Intelligent myocardial infarction risk assessment method based on cardiopulmonary exercise test data

Abstract

The invention discloses an intelligent myocardial infarction risk assessment method based on cardiopulmonary exercise test data, and relates to the field of crossing of medical science and artificial intelligence, and the method comprises the steps: carrying out the data stage division of the obtained cardiopulmonary exercise test data, obtaining cardiopulmonary exercise test stage data, carrying out the abnormal value processing, and obtaining a cardiopulmonary exercise test stage data; the preprocessed cardiopulmonary exercise test data are obtained; oxygen pulse parameters are calculated, data reconstruction is conducted on the oxygen pulse parameters through a multi-scale fitting method, and the ratio of the slope of the back slope to the slope of the front slope is calculated; constructing a regression fitting equation by using a logarithmic regression model, extracting a slope coefficient of the regression fitting equation as an oxygen uptake efficiency slope, and performing multi-level enhancement processing to obtain an oxygen uptake efficiency slope comprehensive feature; and constructing an expanded joint feature vector, and inputting the joint feature vector into a myocardial infarction risk prediction model for myocardial infarction risk prediction. According to the method, comprehensive characterization of the dynamic regulation capability and the metabolic efficiency level of the cardiopulmonary system in the exercise load process is realized, and the accuracy and the stability of risk prediction are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of medicine and artificial intelligence, in particular to a method for intelligent assessment of myocardial infarction risk based on cardiopulmonary exercise test data. BACKGROUND

[0002] The cardiopulmonary system is the core of the human body to maintain life activities, and is directly related to the efficiency and stability of the whole body blood circulation and gas exchange in the body. Among them, myocardial infarction (referred to as heart attack) is a common acute coronary event, usually caused by acute coronary artery occlusion. Systematic and dynamic assessment of cardiopulmonary function in individuals not only has important clinical value in the risk assessment of heart attack, but also has wide application needs in rehabilitation intervention and health management.

[0003] In recent years, the oxygen uptake efficiency slope (Oxygen Uptake Efficiency Slope, OUES) has gradually become a new functional evaluation index in cardiopulmonary exercise test (CPET). This index can reflect the coupling efficiency of ventilation (VE) and oxygen uptake (VO2) under exercise load, and has certain advantages in quantitative analysis of cardiopulmonary function. However, existing studies have shown that OUES as a single index has limitations, that is, it is difficult to represent the dynamic law evolving with time during exercise, and cannot fully reflect the adjustment speed and recovery ability of the cardiopulmonary system; at the same time, this index is easily disturbed by static differences such as basal metabolic level and individual training state, resulting in insufficient stability and clinical universality of the model.

[0004] In view of the problems in the related art, no effective solution has been proposed so far. SUMMARY

[0005] In view of the problems in the related art, the present application proposes a method for intelligent assessment of myocardial infarction risk based on cardiopulmonary exercise test data to overcome the above technical problems existing in the prior art.

[0006] To this end, the specific technical solutions adopted by the present application are as follows:

[0007] A method for intelligent assessment of myocardial infarction risk based on cardiopulmonary exercise test data, the method comprising the following steps:

[0008] S1, dividing the obtained cardiopulmonary exercise test data into data stages to obtain cardiopulmonary exercise test stage data, and performing abnormal value processing on the cardiopulmonary exercise test stage data to obtain preprocessed cardiopulmonary exercise test data;

[0009] S2, calculate oxygen pulse parameters based on the pre-processed cardiopulmonary exercise test data, reconstruct the oxygen pulse parameters by using a multi-scale fitting method to obtain reconstructed oxygen pulse parameters, and calculate the ratio of the post-slope to the pre-slope according to the reconstructed oxygen pulse parameters;

[0010] S3, construct a regression fitting equation by using a logarithmic regression model according to the pre-processed cardiopulmonary exercise test data, extract the slope coefficient of the regression fitting equation as the oxygen uptake efficiency slope, and perform multi-level enhancement processing on the oxygen uptake efficiency slope to obtain an oxygen uptake efficiency slope comprehensive feature;

[0011] S4, construct an extended joint feature vector according to the ratio of the post-slope to the pre-slope and the oxygen uptake efficiency slope comprehensive feature, input the extended joint feature vector into a myocardial infarction risk prediction model to perform myocardial infarction risk prediction, and divide the myocardial infarction risk prediction result into risk levels to obtain a myocardial infarction risk level.

