Method for realizing exercise health management based on individualized cardiopulmonary function digital twinborn model

Abstract

The invention discloses a method for realizing exercise health management based on an individualized cardiopulmonary function digital twinborn model. According to the method, CPET (cardiopulmonary function exercise test) data of a user is acquired as a reference and is input into a cardiopulmonary function digital twinborn model, so that the individual cardiopulmonary function limit and sudden acute disease risk prediction of the user is generated, and the purpose of individualized exercise health management is achieved. According to the technical scheme, firstly, a partial differential equation set is adopted to describe an oxygen metabolism chain of respiration-circulation-functional organs, then a digital twinning neural network is used for fitting the metabolism chain, individual characteristics and exercise load parameters are input, and an oxygen metabolism result and human body internal functional characteristic parameters are output; furthermore, the output of the digital twin network is input to the prediction network, so that the cardiovascular disease risk prediction is realized. The generated personalized cardiopulmonary function digital twin model provides a powerful tool for developing more scientific and efficient physical training and early intervention and precise prevention of diseases.

Description

Technical Field

[0001] The present invention belongs to the field of cardiopulmonary exercise testing, and designs an individualized digital twin model of cardiopulmonary function. By means of the data obtained by the user through cardiopulmonary function exercise testing, the physiological mechanism model, and the neural network algorithm, the construction of the digital twin model is realized, so that the model simulates the user's individual in terms of physiological functions, generates an individualized digital twin model of cardiopulmonary function for the user, and effectively predicts the individual cardiopulmonary function limit and the risk of sudden acute diseases at the same time. It can be used to carry out more scientific and efficient physical fitness training, early intervention and precise prevention of diseases. Background Art

[0002] Cardiopulmonary exercise testing is an objective, quantitative, and non-invasive examination method, including a power treadmill, a suitable mask, a gas metabolism analyzer, an electrocardiogram monitoring system, etc., which can monitor the key indicators of the subject at each stage from rest, warm-up, exercise, to recovery. However, current research, such as the research of the Zhang Xiaohong team published in the Chinese Journal of Gerontology, "Construction of a Risk Prediction Model for Coronary Artery Stenosis Using Cardiopulmonary Exercise Testing", explores the relationship between cardiopulmonary exercise testing and coronary artery stenosis and establishes a risk prediction model. T-tests and Logistic regression analysis are used to analyze the risk factors that may affect coronary artery stenosis, and a nomogram model is constructed and internally verified by the Bootstrap method for risk prediction; the research of the Duli team published in the Journal of Inner Mongolia Medical University, "Prediction of Postoperative Complications and Pulmonary Function in Patients Undergoing Hepatobiliary Surgery by Cardiopulmonary Exercise Testing", mainly focuses on patients undergoing hepatobiliary surgery, and the postoperative complications can be predicted through the indicators of cardiopulmonary exercise testing. Normal distribution, t-tests, Wilcoxon rank sum tests, and χ2 tests are used. Professional statistical methods are used for comparative analysis; however, the indicator selection is limited, mainly focusing on a few indicators such as VO2max / pred, VO2max / kg, AT, O2, and HR / pred, and the detailed data records provided by cardiopulmonary exercise testing equipment are not fully utilized. Most current research stays at data processing, and disease diagnosis still requires professional analysis by doctors. If machine learning neural network technology is used, through the individualized construction of a digital twin model of cardiopulmonary function, the automatic evaluation of the body state and the risk of disease by the machine can be realized, providing help and suggestions for sports health management.

[0003] To construct an individualized digital twin model of cardiopulmonary function, the present invention requires measurement data from the cardiopulmonary exercise test of the subject (such as metabolic equivalent, electrocardiogram, blood pressure, blood oxygen saturation, gas metabolism, respiratory rate, respiratory reserve, oxygen pulse, etc.), as well as corresponding health status diagnoses, including cardiopulmonary function limits (such as maximum heart rate, maximum oxygen uptake, forced vital capacity, etc.) and disease judgments (such as hypertension, heart valve disease, pulmonary hypertension, pulmonary artery insufficiency, obesity, etc.). Through some of the above data, more parameters (such as alveolar gas partial pressure, respiratory rhythm, neural drive ventilation volume, pulmonary peripheral circulation flow, etc.) are first obtained through a physiological mechanism model, and then combined with all the data and each index to guide machine learning to obtain a digital twin model of cardiopulmonary function. At the same time, it can probabilistically evaluate and risk grade potential diseases in an individualized manner, providing a scientific health management plan and exercise prescription advice for individuals. Summary of the Invention

[0004] The present invention proposes a method for realizing exercise health management based on an individualized digital twin model of cardiopulmonary function. Taking the results of the cardiopulmonary exercise test of the subject as a reference, a digital twin model is constructed through a neural network to simulate the physiological functions of the cardiopulmonary system during exercise, and an exercise health management plan for individual subjects is generated to achieve the purpose of intelligent individual diagnosis and health intervention.

[0005] The technical solution of the present invention:

[0006] A method for realizing exercise health management based on an individualized digital twin model of cardiopulmonary function is as follows:

[0007] The first step is to construct a digital twin model of the physiological mechanism of the human cardiopulmonary function

[0008] First, construct a physical model describing the physiological mechanism of the human cardiopulmonary function, and then use the first neural network algorithm to learn the complex physical model to generate a lightweight digital model;

[0009] (1) The physical model of the physiological mechanism of the human cardiopulmonary function includes the cardiovascular system, respiratory mechanics system, gas exchange system, cardiovascular control system, and respiratory control system.

[0010] Furthermore, the cardiovascular system is used to simulate the blood flow in the circulatory system and the dynamic changes of the cardiovascular system during exercise. It includes a heart model, a pulmonary circulation model, a systemic circulation model, a vascular bed model, and a venous return model. The heart model contains the left and right atria and ventricles, and considers the blood flow through the mitral valve, aortic valve, tricuspid valve, and pulmonary valve; the pulmonary circulation model includes the pulmonary artery, peripheral blood vessels, and pulmonary veins; the systemic circulation model includes the systemic artery, peripheral blood vessels, systemic veins, and vena cava; the vascular bed model is divided into the vascular beds of active and resting muscles, and considers the local vasodilation mechanism during exercise; the venous return model includes the muscle pump and the respiratory pump. The muscle pump refers to the effect of muscle contraction on venous return, and the respiratory pump refers to the effect of respiration on venous return. Preferably, the cardiovascular system includes a heart model, pulmonary circulation, systemic circulation, respiratory pump, and muscle pump.

[0011] Furthermore, the cardiovascular control system is used to simulate the regulation of the neuroregulatory mechanism on the cardiovascular system. It includes a local blood flow regulation model, a central nervous system ischemic response model, afferent and efferent nerve pathways, and a reflex control model of cardiovascular effectors. Among them, the local blood flow regulation model simulates the vasodilation response of the vascular beds of the brain, coronary arteries, and skeletal muscles; the central nervous system ischemic response model simulates the effect of hypoxia on heart rate and systemic circulation; the afferent and efferent nerve pathways simulate the carotid sinus baroreceptors, peripheral chemoreceptors, and pulmonary stretch reflex; the reflex control model of cardiovascular effectors simulates the regulation of the sympathetic and parasympathetic nerves on heart rhythm, ventricular contractility, peripheral vascular resistance, and venous capacitance. Cardiovascular control also includes the influence of respiratory central neuromuscular drive on the efferent pathway and evaluates the central command mechanism as a metabolic regulatory response.

[0012] Furthermore, the respiratory mechanics system is used to simulate the generation of airflow and the changes in lung volume during respiration, as well as the adjustment of respiratory frequency and tidal volume during exercise. It includes a model representing the mechanical properties of the lungs and a model representing the mechanical properties of the upper respiratory tract.

[0013] Furthermore, the respiratory control system uses information from central and peripheral chemoreceptors and metabolic drive components related to ventilation to estimate ventilation demand and constitutes a ventilation controller. At the same time, the respiratory controller also includes a respiratory pattern optimization method, which minimizes the cyclic respiratory work by adjusting the respiratory pattern in each respiratory cycle and constitutes a respiratory pattern optimizer.

