cough, physiologically decoupling high-fidelity biomarker quantitative data and patient baseline clinical characteristics, mapping them into two feature sets: Mechanical hypoxia characteristic group The study combined patients' body mass index (BMI), neck circumference parameters, and high-fidelity ox-LDL quantitative data. Obesity is a physical cause of upper airway collapse, and ox-LDL characterizes the depth of chronic intermittent hypoxia.

[0034] Inflammatory barrier feature group The Lester Cough Questionnaire (LCQ) score, DeMeester reflux severity score, and high-fidelity anti-IL-6Ra antibody quantitative data were combined. High expression of anti-IL-6Ra antibody suggests uncontrolled systemic inflammation, which may damage smooth muscle function and lead to esophageal sphincter relaxation.

[0035] Then, The self-attention encoder, input to the hypoxia-driven flow module in the two-stream network, outputs a pleural suction potential energy feature vector characterizing the high-frequency, large negative pressure generated in the pleural cavity during nocturnal apnea. .Will The input is fed into the multilayer perceptron (MLP) feature extraction layer of the inflammatory barrier flow module, and the output is an anti-reflux barrier vulnerability feature vector characterizing the degree of damage to the physical anti-reflux barrier in the digestive tract. .

[0036] Step S400: Perform feature game and adaptive hierarchical output through pathological adversarial cross-gating units. The dual-flow network incorporates pathological adversarial cross-gating units to simulate the nonlinear game fusion of the "negative pressure suction effect in the pleural cavity" overcoming the "esophageal sphincter anti-reflux barrier." Its logical formula is expressed as follows: Probability of recovery from complications of combined reflux: in, and The weight matrix is ​​a learnable matrix. This indicates the absolute defensive tensile strength of the anti-backflow barrier. This is a parameter for individualized pathological threshold bias; when it represents the simulated negative pressure suction force. Greater than represents the simulated barrier's defensive power When the difference within the brackets is positive, the activation function outputs a high-risk probability close to 1; otherwise, it outputs a low-risk probability. In addition, the steps for outputting OSA severity rating results include: constructing a game feature fusion representation: in, Indicates feature splicing, Representing element-wise collaborative multiplication; fusing the game-theoretic features into a representation. The input is fed into a multilayer perceptron classifier, which outputs the classification probability distribution of the patient's mild, moderate, or severe OSA through the Softmax function.

[0037] Example 2: OSA Hierarchy System Please see Figure 3 This application also provides an OSA classification system based on multidimensional biomarker instantaneous signal denoising and physiological adversarial network, the system comprising: Signal acquisition module (100): used to acquire the original optical absorption signal sequence of multidimensional markers obtained by scanning multiple colloidal gold test strips by the base-level instant detection device, and simultaneously acquire the temperature and humidity sequences of the detection environment.

[0038] Physical compensation denoising module (200): used to input the original optical absorption signal sequence, the temperature sequence and the humidity sequence into a pre-trained physical information-guided thermodynamic-optical compensation model to perform noise floor subtraction and cross-modal denoising to obtain high-fidelity biomarker quantitative data.

[0039] The pathological decoupling dual-flow module (300) is used to physiologically decouple the high-fidelity biomarker quantitative data and the patient's baseline clinical characteristics, mapping them into a mechanical hypoxia feature group characterizing intrathoracic negative pressure and an inflammatory barrier feature group characterizing tissue defense, and inputting them into a pre-constructed dual-flow network to extract independent latent variable potential features.

[0040] Adversarial game decision module (400): Used to perform nonlinear game fusion of the latent variable potential energy characteristics through the pathological adversarial cross-gating unit built into the dual-stream network to simulate the physical adversarial mechanism of human pathology, and output the severity classification result of obstructive sleep apnea (OSA) and the risk assessment probability of combined reflux chronic cough.

[0041] The specific implementation methods of the above modules correspond to steps S100 to S400 in Embodiment 1, and will not be repeated here.

[0042] Example 3: Electronic Equipment This application also provides an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the OSA classification method based on multidimensional biomarker instantaneous signal denoising and physiological adversarial networks as described in Embodiment 1.