[0012] Further, the obtained cardiopulmonary exercise test data is divided into data stages to obtain cardiopulmonary exercise test stage data, and the cardiopulmonary exercise test stage data is processed for abnormal values to obtain pre-processed cardiopulmonary exercise test data, including:

[0013] S11, obtain the cardiopulmonary exercise test data, and perform preliminary abnormal value processing on the cardiopulmonary exercise test data according to a pre-set physiological parameter interval to obtain initial cardiopulmonary exercise test data;

[0014] S12, divide the initial cardiopulmonary exercise test data into data stages by using a staged hidden semi-Markov model to obtain cardiopulmonary exercise test stage data;

[0015] The cardiopulmonary exercise test stage data includes: resting stage data, incremental load stage data, peak stage data, and recovery stage data;

[0016] S13, process the resting stage data, incremental load stage data, peak stage data, and recovery stage data for abnormal values respectively, and verify the abnormal value processing results for quality to obtain pre-processed cardiopulmonary exercise test stage data.

[0017] Further, the initial cardiopulmonary exercise test data is divided into data stages by using a staged hidden semi-Markov model to obtain cardiopulmonary exercise test stage data, including:

[0018] S121, fit the duration distribution of the cardiopulmonary exercise test stage based on pre-defined one-way stage transition and observation sequence fitting, and construct a staged hidden semi-Markov model;

[0019] S122, according to the pre-labeled cardiopulmonary exercise test data, the feature dimension of the observation sequence is expanded to obtain the expanded observation sequence, and the expectation maximization algorithm is used to train the staged hidden semi-Markov model based on the expanded observation sequence, and the trained staged hidden semi-Markov model is obtained;

[0020] S123, the initial cardiopulmonary exercise test data is decoded by using the improved Viterbi algorithm to obtain the hidden state sequence, and the stage boundary is obtained according to the hidden state sequence, and the stage reasoning and correction of the initial cardiopulmonary exercise test data are carried out based on the stage boundary, and the cardiopulmonary exercise test stage data is obtained.

[0021] Further, the oxygen pulse parameter is calculated based on the preprocessed cardiopulmonary exercise test data, the multi-scale fitting method is used for data reconstruction of the oxygen pulse parameter, and the reconstructed oxygen pulse parameter is obtained, and the ratio of the post-slope to the pre-slope is calculated according to the reconstructed oxygen pulse parameter, including:

[0022] S21, based on the preprocessed cardiopulmonary exercise test data, the oxygen pulse parameter is calculated, the wavelet function is used for wavelet decomposition of the oxygen pulse parameter, and the stationary sequence is obtained, and the dominant trend signal is extracted according to the stationary sequence;

[0023] S22, the linear fitting model is used for linear fitting of the dominant trend signal, and the global optimization of the linear fitting result is carried out by using the simulated annealing algorithm with the minimum regression loss function as the target, and the trend function is obtained;

[0024] S23, based on the trend function, the data of the oxygen pulse parameter is reconstructed to obtain the reconstructed oxygen pulse parameter, the exercise stage time range is defined according to the preprocessed cardiopulmonary exercise test data, and the ratio of the post-slope to the pre-slope is calculated based on the reconstructed oxygen pulse parameter and the exercise stage time range.

[0025] Further, the oxygen pulse parameter is calculated based on the preprocessed cardiopulmonary exercise test data, the wavelet function is used for wavelet decomposition of the oxygen pulse parameter, and the stationary sequence is obtained, and the dominant trend signal is extracted according to the stationary sequence, including:

[0026] S211, based on the preprocessed cardiopulmonary exercise test data, the oxygen pulse parameter is calculated, the wavelet function is used for wavelet decomposition of the oxygen pulse parameter, and the oxygen pulse approximate component is obtained, the stationary sequence is obtained by enhancing the Dickey-Fuller unit root test on the oxygen pulse approximate component, and the judgment parameter is obtained;

[0027] S212, the significance test is carried out on the judgment parameter, and the sequence screening is carried out on the oxygen pulse approximate component according to the significance test result to obtain the stationary sequence;

[0028] S213, the frequency spectrum of the stationary sequence is analyzed by using the fast Fourier transform, and the dominant trend is extracted according to the frequency spectrum analysis result to obtain the dominant trend signal.