[0014] Furthermore, the gas exchange system is used to simulate the transport and exchange of oxygen and carbon dioxide during exercise, as well as to maintain the stability of gas concentrations in the blood. It includes a gas mixing and exchange model, a gas transport model, and a metabolic kinetics model. Among them, the gas mixing and exchange model considers the mixing of alveolar gases, blood-gas exchange, and gas exchange at the tissue level; the gas transport model simulates the transport and metabolism of oxygen and carbon dioxide in the brain and tissues; the metabolic kinetics model regulates the respiratory demand through the metabolic rate ratio and metabolism-related neural drive. The cardiovascular system and respiratory mechanics interact through the gas exchange system to regulate the concentration of blood gases.

[0015] Furthermore, the specific information interaction relationship between the physical models of the physiological mechanism of the human cardiopulmonary function is as follows: For the cardiovascular control system: Metabolic regulation requires the external input of carbon dioxide exhalation volume and anaerobic threshold AT, and outputs the relative intensity I of aerobic exercise; The neuromuscular drive of the respiratory system requires the ventilation volume VE output by the ventilation controller, the induction time TI output by the respiratory pattern optimizer, and the blood flow rate BF, and outputs the central respiratory neuromuscular drive response Nt; The efferent pathway requires I output by metabolic regulation, Nt output by the neuromuscular drive of the respiratory system, the controlled variable θ output by the ischemic response of the central nervous system aj and multiple afferent activities f output by the afferent pathway aj , and outputs multiple efferent activities f sj and the spike frequency f of the vagus nerve efferent fibers v ; The effector of reflex control requires I output by metabolic regulation, f output by the efferent pathway aj and f v and the coronary artery region hp, skeletal resting muscle region rmp, active muscle region amp, and brain region bp fluid group pressure R output by local blood flow control jp , and outputs the smooth muscle outer region ep, spleen region sp, coronary artery region hp, skeletal resting muscle region rmp, active muscle region amp, and brain region bp fluid group pressure R output by local blood flow control jp,t , heart rate HR, maximum systolic instantaneous left and right ventricular elasticity E max,jv , left atrial blood flow Q la and skeletal peripheral resistance R amp,n ; The ischemic response of the central nervous system requires the partial pressure of carbon dioxide in the brain PaCO2 and the partial pressure of oxygen in the brain PaO2 output by gas exchange and mixing, and outputs the controlled variable θ sj ; The afferent pathway requires PaCO2, PaO2 output by gas exchange and mixing, tidal volume VT output by pulmonary mechanics, and P output by systemic circulation SA , and outputs f aj ; Local blood flow control requires the skeletal peripheral resistance R output by the effector of reflex control amp,n, I of metabolic regulation output, PaCO2 of gas exchange and mixing output, and inflow Q of systemic circulation output jp , output R jp .

[0016] For the cardiovascular system: The systemic circulation requires the relative intensity I of aerobic exercise of metabolic regulation output, the circulating blood volume V of the effector output of reflex control u,jv , the pressure R of each partition fluid group at time T of the effector output of reflex control jp,t , the extravascular pressure P of the active muscle veins of the muscle pump output im , the abdominal pressure P of the respiratory pump output abd , the intrathoracic pressure P of the respiratory pump output thor , the pulmonary artery blood volume V of the pulmonary circulation output pulmonary , the left ventricular blood flow Q of the heart output lv , the right atrial pressure P of the heart output ra and the heart blood volume V of the heart output heart , output the right atrial blood flow Q ra , the systemic arterial pressure P SA and inflow Q jp ; The heart requires the right atrial blood flow Q of the systemic circulation output ra , the heart rate HR of the effector output of reflex control, the left and right ventricular elasticity E at the maximum contraction moment of the effector output of reflex control max,jv , the left atrial blood flow Q of the pulmonary circulation output la and the P of the respiratory pump output thor , output the left ventricular blood flow Q lv , the right atrial pressure P ra , the heart blood volume V heart and the right ventricular blood flow Q rv ; The pulmonary circulation requires the right ventricular blood flow Q of the heart output rv and the intrathoracic pressure P of the respiratory pump output thor , output the pulmonary peripheral circulation blood flow Q pp , the pulmonary artery blood volume V pulmonary and the left atrial blood flow Q la ; The respiratory pump requires the induction time TI of the respiratory pattern optimizer output, the tidal volume VT of the pulmonary mechanics output, and the blood flow rate BF of the respiratory pattern optimizer output, and outputs the abdominal pressure P abd and the intrathoracic pressure P thor ; The muscle pump requires no input and outputs the extravascular pressure P of the active muscle veins im .

[0017] For the respiratory control system: The mean detector requires the partial pressure of oxygen in the brain PaO2 for gas exchange and mixed output, the partial pressure of carbon dioxide in the brain PaCO2 for gas exchange and mixed output, and the pressure of carbon dioxide in the cerebral venous blood PbCO2 for gas transport output, and outputs the average value of PaO2 PamO2, the average value of PaCO2 PamCO2, and the average value of PbCO2 PbmCO2; the ventilation controller requires the average value of PaO2 PamO2, the average value of PaCO2 PamCO2, the average value of PbCO2 PbmCO2 output by the mean detector, the ventilation neural drive component MRV related to metabolism output by the metabolic kinetics, and the blood flow rate BF output by the respiratory pattern optimizer, and outputs the ventilation volume VE; the respiratory pattern optimizer requires the ventilation volume VE output by the ventilation controller and the muscle pressure signal P output by the pulmonary mechanics musc , and outputs the induction time TI, the blood flow rate BF, and the neural drive response Nd.

[0018] For the gas exchange system: Gas exchange and mixing requires the blood volume V output by the pulmonary mechanics, the inflow Q output by the systemic circulation jp , the blood flow in the left atrium Q output by the pulmonary circulation la , the inhaled oxygen concentration FiO2 input from the outside, the inhaled carbon dioxide concentration FiCO2 input from the outside, the atmospheric pressure P input from the outside atm , the blood flow in the peripheral pulmonary circulation Q output by the pulmonary circulation pp , the tidal volume VT output by the pulmonary mechanics, the mixed venous carbon dioxide concentration CvCO2 output by the gas transport, and the mixed venous oxygen concentration CvO2 output by the gas transport, and outputs the arterial carbon dioxide concentration CaCO2 and the arterial oxygen concentration CaO2; gas transport requires the partial pressure of carbon dioxide in the brain PaCO2 output by gas exchange and mixing, the arterial carbon dioxide concentration CaCO2 output by gas exchange and mixing, the arterial oxygen concentration CaO2 output by gas exchange and mixing, the inflow Q output by the systemic circulation jp , the carbon dioxide uptake VCO2 input from the outside, the tidal volume VT output by the pulmonary mechanics, and the oxygen uptake VO2 input from the outside, and outputs the mixed venous carbon dioxide concentration CvCO2, the mixed venous oxygen concentration CvO2, the metabolic rate of carbon dioxide production MRTCO2, the metabolic rate of oxygen production MRTO2, and the pressure of carbon dioxide in the cerebral venous blood PbCO2; metabolic kinetics requires the metabolic rate of oxygen production MRTO2 output by gas transport and the metabolic rate of carbon dioxide production MRTCO2 output by gas transport, and outputs the ventilation neural drive component MRV related to metabolism.

[0019] For the respiratory dynamics system: The upper respiratory tract requires the blood volume V output by the pulmonary mechanics, the pleural pressure P output by the pulmonary mechanics pl , and the tidal volume VT output by the pulmonary mechanics, and outputs the airway flow gain G AW; Pulmonary mechanics requires the airway flow gain G of the upper respiratory tract output AW , the tracheal pressure P input from the outside ao and the neural drive response Nd output by the respiratory pattern optimizer, and outputs the blood volume V, pleural pressure P pl , muscle pressure signal P musc and tidal volume VT.