[0043] Example 4: Storage Medium This application also provides a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the OSA classification method based on multidimensional marker instantaneous signal denoising and physiological adversarial network as described in Embodiment 1.

[0044] Example 5: Computer Program Product This application also provides a computer program product that, when run on a computer, causes the computer to execute the OSA classification method based on multidimensional marker instantaneous signal denoising and physiological adversarial network as described in Embodiment 1.

[0045] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. An OSA classification method based on multidimensional biomarker instantaneous signal denoising and physiological adversarial networks, characterized in that, Includes the following steps: The original optical absorption signal sequence of multidimensional markers obtained by scanning multiple colloidal gold test strips with a grassroots instant detection device was acquired, and the temperature and humidity sequences of the detection environment were acquired simultaneously. The original optical absorption signal sequence, the temperature sequence, and the humidity sequence are input into a pre-trained physical information-guided thermodynamic-optical compensation model to perform noise floor subtraction and cross-modal denoising to obtain high-fidelity quantitative data of biomarkers. Based on the pathological and anatomical mechanism of obstructive sleep apnea complicated with reflux chronic cough, the high-fidelity biomarker quantitative data and the patient's baseline clinical characteristics are physiologically decoupled and mapped to a mechanical hypoxia feature group characterizing intrathoracic negative pressure and an inflammatory barrier feature group characterizing tissue defense. The data are then input into a pre-constructed two-stream network to extract independent latent variable potential energy features. By using the pathological adversarial cross-gating unit built into the dual-flow network, the latent variable potential energy characteristics are subjected to nonlinear game fusion to simulate the physical adversarial mechanism of human pathology, and the severity classification results of obstructive sleep apnea and the risk assessment probability of combined reflux chronic cough are output.

2. The OSA classification method based on multidimensional biomarker instantaneous signal denoising and physiological adversarial network according to claim 1, characterized in that, The multidimensional biomarkers include at least an anti-IL-6Ra antibody representing immune imbalance and oxidized low-density lipoprotein (ox-LDL) representing lipid peroxidation damage; the step of inputting the original optical absorption signal sequence, the temperature sequence, and the humidity sequence into a pre-trained physical information-guided thermodynamic-optical compensation model for noise floor subtraction and cross-modal denoising includes: The theoretical drift characteristics are extracted based on the thermodynamic kinetic equations of colloidal gold particle aggregation, and the formula is as follows: Obtain the baseline-removed residual signal: ; in, Indicates time The theoretical background of optical drift, and They represent time respectively Temperature and humidity, This represents the apparent activation energy of a combination reaction. Let be the ideal gas constant. , , All are physical constant matrices. This represents the instrument's inherent dark current noise; the baseline-de-baseline residual signal As a query vector matrix, environmental disturbance features are mapped to key and value vector matrices, cross-modal self-attention calculation is performed, and the high-fidelity biomarker quantitative data is output.

3. The OSA classification method based on multidimensional biomarker instantaneous signal denoising and physiological adversarial network according to claim 1, characterized in that, The step of physiologically decoupling the high-fidelity biomarker quantitative data from the patient's baseline clinical characteristics includes: The patient's body mass index, neck circumference parameters, and high-fidelity ox-LDL quantitative data were combined to construct the mechanical hypoxia characteristic group. It is used to characterize the degree of mechanical stenosis of the upper airway and the depth of chronic intermittent hypoxia; The Lester Cough Questionnaire (LCQ) score, DeMeester reflux severity score, and high-fidelity anti-IL-6Ra antibody quantitative data of the patients were combined to construct the inflammatory barrier characteristic group. It is used to characterize the degree of loss of lower esophageal sphincter tone and the hypersensitive state of airway mucosal inflammation.