[0029] Further, a regression fitting equation is constructed according to the preprocessed cardiopulmonary exercise test data by using a logarithmic regression model, and a slope coefficient of the regression fitting equation is extracted as an oxygen uptake efficiency slope, and the oxygen uptake efficiency slope is processed by multi-level enhancement to obtain an oxygen uptake efficiency slope comprehensive feature including:

[0030] S31, a regression fitting equation is constructed according to the preprocessed cardiopulmonary exercise test data by using a logarithmic regression model, and a slope coefficient of the regression fitting equation is extracted as an oxygen uptake efficiency slope;

[0031] S32, a ventilation threshold is set based on the preprocessed cardiopulmonary exercise test data by using a ratio inflection point method, a metabolic stage is divided according to the ventilation threshold, and a segmented weight is generated, the oxygen uptake efficiency slope is processed by weighted fusion based on the segmented weight, and a weighted oxygen uptake efficiency slope is obtained;

[0032] S33, a time sequence is extracted from the preprocessed cardiopulmonary exercise test data based on recovery period dynamics analysis, a nonlinear invariant is obtained, and the nonlinear invariant is processed by multi-modal fusion to obtain a nonlinear enhanced oxygen uptake efficiency slope;

[0033] S34, the oxygen uptake efficiency slope, the weighted oxygen uptake efficiency slope and the nonlinear enhanced oxygen uptake efficiency slope are integrated by using a normalized weighted fusion method to obtain an oxygen uptake efficiency slope comprehensive feature.

[0034] Further, the oxygen uptake efficiency slope, the weighted oxygen uptake efficiency slope and the nonlinear enhanced oxygen uptake efficiency slope are integrated by using a normalized weighted fusion method to obtain an oxygen uptake efficiency slope comprehensive feature including:

[0035] S331, a time period is defined based on the preprocessed cardiopulmonary exercise test data by using recovery period dynamics analysis, and a time sequence is extracted from the preprocessed cardiopulmonary exercise test data according to the time period definition result to obtain a nonlinear invariant;

[0036] S332, the nonlinear invariant is processed by denoising to obtain a denoised nonlinear invariant, and the denoised nonlinear invariant is standardized by using a Z-score method to obtain a standard nonlinear invariant;

[0037] S333, a phase space reconstruction parameter is optimized based on the standard nonlinear invariant and the preprocessed cardiopulmonary exercise test data, a dynamics invariant is extracted according to the optimization result, and the dynamics invariant is processed by multi-modal fusion to obtain a nonlinear enhanced oxygen uptake efficiency slope.

[0038] Further, the phase space reconstruction parameters are optimized based on the standard nonlinear invariants and the pretreated cardiopulmonary exercise test data, and the kinetic invariants are extracted according to the optimization results, the kinetic invariants are fused in multiple modes, and the nonlinear enhanced oxygen uptake efficiency slope is obtained, including:

[0039] S3331, determining the time delay and embedding dimension based on the standard nonlinear invariants and the pretreated cardiopulmonary exercise test data, optimizing the phase space reconstruction parameters through the time delay and embedding dimension, and obtaining the trajectory optimization result;

[0040] S3332, extracting the maximum Lyapunov exponent, correlation dimension and cross-sectional offset rate according to the trajectory optimization result, and integrating the maximum Lyapunov exponent, correlation dimension and cross-sectional offset rate to obtain the kinetic invariant;

[0041] S3333, performing dynamic time warping on the trajectory optimization result, constructing a tensor according to the dynamic time warping result, and performing multi-modal phase space fusion based on the tensor and the kinetic invariant to obtain the nonlinear enhanced oxygen uptake efficiency slope.