[0020] The physical model of the physiological mechanism of the human cardiopulmonary function is modeled and simulated through the Simulink library in MATLAB. First, five large module frameworks are constructed by adding atomic subsystem modules; considering the complex characteristics of the formulas in the model, S functions are used to write the formulas and parameters in each sub-module respectively, set the inputs and outputs according to the model requirements, and record them in the corresponding S-Function modules; connect the input and output interfaces of each S function block according to the parameter exchange situation in the model, and the construction of the entire human cardiopulmonary function physiological model can be realized.

[0021] (2) The first neural network algorithm adopts a sequence prediction model based on the long short-term memory network LSTM;

[0022] The input data includes: human body parameters: muscle contraction duration, muscle contraction-relaxation cycle duration; directly measured quantities in cardiopulmonary exercise tests: VCO2, VO2, FiCO2, FiO2, AT; environmental input quantities: P atm , Pao; external environment: temperature T (degrees Celsius), humidity RH (percentage), noise level L P (decibels); exercise load: Load;

[0023] The output data includes: forced vital capacity, right atrial flow, left ventricular flow, left atrial flow, right ventricular flow, intrathoracic pressure, pulmonary peripheral circulation flow, minimized cost function, neural drive ventilation, alveolar gas partial pressure, metabolic equivalent, respiratory exchange ratio, load intensity, heart rate, reserve heart rate, oxygen pulse, stroke volume, systolic blood pressure, pulmonary artery diastolic pressure, blood oxygen saturation, oxygen uptake, oxygen uptake per kg of muscle, carbon dioxide output, minute ventilation volume, respiratory frequency, respiratory reserve, oxygen ventilation equivalent, carbon dioxide ventilation equivalent, end-tidal oxygen partial pressure, end-tidal carbon dioxide partial pressure; further, since the physiological indicators change dynamically with the exercise intensity, the input data is presented in the form of a time series, and the specific structure is a three-dimensional tensor, and its dimensions are defined as the number of samples (n), time step (Δt, that is, the observed values at different time points), and the number of features (y i ), that is, (n, Δt, y i ). In order to adapt to the processing requirements of the neural network, the input data needs to be processed by the Z-score standardization method:

[0024]

[0025] Among them, y i is the original data point, μ is the mean of the data, and σ is the standard deviation of the data. The preprocessed data set is divided into a training set and a validation set proportionally.

[0026] The implementation of the first neural network algorithm is divided into two steps: the pre-training stage, learning the general laws of the cardiopulmonary exercise process from a large range of population data to establish a basic framework; the secondary training stage, fine-tuning based on individual real data, optimizing the network parameters to precisely adapt to the characteristics of individuals, and obtaining a personalized digital twin model;

[0027] Furthermore, the specific network training stage is as follows:

[0028] 1) Pre-training:

[0029] The input data comes from abnormal data with randomly modified indicators of normal people, as well as the indicator results measured by the subjects through cardiopulmonary exercise testing (CPET);

[0030] The output data comes from the indicator results measured by some subjects through cardiopulmonary exercise testing (CPET), as well as the internal function parameters of the human body obtained during the calculation process of the physiological mechanism model for the indicator results measured by other subjects through cardiopulmonary exercise testing (CPET).

[0031] First, initialize the network weights, perform forward propagation with the corresponding parameters in the input data set. The data passes through three LSTM hidden layers in sequence. In each hidden layer, it is jointly processed by weighted summation and a non-linear activation function; finally, a predicted value is generated in the output layer.

[0032] Regarding the difference between the network predicted value and the value generated by the simulation data set. The loss function is calculated by the mean square error and is described as:

[0033]

[0034] Among them, L represents the loss function; y i can represent the internal function parameters of the human body calculated in the previous step and the true value of the CPET test data; represents the corresponding predicted value of the neural network (1st) pre-training; N refers to the number of output data of the neural network (1st);

[0035] In the backpropagation process, first, it is necessary to calculate the gradient of the loss function with respect to the network output, and then calculate the gradients of each weight and bias layer by layer through the chain rule. These gradients represent the contribution degree of each parameter to the final error. Then, through the stochastic gradient descent algorithm, the gradients are used to update the network weights to reduce the loss, so that the network continuously adjusts its weight parameters to minimize the loss function. To prevent overfitting, a regularization technique (Dropout) is adopted after each hidden layer: during training, some neuron outputs in the network are randomly set to zero to avoid the model being too complex or overfitting to the training data.

[0036] By repeatedly iterating the above forward propagation and backpropagation until the loss function converges stably.

[0037] 2) Secondary training:

[0038] For the dataset of secondary training, the input and output only need the real index results measured by the subject through cardiopulmonary exercise testing (CPET), and do not require the internal function parameters of the human body calculated by the physiological mechanism model in the first step.

[0039] The process of secondary training is as follows:

[0040] First, use the pre-trained network parameters as the initial weights. The input is the same as in the pre-training stage, and the same network structure is passed through to generate predicted values at the output layer.

[0041] Then, use the loss function to calculate the CPET test index parameters according to the mean square error in formula (2) for the difference between the network predicted value and the real measured value. The backpropagation process is the same as in the pre-training stage. By minimizing the loss function, the network output can more accurately fit the individual's test data and optimize the network performance.

[0042] By repeatedly iterating the above forward propagation and backpropagation until the loss function converges stably.

[0043] In the second step, endow the digital twin model with the function of sports health management

[0044] The second neural network still uses the LSTM method. Input data: including the output data used by the first neural network in the second training; Output data: the CPET diagnosis results of the subject. The preprocessed data is divided into a training set and a validation set according to a ratio. For the loss function, for disease prediction and health advice at the same time, it is necessary to establish loss functions separately and assign weights to combine them into a total loss function.

[0045] Among them, the LSTM layer implements the gating mechanism through the sigmoid function to control the flow and forgetting of information; at the same time, the tanh function is used for non-linear transformation of the candidate memory cells and output values to capture complex dependencies in the time series.

[0046] Further, since the physiological indicators change dynamically with the exercise intensity, the input data is presented in the form of a time series, and its specific structure is a three-dimensional tensor, whose dimensions are defined as the number of samples (n), the time step (Δt, i.e., the observed values at different time points), and the number of features (x i ), i.e., (n, Δt, x i ). To meet the processing requirements of the neural network, the input data needs to be processed by the Z-score normalization method:

[0047]

[0048] where x i is the original data point, μ is the mean of the data, and σ is the standard deviation of the data.

[0049] After being trained by the neural network (2nd), the output needs to pass through the softmax function:

[0050]

[0051] where Soft(z) i is the result of the i-th element in the vector z after being processed by the Softmax function, so as to output the disease corresponding probability distribution and exercise training suggestions in the specified disease order.

[0052] The effects and benefits of the present invention are as follows: By means of the relevant indicators and data provided by the subject through the cardiopulmonary exercise test, the physiological mechanism simulation of the cardiopulmonary function is realized, the construction of an individualized digital twin model is achieved, the cardiopulmonary function limit and the risk of sudden acute diseases of an individual are predicted, and the purpose of giving exercise health management suggestions is achieved. The digital twin model can be used as a virtual digital representation of the subject's own physiological mechanism, and machine evaluation can replace doctor diagnosis. It can be used in the fields of more scientific and efficient physical fitness training, early intervention and precise prevention of diseases, etc. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 is the technical principle flow chart of the present invention.

[0054] Figure 2 is a schematic diagram of the physiological model of the human cardiopulmonary function.

[0055] Figure 3 is the internal principle diagram of the neural network; (a) is the first neural network; (b) is the second neural network. DETAILED DESCRIPTION OF THE INVENTION

[0056] In order to make the purpose and technical solution of the present invention clearer, the present invention will be further described in detail below in combination with specific parameter examples.

[0057] A method for realizing exercise health management based on an individualized digital twin model of cardiopulmonary function, the steps are as follows:

[0058] The first step is to construct a digital twin model of the physiological mechanism of the human cardiopulmonary function

[0059] The digital twin model of the physiological mechanism of the human cardiopulmonary function is realized by integrating physical modeling and deep learning algorithms, and can be personalized for the cardiopulmonary function of the subject to update important index data.