4. The OSA classification method based on multidimensional biomarker instantaneous signal denoising and physiological adversarial network according to claim 3, characterized in that, The dual-flow network includes a hypoxia-driven flow module and an inflammatory barrier flow module; The step of extracting independent latent variable potential features from the input to the pre-constructed two-stream network includes: Will The input is fed into the self-attention encoder of the hypoxia-driven flow module, and the output is a pleural suction potential energy feature vector characterizing the high-frequency, large negative pressure generated in the pleural cavity during nocturnal apnea. ; Will The input is fed into the multilayer perceptron feature extraction layer of the inflammatory barrier flow module, and the output is an anti-reflux barrier vulnerability feature vector characterizing the degree of damage to the digestive tract's anti-reflux physical barrier. .

5. The OSA classification method based on multidimensional biomarker instantaneous signal denoising and physiological adversarial network according to claim 4, characterized in that, In the aforementioned pathological adversarial cross-gating unit, the nonlinear game-theoretic fusion logic formula simulating the "negative pressure suction effect in the pleural cavity" breaking through the "esophageal sphincter anti-reflux barrier" is expressed as follows: Probability of recovery from complications of combined reflux: in, and The weight matrix is ​​a learnable matrix. This indicates the absolute defensive tensile strength of the anti-backflow barrier. This is a parameter for individualized pathological threshold bias; when it represents the simulated negative pressure suction force. Greater than represents the simulated barrier's defensive power When the activation function outputs a high-risk probability that is close to 1, the activation function outputs a high-risk probability that is close to 1.

6. The OSA classification method based on multidimensional biomarker instantaneous signal denoising and physiological adversarial network according to claim 4, characterized in that, The step of outputting the severity grading results of obstructive sleep apnea (OSA) includes: Constructing a game feature fusion representation: in, Indicates feature splicing, Representing element-wise collaborative multiplication; fusing the game-theoretic features into a representation. The data is input into a multilayer perceptron classifier, which outputs the classification probability distribution of the patient's mild, moderate, or severe OSA through the Softmax function.

7. An OSA hierarchical system based on multidimensional biomarker instantaneous signal denoising and physiological adversarial network, characterized in that, include: The signal acquisition module is used to acquire the original optical absorption signal sequence of multidimensional markers obtained by scanning multiple colloidal gold test strips with the basic real-time detection device, and simultaneously acquire the temperature and humidity sequences of the detection environment. The physical compensation and denoising module is used to input the original optical absorption signal sequence, the temperature sequence, and the humidity sequence into a pre-trained physical information-guided thermodynamic-optical compensation model to perform noise floor subtraction and cross-modal denoising to obtain high-fidelity quantitative data of biomarkers. The pathological decoupling dual-stream module is used to physiologically decouple the high-fidelity biomarker quantitative data and the patient's baseline clinical characteristics, mapping them into a mechanical hypoxia feature group characterizing intrathoracic negative pressure and an inflammatory barrier feature group characterizing tissue defense, and inputting them into a pre-constructed dual-stream network to extract independent latent variable potential energy features. The adversarial game decision-making module is used to perform nonlinear game fusion of the latent variable potential energy characteristics through the pathological adversarial cross-gating unit built into the dual-stream network to simulate the physical adversarial mechanism of human pathology, and output the severity classification result of obstructive sleep apnea and the risk assessment probability of combined reflux chronic cough.

8. The OSA classification system based on multidimensional biomarker instantaneous signal denoising and physiological adversarial network according to claim 7, characterized in that, The adversarial game decision-making module is specifically used for: Through formula The calculations are performed, where the probability of reflux complications breaking through is output when the latent variable representing the simulated negative pressure suction force is greater than the latent variable representing the simulated barrier defense force; and a game-theoretic feature fusion representation is constructed through collaborative multiplication and feature concatenation. The OSA severity rating results are output using the Softmax classification head.

9. An electronic device, characterized in that, include: The memory, the processor, and the computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the steps of the OSA classification method based on multidimensional marker instantaneous signal denoising and physiological adversarial network as described in any one of claims 1 to 6.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores instructions that, when executed on a computer, cause the computer to perform the OSA classification method based on multidimensional marker instantaneous signal denoising and physiological adversarial network as described in any one of claims 1 to 6.

Images