[0042] Further, an extended joint feature vector is constructed according to the ratio of the post-slope to the pre-slope and the oxygen uptake efficiency slope comprehensive characteristics, the extended joint feature vector is input into the myocardial infarction risk prediction model to predict the myocardial infarction risk, and the myocardial infarction risk prediction result is classified according to the risk level to obtain the myocardial infarction risk level, including:

[0043] S41, square expanding the ratio of the post-slope to the pre-slope and the oxygen uptake efficiency slope comprehensive characteristics, and constructing an extended joint feature vector based on the square expansion result;

[0044] S42, constructing a multi-layer perception machine model based on the extended joint feature vector, dividing the extended joint feature vector into a training set and a test set, training the multi-layer perception machine model using the training set, and obtaining a myocardial infarction risk prediction model;

[0045] S43, using the myocardial infarction risk prediction model to predict the myocardial infarction risk of the test set, and classifying the myocardial infarction risk prediction result according to the preset risk level division rule to obtain the myocardial infarction risk level.

[0046] Further, a multi-layer perception machine model is constructed based on the extended joint feature vector, the extended joint feature vector is divided into a training set and a test set, the multi-layer perception machine model is trained using the training set, and a myocardial infarction risk prediction model is obtained, including:

[0047] S421, constructing a multi-layer perception machine model based on the extended joint feature vector, and initializing the parameters of the multi-layer perception machine model to obtain an initial multi-layer perception machine model;

[0048] S422, divide the extended joint feature vector into a training set and a test set, train the initial multilayer perceptron model using the training set, and obtain a preliminary myocardial infarction risk prediction result;

[0049] S423, calculate the prediction error using the cross-entropy loss function according to the preliminary myocardial infarction risk prediction result, and optimize the initial multilayer perceptron model using back propagation according to the error calculation result, to obtain a myocardial infarction risk prediction model.

[0050] The beneficial effects of the present application are:

[0051] 1. The present application fuses the slope ratio of the post-slope and the pre-slope and the oxygen uptake efficiency slope comprehensive feature, combines the multilayer perceptron model to model the complex nonlinear relationship, realizes the comprehensive representation of the dynamic regulation ability and the metabolic efficiency level of the cardiopulmonary system in the process of exercise load, thereby improving the accuracy and stability of risk prediction, and promoting the development of cardiopulmonary exercise test data to data-driven intelligent direction. At the same time, the model output is presented in the form of risk probability, which enhances the data interpretability and practical application value of the results. Avoid the problems of difficulty in nonlinear feature modeling and insufficient accuracy of traditional methods in myocardial infarction risk prediction, realize the accurate modeling and intelligent discrimination of individual myocardial infarction risk.

[0052] 2. The present application quantifies the dynamic regulation ability of the cardiopulmonary system in the process of load and recovery through the extended joint feature vector. At the same time, the oxygen uptake efficiency slope is enhanced in multiple levels, the correlation is improved through metabolic stage weighting, the nonlinear features are extracted combined with recovery period dynamics analysis, and finally the comprehensive OUES index is formed, realizing the intelligent prediction and grading of myocardial infarction risk, providing efficient and accurate decision support for myocardial infarction risk assessment. BRIEF DESCRIPTION OF DRAWINGS

[0053] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0054] Figure 1 is a flow chart of a myocardial infarction risk intelligent evaluation method based on cardiopulmonary exercise test data according to an embodiment of the present application;

[0055] Figure 2 is a curve graph of O2-pulse and heart rate changing with time in a typical cardiopulmonary test process according to a myocardial infarction risk intelligent evaluation method based on cardiopulmonary exercise test data according to an embodiment of the present application;

[0056] Figure 3 is a trend segmentation and slope fitting result schematic diagram of O2-pulse after filtering processing in an intelligent assessment method of myocardial infarction risk based on cardiopulmonary exercise test data according to an embodiment of the application.

[0057] Figure 4 is an MLP neural network structure schematic diagram in an intelligent assessment method of myocardial infarction risk based on cardiopulmonary exercise test data according to an embodiment of the application.