[0060] The specific implementation method can be divided into two processes. First, a physical model describing the physiological mechanism of the human cardiopulmonary function is constructed, and then the first neural network algorithm is used to learn the complex physical model to generate a lightweight digital model;

[0061] (1) The physical model of the physiological mechanism of the human cardiopulmonary function is constructed based on an integrated mathematical model, including five modules: the cardiovascular system, the respiratory mechanics system, the gas exchange system, the cardiovascular controller, and the respiratory controller. Each module contains different sub-modules. This model is based on previously validated cardiovascular and respiratory models and incorporates mechanisms related to the dynamics of aerobic exercise. It can predict the parameters related to the internal cardiopulmonary function of healthy people during rest and aerobic exercise.

[0062] The five modules in the physical model of the human cardiopulmonary function are described as follows:

[0063] Furthermore, the cardiovascular system is used to simulate the flow of blood in the circulatory system and the dynamic changes of the cardiovascular system during exercise. It includes a heart model, a pulmonary circulation model, a systemic circulation model, a vascular bed model, and a venous return model. The heart model contains the left and right atria and ventricles, considering the blood flow through the mitral valve, aortic valve, tricuspid valve, and pulmonary valve; the pulmonary circulation model includes the pulmonary artery, peripheral blood vessels, and pulmonary veins; the systemic circulation model includes the systemic artery, peripheral blood vessels, systemic veins, and vena cava; the vascular bed model is divided into the vascular beds of active and resting muscles, considering the local vasodilation mechanism during exercise; the venous return model includes the muscle pump and the respiratory pump. The muscle pump refers to the effect of muscle contraction on venous return, and the respiratory pump refers to the effect of respiration on venous return. Preferably, the cardiovascular system includes a heart model, pulmonary circulation, systemic circulation, respiratory pump, and muscle pump.

[0064] Furthermore, the Cardiovascular Controller is used to simulate the regulation of the cardiovascular system by the neuromodulation mechanism, especially the regulation of heart rate and blood pressure during exercise. It includes a local blood flow regulation model, a central nervous system ischemic response model, afferent and efferent neural pathways, and a reflex control model of cardiovascular effectors. Among them, the local blood flow regulation model simulates the vasodilation response of the blood vessels in the brain, coronary arteries, and skeletal muscle vascular beds; the central nervous system ischemic response model simulates the effects of hypoxia on heart rate and systemic circulation; the afferent and efferent neural pathways simulate the carotid sinus baroreceptors, peripheral chemoreceptors, and pulmonary stretch reflex; the reflex control model of cardiovascular effectors simulates the regulation of heart rhythm, ventricular contractility, peripheral vascular resistance, and venous capacitance by the sympathetic and parasympathetic nerves. Cardiovascular control also includes the influence of respiratory central neuromuscular drive on the efferent pathway and evaluates the central command mechanism as a metabolic regulatory response.

[0065] Furthermore, the Respiratory Mechanics is used to simulate the generation of air flow and the change of lung volume during respiration, as well as the adjustment of respiratory frequency and tidal volume during exercise. It includes a model representing the mechanical properties of the lungs and a model representing the mechanical properties of the upper respiratory tract.

[0066] Furthermore, the Respiratory Controller uses the information of central chemoreceptors, peripheral chemoreceptors, and metabolic drive components related to ventilation to estimate the ventilation demand and constitutes a ventilation controller. At the same time, the respiratory controller also includes a respiratory pattern optimization method, which minimizes the cyclic respiratory work by adjusting the respiratory pattern in each respiratory cycle and constitutes a respiratory pattern optimizer.

[0067] Furthermore, the Gas Exchange System is used to simulate the transport and exchange of oxygen and carbon dioxide during exercise, as well as to maintain the stability of gas concentration in the blood. It includes a gas mixing and exchange model, a gas transport model, and a metabolic kinetics model. Among them, the gas mixing and exchange model considers the mixing of alveolar gases, blood-gas exchange, and gas exchange at the tissue level; the gas transport model simulates the transport and metabolism of oxygen and carbon dioxide in the brain and tissues; the metabolic kinetics model regulates the respiratory demand through the metabolic rate ratio and metabolism-related neural drive. The cardiovascular system and respiratory mechanics interact through the gas exchange system to regulate the concentration of blood gases.

[0068] Furthermore, the information interaction relationship between the physical models of the physiological mechanism of the human cardiopulmonary function is specifically as follows: For the cardiovascular control system: Metabolic regulation requires an external input of carbon dioxide exhalation volume And the anaerobic threshold AT, output the relative intensity I of aerobic exercise; the neuromuscular drive of the respiratory system requires the ventilation volume VE output by the ventilation controller, the induction time TI output by the respiratory pattern optimizer, and the blood flow rate BF, and output the central respiratory neuromuscular drive response Nt; the efferent pathway requires I output by metabolic regulation, Nt output by the neuromuscular drive of the respiratory system, and the controlled variable θ output by the central nervous system ischemic response aj And a variety of afferent activities f output by the afferent pathway aj , output a variety of efferent activities f sj And the spike frequency f of the vagus nerve efferent fibers v ; The effector of reflex control requires I output by metabolic regulation, f output by the efferent pathway aj And f v And the coronary artery region hp, the skeletal resting muscle region rmp, the active muscle region amp, and the brain region bp fluid group pressure R output by local blood flow control jp , output the smooth muscle outer region ep, the spleen region sp, the coronary artery region hp, the skeletal resting muscle region rmp, the active muscle region amp, and the brain region bp fluid group pressure R output by local blood flow control jp,t , heart rate HR, the left and right ventricular elasticity E at the instant of maximum contraction max,jv , the left atrial blood flow Q la And the skeletal peripheral resistance R amp,n ; The central nervous system ischemic response requires the cerebral carbon dioxide partial pressure PaCO2 and the cerebral oxygen partial pressure PaO2 output by gas exchange and mixing, and output the controlled variable θ sj ; The afferent pathway requires PaCO2, PaO2 output by gas exchange and mixing, the tidal volume VT output by pulmonary mechanics, and P output by the systemic circulation SA , and output f aj ; Local blood flow control requires the skeletal peripheral resistance R output by the effector of reflex control amp,n , I output by metabolic regulation, PaCO2 output by gas exchange and mixing, and the inflow volume Q output by the systemic circulation jp , and output R jp .

[0069] For the cardiovascular system: The systemic circulation requires the relative intensity I of aerobic exercise output by metabolic regulation, the circulating blood volume V output by the effector of reflex control u,jv , the fluid group pressure R of each partition at time T output by the effector of reflex control jp,t , the extravascular pressure P of the active muscle veins output by the muscle pump im , the abdominal pressure P output by the respiratory pump abd , the intrathoracic pressure P output by the respiratory pump thor , the pulmonary artery blood volume V output by the pulmonary circulation pulmonary , the left ventricular blood flow Q output by the heart lv, the right atrial pressure P of cardiac output ra and the cardiac blood volume V of cardiac output heart , outputting the right atrial blood flow Q ra , the systemic arterial pressure P SA and the inflow Q jp ; the heart requires the right atrial blood flow Q of systemic circulation output ra , the heart rate HR output by the effector of reflex control, the left and right ventricular elasticity E at the instant of maximum contraction output by the effector of reflex control max,jv , the left atrial blood flow Q of pulmonary circulation output la and the P output by the respiratory pump thor , outputting the left ventricular blood flow Q lv , the right atrial pressure P ra , the cardiac blood volume V heart and the right ventricular blood flow Q rv ; the pulmonary circulation requires the right ventricular blood flow Q of cardiac output rv and the intrathoracic pressure P output by the respiratory pump thor , outputting the pulmonary peripheral circulation blood flow Q pp , the pulmonary artery blood volume V pulmonary and the left atrial blood flow Q la ; the respiratory pump requires the induction time TI output by the respiratory pattern optimizer, the tidal volume VT output by the pulmonary mechanics, and the blood flow rate BF output by the respiratory pattern optimizer, and outputs the abdominal pressure P abd and the intrathoracic pressure P thor ; the muscle pump requires no input and outputs the extravascular pressure P of the active muscle vein im .

[0070] For the respiratory control system: The mean detector requires the partial pressure of oxygen in the brain PaO2 output by gas exchange and mixing, the partial pressure of carbon dioxide in the brain PaCO2 output by gas exchange and mixing, and the pressure of carbon dioxide in the cerebral venous blood PbCO2 output by gas transport, and outputs the average value of PaO2 PamO2, the average value of PaCO2 PamCO2, and the average value of PbCO2 PbmCO2; The ventilation controller requires the average value of PaO2 PamO2 output by the mean detector, the average value of PaCO2 PamCO2, the average value of PbCO2 PbmCO2, the ventilation neural drive component MRV related to metabolism output by metabolic kinetics, and the blood flow rate BF output by the respiratory pattern optimizer, and outputs the ventilation volume VE; The respiratory pattern optimizer requires the ventilation volume VE output by the ventilation controller and the muscle pressure signal P output by the pulmonary mechanics musc , and outputs the induction time TI, the blood flow rate BF, and the neural drive response Nd.

[0071] For the gas exchange system: Gas exchange and mixing requires the blood volume V output by pulmonary mechanics and the inflow Q of systemic circulation output jp, the left atrial blood flow Q output by the pulmonary circulation la , the inhaled oxygen concentration FiO2 input from the outside world, the inhaled carbon dioxide concentration FiCO2 input from the outside world, and the atmospheric pressure P input from the outside world atm , the pulmonary peripheral circulation blood flow Q output by the pulmonary circulation pp , the tidal volume VT output by pulmonary mechanics, the mixed venous carbon dioxide concentration CvCO2 output by gas transport, and the mixed venous oxygen concentration CvO2 output by gas transport, outputting the arterial carbon dioxide concentration CaCO2 and the arterial oxygen concentration CaO2; gas transport requires the partial pressure of carbon dioxide in the brain PaCO2, the arterial carbon dioxide concentration CaCO2, the arterial oxygen concentration CaO2 output by gas exchange and mixing, and the inflow Q output by the systemic circulation jp , the carbon dioxide uptake VCO2 input from the outside world, the tidal volume VT output by pulmonary mechanics, and the oxygen uptake VO2 input from the outside world, outputting the mixed venous carbon dioxide concentration CvCO2, the mixed venous oxygen concentration CvO2, the metabolic production rate of carbon dioxide MRTCO2, the metabolic production rate of oxygen MRTO2, and the pressure of carbon dioxide in the cerebral venous blood PbCO2; metabolic kinetics requires the metabolic production rate of oxygen MRTO2 output by gas transport and the metabolic production rate of carbon dioxide MRTCO2 output by gas transport, outputting the ventilation neural drive component MRV related to metabolism.

[0072] For the respiratory dynamics system: the upper respiratory tract requires the blood volume V output by pulmonary mechanics, the pleural pressure P output by pulmonary mechanics pl and the tidal volume VT output by pulmonary mechanics, outputting the airway flow gain G AW ; pulmonary mechanics requires the airway flow gain G output by the upper respiratory tract AW , the tracheal pressure P input from the outside world ao and the neural drive response Nd output by the respiratory pattern optimizer, outputting the blood volume V, the pleural pressure P pl , the muscle pressure signal P musc and the tidal volume VT.

[0073] The physical model of the physiological mechanism of the human cardiopulmonary function is modeled and simulated through the Simulink library in MATLAB. First, five large module frameworks are constructed by adding atomic subsystem modules; considering the complex characteristics of the formulas in the model, S functions are used to write the formulas and parameters in each sub-module respectively, set the inputs and outputs according to the model requirements, and record them in the corresponding S-Function modules; connect the input and output interfaces of each S function block according to the exchange of parameters in the model, and the construction of the entire human cardiopulmonary function physiological model can be realized.

[0074] Input the required parameters in Simulink for physiological model simulation to calculate a series of internal human function parameters; the results of each index measured by the subject through a cardiopulmonary exercise test (CPET) together provide a data basis for the neural network (1st).

[0075] The input parameters required to run the simulation are: directly measured quantities in the cardiopulmonary exercise test: VCO2 (ml), VO2 (ml), FiCO2 (%), FiO2 (%), AT (1 / min); environmental input quantities: P atm (mmHg), Pao (cmH2O);

[0076] Table 1 List of pre-set parameters in the physiological model

[0077]

[0078] Table 2 List of parameters calculated in the physiological model

[0079]

[0080] Table 3 List of parameters measured by the cardiopulmonary exercise test

[0081]

[0082]

[0083] (2) The first neural network algorithm obtains a digital twin model by machine learning the physiological mechanism model of cardiopulmonary function, making the model change from complex to lightweight to improve the calculation efficiency.

[0084] This neural network structure is fixed and is a sequence prediction model based on the long short-term memory network LSTM;

[0085] The input data includes:

[0086] Human parameters: muscle contraction duration, muscle contraction-relaxation cycle duration;

[0087] Directly measured quantities in the cardiopulmonary exercise test: VCO2, VO2, FiCO2, FiO2, AT;

[0088] Environmental input quantities: P atm , Pao;

[0089] External environment: T, RH, L P ;

[0090] Exercise load: Load;

[0091] Output data include: forced vital capacity, right atrial flow, left ventricular flow, left atrial flow, right ventricular flow, intrathoracic pressure, pulmonary peripheral circulation flow, minimization cost function, neural driven ventilation, alveolar gas partial pressure, metabolic equivalent, respiratory exchange rhythm, load intensity, heart rate, reserve heart rate, oxygen pulse, stroke volume, systolic pressure, pulmonary artery diastolic pressure, blood oxygen saturation, oxygen uptake, oxygen uptake per kg muscle, carbon dioxide excretion, minute ventilation, respiratory rate, respiratory reserve, oxygen ventilation equivalent, carbon dioxide ventilation equivalent, end-tidal oxygen partial pressure, end-tidal carbon dioxide partial pressure;

[0092] Since physiological indicators change dynamically with exercise intensity, the input data is expressed in the form of time series. The specific structure is a three-dimensional tensor, whose dimensions are defined as the number of samples (n), the time step (Δt, i.e., the observation value at different time points), the number of features (y i ), that is, (n,Δt,y i ). In order to meet the processing requirements of neural networks, the input data needs to be processed by Z-score standardization method:

[0093]

[0094] Among them, y i is the original data point, μ is the mean of the data, and σ is the standard deviation of the data. After standardization, the data will have a distribution with a mean of 0 and a standard deviation of 1, which helps to reduce the impact of outliers and speed up the convergence speed during model training. The preprocessed dataset is divided into a training set and a validation set in a ratio of 7:3.

[0095] The first neural network algorithm is implemented in two steps: the pre-training stage, learning the general laws of the cardiopulmonary exercise process from a wide range of population data and establishing a basic framework; the secondary training stage, fine-tuning based on individual real data, optimizing network parameters to make it accurately adapt to the characteristics of individuals, and obtaining a personalized digital twin model;

[0096] Furthermore, the specific network training stages are as follows:

[0097] 1) Pre-training: Before the model formally performs personalized cardiopulmonary function simulation tasks, it needs to be trained and learned on a large-scale dataset to grasp the general laws of cardiopulmonary function in a wide range of people.

[0098] The pre-trained dataset needs to use the physical model of the physiological mechanism of human cardiopulmonary function in step (1) to generate a diverse simulated dataset that can represent most of the population by randomly adjusting the constant parameters preset in the model within a reasonable range.

[0099] The pre-training process is as follows:

[0100] The input data is sourced from randomly modified abnormal data of normal person's index parameters and the results of various indicators measured by the subjects through cardiopulmonary exercise testing (CPET);

[0101] The output data is sourced from the results of various indicators measured by some subjects through cardiopulmonary exercise testing (CPET), and the internal function parameters of the human body obtained during the calculation process of the physiological mechanism model in step (1) for the results of various indicators measured by other subjects through cardiopulmonary exercise testing (CPET).

[0102] First, initialize the network weights, perform forward propagation on the corresponding parameters in the input dataset. The data passes through three LSTM hidden layers in sequence. In each hidden layer, it is jointly processed by weighted summation and non-linear activation functions; finally, a predicted value is generated in the output layer.

[0103] Then, use a loss function to measure the difference between the CPET test index parameters of the subjects and the internal function parameters of the human body calculated by the physical model of the physiological mechanism of the human cardiopulmonary function in step (1) for the difference between the network predicted value and the generated value of the simulated dataset. The loss function is calculated by the mean squared error and is described as:

[0104]

[0105] where, L represents the loss function; y i can represent the internal function parameters of the human body calculated in the previous step and the true value of the CPET test data; represents the corresponding predicted value pre-trained by the neural network (1st); N refers to the number of output data of the neural network (1st);

[0106] The smaller the loss value, the closer the prediction of the model is to the real data. By calculating the loss function, the network can understand the accuracy of the current prediction and provide a basis for the backpropagation process.

[0107] In the backpropagation process, first, it is necessary to calculate the gradient of the loss function with respect to the network output, and calculate the gradient of each weight and bias layer by layer through the chain rule. These gradients represent the contribution degree of each parameter to the final error; then, through the stochastic gradient descent algorithm, the gradients are used to update the network weights to reduce the loss, so that the network continuously adjusts its weight parameters to minimize the loss function; to prevent overfitting, a regularization technique (Dropout) is adopted after each hidden layer: during the training process, randomly set the outputs of some neurons in the network to zero to avoid the model being too complex or overfitting to the training data.

[0108] By repeatedly iterating the above forward propagation and backward propagation until the loss function converges stably. This training result can basically establish the weight parameters in the neural network, directly reflecting the learning result of the model from the large-scale dataset, thus constructing a basic model that can describe the general laws of cardiopulmonary function.

[0109] 2) Secondary training: In the pre-training stage, the network can initially master the general laws of cardiopulmonary function of most people. However, due to the differences in individual cardiopulmonary function, the current model cannot accurately simulate the physiological data of a certain real individual. To achieve personalized fitting, the network must be retrained to construct a digital physiological model for the subject individual.

[0110] For the dataset of secondary training, the input and output only need the real index results measured by the subject through the cardiopulmonary exercise test (CPET), and do not require the internal function parameters of the human body calculated by the physiological mechanism model in the first step. Although the output data in the training set is incomplete, this does not affect the goal of the model to achieve individualized fitting. The CPET test data can directly quantify the physical condition of the subject and reflect its individual physiological characteristics to a certain extent. Therefore, even without complete functional data, only based on the CPET data for model fitting can still significantly improve the personalized accuracy of the model.

[0111] The process of secondary training is as follows:

[0112] First, use the network parameters of pre-training as the initial weights. The input is the same as that in the pre-training stage, and the prediction value is generated in the output layer through the same network structure. In this process, the pre-training stage provides the network with preliminary capabilities, enabling it to have certain performance; introducing individualized data provides a training basis for specific individuals in secondary training, thus further improving the adaptability and accuracy of the model.

[0113] Then, use the loss function to calculate the CPET test index parameters according to formula (2) for the difference between the network prediction value and the real measurement value. The backward propagation process is the same as that in the pre-training stage. By minimizing the loss function, the network output can more accurately fit the test data of the individual and optimize the network performance.

[0114] By repeatedly iterating the above forward propagation and backward propagation until the loss function converges stably. This training result can fully fit the individualized data, enabling the weight parameters in the neural network to reach the optimal adjustment for each individual and realizing the individualized model adaptation ability.

[0115] At this time, the individualized neural network (1st) after pre-training and secondary training can efficiently calculate the predicted values of the internal function parameters of the human body, thus completely replacing the cardiopulmonary function physiological mechanism model. The complexity of the model is significantly reduced and the computational amount is greatly reduced. This output result and the known CPET measurement data of the subject are jointly transmitted to the next neural network, realizing the efficient and accurate determination of individualized cardiopulmonary function-related parameters.

[0116] The second step is to endow the digital twin model with the function of sports health management

[0117] To further realize individualized disease risk assessment and exercise recommendations based on digital twin technology, a second neural network needs to be constructed. This network will use the output results of the first individualized neural network and the cardiopulmonary exercise test (CPET) data of the subject to predict the potential disease risk of the individual and provide personalized exercise recommendations. The basic principle of training the second neural network is basically the same as that of the first one. The following is the training method of the second neural network:

[0118] The second neural network still uses the LSTM method. Input data: including the output data used by the first neural network in the second training; Output data: the CPET diagnosis results of the subject. For the division of the data set, the preprocessed data is divided into a training set and a validation set according to a ratio of 7:3. For the loss function, since disease prediction and health advice are carried out simultaneously, loss functions need to be established separately and weighted to form a total loss function.

[0119] Among them, the LSTM layer realizes the gating mechanism through the sigmoid function to control the flow and forgetting of information; at the same time, the tanh and function are used to perform non-linear transformation on the candidate memory cells and output values to capture the complex dependencies in the time series.

[0120] Since the physiological indicators change dynamically with the exercise intensity, the input data is presented in the form of a time series. The specific structure is a three-dimensional tensor, and its dimensions are defined as the number of samples (n), the time step (Δt, that is, the observed values at different time points), and the number of features (x i ), that is, (n, Δt, x i ). In order to adapt to the processing requirements of the neural network, the input data needs to be processed by the Z-score standardization method:

[0121]

[0122] Among them, x iis the original data point, μ is the mean of the data, and σ is the standard deviation of the data. After standardization, the data will have a distribution with a mean of 0 and a standard deviation of 1, which helps to reduce the impact of outliers and speeds up the convergence rate during model training; in addition, naming the corresponding predicted diseases by serial numbers helps to lighten the work of the neural network.

[0123] After being trained by the neural network (2nd), the output needs to pass through the softmax function:

[0124]

[0125] where Soft(z) i is the result of the i-th element in the vector z after being processed by the Softmax function, so as to output the corresponding probability distribution of the disease and the sports training suggestions in the specified disease order.

[0126] The neural network (2nd) generates predicted values for the individual's cardiopulmonary function limit and the risk of sudden acute diseases by receiving the individual cardiopulmonary function-related parameters output by the neural network (1st). It provides scientific and efficient physical training suggestions for users in a personalized manner and conducts disease risk assessments, so as to enable the digital twin model to have the function of sports health management.

[0127] Table 4 List of Disease Risk Predictions

[0128]

[0129]

[0130]

[0131] Table 5 List of Sports Training Suggestions

[0132]

[0133]

Claims

1. A method for sports health management based on an individualized cardiopulmonary function digital twin model, characterized in that: Here are the steps: The first step is to build a digital twin model of the physiological mechanism of human cardiopulmonary function First, a physical model describing the physiological mechanism of human cardiopulmonary function is constructed. Then, the first neural network algorithm is used to learn the complex physical model to generate a lightweight digital model. (1) The physical model of the physiological mechanism of human cardiopulmonary function includes the cardiovascular system, respiratory mechanics system, gas exchange system, cardiovascular control system, and respiratory control system; The physical model of the physiological mechanism of human cardiopulmonary function is modeled and simulated through the Simulink library in MATLAB. First, five large module frameworks are constructed by adding atomic subsystem modules. Considering the complex characteristics of the formulas in the model, the S-function is used to write the formulas and parameters in each submodule respectively, and the input and output are set according to the model requirements and recorded in the corresponding S-Function module. The input and output interfaces of each S-function block are interconnected according to the exchange of parameters in the model, so that the entire human cardiopulmonary function physiological model can be built. (2) The first neural network algorithm uses a sequence prediction model based on the long short-term memory network LSTM; Input data include: human parameters: muscle contraction duration, muscle contraction-relaxation cycle duration; cardiopulmonary exercise test directly measured quantities: VCO2, VO2, FiCO2, FiO2, AT; environmental input: P atm , Pao; External environment: temperature T, humidity RH, noise level L P ; Exercise load: Load; The output data include: forced vital capacity, right atrial flow, left ventricular flow, left atrial flow, right ventricular flow, intrathoracic pressure, pulmonary peripheral circulation flow, minimization cost function, neural driven ventilation, alveolar gas partial pressure, metabolic equivalent, respiratory exchange law, load intensity, heart rate, reserve heart rate, oxygen pulse, stroke volume, systolic pressure, pulmonary artery diastolic pressure, blood oxygen saturation, oxygen uptake, oxygen uptake per kg muscle, carbon dioxide excretion, minute ventilation, respiratory rate, respiratory reserve, oxygen ventilation equivalent, carbon dioxide ventilation equivalent, end-tidal oxygen partial pressure, end-tidal carbon dioxide partial pressure; further, since the physiological indicators change dynamically with exercise intensity, the input data is expressed in the form of time series, and the specific structure is a three-dimensional tensor, whose dimensions are defined as the number of samples n, the time step Δt, that is, the observation value at different time points, and the number of features y i ; In order to meet the processing requirements of neural networks, the input data needs to be processed by Z-score standardization method: Among them, y i is the original data point, μ is the mean of the data, and σ is the standard deviation of the data; the preprocessed data set is divided into a training set and a validation set in proportion; The first neural network algorithm is implemented in two steps: the pre-training stage, learning the general laws of the cardiopulmonary exercise process from a wide range of population data and establishing a basic framework; the secondary training stage, fine-tuning based on individual real data, optimizing network parameters to make it accurately adapt to the characteristics of individuals, and obtaining a personalized digital twin model; The second step is to give the digital twin model sports health management functions The second neural network still uses the LSTM method; input data: including the output data used by the first neural network in the second training; output data: the CPET diagnosis results of the subjects; the preprocessed data is divided into a training set and a validation set in proportion; for the loss function, disease prediction and health advice are performed simultaneously, and loss functions need to be established separately and weighted to form a total loss function; The LSTM layer uses the sigmoid function to implement a gating mechanism to control the flow and forgetting of information. At the same time, the tanh and function are used to perform nonlinear transformations on candidate memory cells and output values ​​to capture complex dependencies in the time series. After training through the second neural network, the output needs to pass through the softmax function: Among them, Soft(z) i It is the result of processing the i-th element in the vector z by the Softmax function, thereby outputting the probability distribution corresponding to the disease and exercise training suggestions according to the prescribed disease order.

2. According to claim 1, a method for implementing sports health management based on an individualized cardiopulmonary function digital twin model is characterized in that: The cardiovascular system is used to simulate the flow of blood in the circulatory system and the dynamic changes of the cardiovascular system during exercise; it includes a heart model, a pulmonary circulation model, a systemic circulation model, a vascular bed model, and a venous return model; the heart model includes the left and right atria and ventricles, and takes into account the blood flow of the mitral valve, aortic valve, tricuspid valve, and pulmonary valve; the pulmonary circulation model includes the pulmonary artery, peripheral blood vessels, and pulmonary veins; the systemic circulation model includes systemic arteries, peripheral blood vessels, systemic veins, and vena cava; the vascular bed model is divided into the vascular beds of active muscles and resting muscles, taking into account the local vasodilation mechanism during exercise; the venous return model includes a muscle pump and a respiratory pump, the muscle pump refers to the effect of muscle contraction on venous return, and the respiratory pump refers to the effect of breathing on venous return; preferably: the cardiovascular system includes a heart model, a pulmonary circulation, a systemic circulation, a respiratory pump, and a muscle pump.

3. According to claim 1, a method for implementing sports health management based on an individualized cardiopulmonary function digital twin model is characterized in that: The cardiovascular control system is used to simulate the regulation of the cardiovascular and cerebrovascular systems by neural regulatory mechanisms; it includes a local blood flow rate regulation model, a central nervous system ischemic response model, afferent and efferent neural pathways, and a reflex control model of cardiovascular effectors; among them, the local blood flow rate regulation model simulates the vasodilation response of the brain, coronary arteries, and skeletal muscle vascular beds; the central nervous system ischemic response model simulates the effects of hypoxia on heart rate and systemic circulation; the afferent and efferent neural pathways simulate the carotid sinus pressure receptors, peripheral chemoreceptors, and lung expansion reflex; the reflex control model of cardiovascular effectors simulates the sympathetic and parasympathetic regulation of heart rhythm, ventricular contractility, peripheral vascular resistance, and venous capacity; cardiovascular control also includes the effects of respiratory central nervous system muscle drive on the efferent pathway, and evaluates the central command mechanism as a metabolic regulation response.

4. According to claim 1, a method for implementing sports health management based on an individualized cardiopulmonary function digital twin model is characterized in that: The respiratory mechanics system is used to simulate the generation of airflow and changes in lung volume during breathing, as well as the adjustment of respiratory rate and tidal volume during exercise; It includes a model representing the mechanical properties of the lungs and a model representing the mechanical properties of the upper respiratory tract.

5. According to claim 1, a method for implementing sports health management based on an individualized cardiopulmonary function digital twin model is characterized in that: The respiratory control system uses information from central chemical receptors, peripheral chemical receptors, and ventilation-related metabolic driving components to estimate ventilation needs, forming a ventilation controller; at the same time, the respiratory controller also includes a respiratory pattern optimization method, which minimizes circulatory breathing work by adjusting the respiratory pattern in each respiratory cycle, forming a respiratory pattern optimizer.

6. According to claim 1, a method for implementing sports health management based on an individualized cardiopulmonary function digital twin model is characterized in that: The gas exchange system is used to simulate the transport and exchange of oxygen and carbon dioxide during exercise, and to maintain the stability of gas concentration in the blood; it includes a gas mixing and exchange model, a gas transport model, and a metabolic kinetic model; among them, the gas mixing and exchange model takes into account the mixing of alveolar gases, blood gas exchange, and gas exchange at the tissue level; the gas transport model simulates the transport and metabolism of oxygen and carbon dioxide in the brain and tissues; the metabolic kinetic model regulates respiratory demand through metabolic rate ratios and metabolism-related neural drives; the cardiovascular system and respiratory mechanics interact through the gas exchange system to regulate blood gas concentrations.

7. According to claim 1, a method for implementing sports health management based on an individualized cardiopulmonary function digital twin model is characterized in that: The information interaction relationship between the physical models of the physiological mechanism of human cardiopulmonary function is specifically as follows: For the cardiovascular control system: metabolic regulation requires external input of the exhaled carbon dioxide volume V CO2 and anaerobic threshold AT, outputting aerobic exercise relative intensity I; respiratory system neuromuscular drive requires ventilation volume VE output by ventilation controller, induction time TI and blood flow rate BF output by respiratory pattern optimizer, outputting central respiratory neuromuscular drive response Nt; efferent pathway requires metabolic regulation output I, respiratory system neuromuscular drive output Nt, and controlled variable θ output by central nervous system ischemic response aj and the various afferent activities of the afferent pathway outputs aj , output various outgoing activities f sj and the spike frequency f of the vagal efferent fibers v ; The effector of reflex control requires metabolic regulation output I and efferent pathway output f aj and f v The coronary artery area hp, the skeletal static muscle area rmp, the active muscle area amp, and the brain area bp fluid group pressure R of the local blood flow control output jp , the output blood flow is controlled locally by the smooth muscle area ep, spleen area sp, coronary artery area hp, skeletal static muscle area rmp, active muscle area amp, brain area bp fluid group pressure R jp,t , heart rate HR, left and right ventricular elasticity E at the moment of maximum contraction max,jv , left atrial blood flow Q la and bone peripheral resistance R amp,n ; The ischemic response of the central nervous system requires gas exchange and mixing to output the brain carbon dioxide partial pressure PaCO2 and brain oxygen partial pressure PaO2, and the output controlled variable θ sj The afferent pathway requires gas exchange and mixing output PaCO2, PaO2, tidal volume VT of lung mechanical output, and P of systemic circulation output SA , output f aj ; Local control of blood flow requires skeletal peripheral resistance R of effector output of reflex control amp,n , metabolic regulation output I, gas exchange and mixing output PaCO2 and systemic circulation output inflow Q jp , output R jp ; For the cardiovascular system: the relative intensity of aerobic exercise I for systemic circulation metabolic regulation output and the circulating blood volume V for effector output of reflex control u,jv , the pressure R of each partition liquid group at time T output by the effector of the reflection control jp,t , the extravascular pressure P of the active muscle vein output by the muscle pump im , the abdominal pressure P output by the respiratory pump abd , the intrathoracic pressure P output by the respiratory pump thor , pulmonary artery blood volume V output of pulmonary circulation pulmonary , left ventricular blood flow Q of cardiac output lv , right atrial pressure P of cardiac output ra and cardiac output of cardiac blood volume V heart , output right atrial blood flow Q ra , systemic arterial pressure P SA and the inflow Q jp ; The right atrial blood flow Q that the heart needs to output from the systemic circulation ra , heart rate HR output by the effector of reflex control, left and right ventricular elasticity E at the moment of maximum contraction output by the effector of reflex control max,jv , left atrial blood flow Q of pulmonary circulation output la and the breathing pump output P thor , output left ventricular blood flow Q lv , right atrial pressure P ra 、Heart blood volume V heart and right ventricular blood flow Q rv ; Pulmonary circulation requires right ventricular blood flow Q rv and the intrathoracic pressure P output by the respiratory pump thor , output pulmonary peripheral circulation blood flow Q pp , pulmonary artery blood volume V pulmonary and left atrial blood flow Q la ; The respiratory pump requires the induction time TI output by the respiratory pattern optimizer, the tidal volume VT output by the lung mechanics, and the blood flow rate BF output by the respiratory pattern optimizer, and outputs the abdominal pressure P abd and intrathoracic pressure P thor ; The muscle pump does not need input, and outputs the extravascular pressure P of the active muscle vein im ; For the respiratory control system: the mean detector needs the brain oxygen partial pressure PaO2 output by gas exchange and mixing, the brain carbon dioxide partial pressure PaCO2 output by gas exchange and mixing, and the pressure of carbon dioxide in cerebral venous blood PbCO2 output by gas transport, and outputs the PaO2 average value PamO2, the PaCO2 average value PamCO2, and the PbCO2 average value PbmCO2; the ventilation controller needs the PaO2 average value PamO2, the PaCO2 average value PamCO2, the PbCO2 average value PbmCO2 output by the mean detector, the metabolism-related ventilation neural drive component MRV output by metabolic dynamics, and the blood flow rate BF output by the respiratory pattern optimizer, and outputs the ventilation volume VE; the respiratory pattern optimizer needs the ventilation volume VE output by the ventilation controller and the muscle pressure signal P output by lung mechanics musc , output induction time TI, blood flow rate BF and neural drive response Nd; For the gas exchange system: gas exchange and mixing require blood volume V output by lung mechanical output and inflow Q output by systemic circulation jp , left atrial blood flow Q of pulmonary circulation output la , the external input oxygen concentration FiO2, the external input carbon dioxide concentration FiCO2, the external input atmospheric pressure P atm , pulmonary circulation output of pulmonary peripheral circulation blood flow Q pp , tidal volume VT of pulmonary mechanics output, mixed venous carbon dioxide concentration CvCO2 of gas transport output and mixed venous oxygen concentration CvO2 of gas transport output, output arterial carbon dioxide concentration CaCO2 and arterial oxygen concentration CaO2; gas transport requires cerebral carbon dioxide partial pressure PaCO2 of gas exchange and mixing output, arterial carbon dioxide concentration CaCO2 of gas exchange and mixing output, arterial oxygen concentration CaO2 of gas exchange and mixing output, and inflow Q of systemic circulation output jp , external input carbon dioxide uptake VCO2, tidal volume VT output by lung mechanics and external input oxygen uptake VO2, output mixed venous carbon dioxide concentration CvCO2, mixed venous oxygen concentration CvO2, metabolic generation rate of carbon dioxide MRTCO2, metabolic generation rate of oxygen MRTO2 and pressure of carbon dioxide in cerebral venous blood PbCO2; metabolic dynamics require metabolic generation rate of oxygen output by gas transport MRTO2 and metabolic generation rate of carbon dioxide output by gas transport MRTCO2, output metabolism-related ventilation neural drive component MRV; For the respiratory dynamics system: the blood volume V of the upper respiratory tract requiring lung mechanical output, the pleural pressure P of the lung mechanical output pl and tidal volume VT of lung mechanical output, output airway flow gain G AW ; Lung mechanics requires airway flow gain G of upper airway output AW , external tracheal pressure P ao And the neural drive response Nd output by the respiratory pattern optimizer, output blood volume V, pleural pressure P pl , muscle pressure signal P musc and tidal volume VT.

8. The method for implementing sports health management based on an individualized cardiopulmonary function digital twin model according to claim 1, characterized in that: The specific network training stages of the first neural network are as follows: 1) Pre-training: The input data comes from the abnormal data of each indicator of normal people that are randomly modified, and the results of each indicator measured by the subjects through cardiopulmonary exercise test (CPET); The output data comes from the results of various indicators measured by some subjects through cardiopulmonary exercise test CPET, and the internal function parameters of the human body obtained by calculating the results of various indicators measured by other subjects through cardiopulmonary exercise test CPET in the first step of the physiological mechanism model; First, the network weights are initialized and the corresponding parameters in the input data set are forward propagated. The data passes through three LSTM hidden layers in sequence. In each hidden layer, it is processed by weighted summation and nonlinear activation function. Finally, the predicted value is generated in the output layer. For the difference between the network predicted values ​​and the values ​​generated by the simulated data set; The loss function is calculated by the mean square error and is described as: Where L represents the loss function; y i It can represent the internal function parameters of the human body calculated in the previous step and the true value of the CPET test data; represents the corresponding predicted value of the first neural network pre-training; N refers to the number of output data of the first neural network; In the back-propagation process, first, the gradient of the loss function relative to the network output needs to be calculated, and the gradient of each weight and bias is calculated layer by layer through the chain rule; these gradients represent the contribution of each parameter to the final error; then, through the stochastic gradient descent algorithm, the gradient updates the network weights to reduce the loss, so that the network continuously adjusts its weight parameters to minimize the loss function; to prevent overfitting, regularization technology is used after each hidden layer: during the training process, the output of some neurons in the network is randomly set to zero to avoid the model being too complex or overfitting the training data; By repeatedly iterating the above forward propagation and back propagation, until the loss function tends to stabilize and converge; 2) Secondary training: The input and output of the secondary training dataset only require the real index results measured by the subjects through cardiopulmonary exercise test (CPET), and do not require the internal functional parameters of the human body calculated by the first step physiological mechanism model; The process of secondary training is as follows: First, the pre-trained network parameters are used as the initial weights, the input is the same as in the pre-training stage, and the predicted value is generated at the output layer through the same network structure; Then, the loss function is used to calculate the CPET test index parameters from the mean square error according to formula (2), which is the difference between the network prediction value and the true measurement value; the back propagation process is consistent with the pre-training stage; by minimizing the loss function, the network output can more accurately fit the individual test data and optimize the network performance; The above forward propagation and back propagation are iterated repeatedly until the loss function converges to a stable state.

9. The method for implementing sports health management based on an individualized cardiopulmonary function digital twin model according to claim 1, characterized in that: In the second step: Since physiological indicators change dynamically with exercise intensity, the input data is expressed in the form of a time series. The specific structure is a three-dimensional tensor, whose dimensions are defined as the number of samples n, the time step Δt, and the number of features x. i ; In order to meet the processing requirements of neural networks, the input data needs to be processed by Z-score standardization method: Among them, x i is the original data point, μ is the mean of the data, and σ is the standard deviation of the data.

Images