[0058] Figure 5 is a myocardial infarction risk prediction flowchart based on an MLP model in an intelligent assessment method of myocardial infarction risk based on cardiopulmonary exercise test data according to an embodiment of the application. DETAILED DESCRIPTION

[0059] To further illustrate the embodiments, the present application provides drawings which are part of the disclosure of the present application, mainly used to illustrate the embodiments, and can explain the operating principle of the embodiments in conjunction with the related description of the specification. Those skilled in the art should understand other possible implementations and advantages of the present application by referring to these contents.

[0060] According to an embodiment of the present application, an intelligent assessment method of myocardial infarction risk based on cardiopulmonary exercise test data is provided.

[0061] The present application will be further described in conjunction with the drawings and specific embodiments, as shown in Figure 1 and Figure 5 The intelligent assessment method of myocardial infarction risk based on cardiopulmonary exercise test data according to the embodiment of the present application includes the following steps:

[0062] S1, the cardiopulmonary exercise test data obtained is divided into data stages to obtain cardiopulmonary exercise test stage data, and the cardiopulmonary exercise test stage data is processed for abnormal values to obtain preprocessed cardiopulmonary exercise test data.

[0063] Specifically, the cardiopulmonary exercise test data obtained is divided into data stages to obtain cardiopulmonary exercise test stage data, and the cardiopulmonary exercise test stage data is processed for abnormal values to obtain preprocessed cardiopulmonary exercise test data, including:

[0064] S11, the cardiopulmonary exercise test data obtained is processed for preliminary abnormal values according to a preset physiological parameter interval to obtain initial cardiopulmonary exercise test data;

[0065] S12, the initial cardiopulmonary exercise test data is divided into data stages by using a staged hidden semi-Markov model to obtain cardiopulmonary exercise test stage data.

[0066] Specifically, the initial cardiopulmonary exercise test data is divided into stages by using a staged hidden semi-Markov model to obtain cardiopulmonary exercise test stage data, including:

[0067] S121, fitting the duration distribution of the cardiopulmonary exercise test stage based on the pre-defined stage one-way transition and observation sequence, and constructing a staged hidden semi-Markov model;

[0068] S122, according to the pre-labeled cardiopulmonary exercise test data, the feature dimension of the observation sequence is expanded to obtain the expanded observation sequence, and the staged hidden semi-Markov model is trained based on the expanded observation sequence using the expectation maximization algorithm to obtain the trained staged hidden semi-Markov model;

[0069] S123, the initial cardiopulmonary exercise test data is decoded by using an improved Viterbi algorithm to obtain a hidden state sequence, and the stage boundary is obtained according to the hidden state sequence, and the initial cardiopulmonary exercise test data is stage reasoned and corrected based on the stage boundary to obtain cardiopulmonary exercise test stage data.

[0070] The cardiopulmonary exercise test stage data includes: resting stage data, incremental load stage data, peak stage data and recovery stage data.

[0071] S13, the resting stage data, the incremental load stage data, the peak stage data and the recovery stage data are respectively subjected to abnormal value processing, and the abnormal value processing results are subjected to quality verification to obtain preprocessed cardiopulmonary exercise test stage data.

[0072] Specifically, the key physiological parameters recorded during the cardiopulmonary exercise test (CPET) process (i.e., cardiopulmonary exercise test data) are collected, including heart rate (HR), oxygen uptake (VO2), ventilation (VE), exercise load (Power) and test time (Time) and other key physiological parameters. In order to improve the accuracy of data processing, this embodiment introduces a staged hidden semi-Markov model (HSMM) and a multi-strategy outlier rejection method to automatically process the CPET process in stages, and applies different outlier rejection strategies according to the physiological characteristics of each stage, and finally forms stable, continuous and clear structure data.

[0073] Specifically, data preprocessing and outlier rejection include extreme outlier rejection of heart-lung exercise test data beyond the preset physiological parameter range, completing the first round of rough screening. Then, the four stages of rest (S1), increasing load (S2), peak (S3), and recovery (S4) of the CPET process are identified using a phased hidden semi-Markov model (HSMM). In each stage, different rejection strategies are applied. In the increasing load stage, a skew-t distribution model is used. In the peak stage, a VO2 platform tolerance zone is set. In the recovery stage, an Ornstein-Uhlenbeck fall-back model is used. Finally, stable, continuous, and clear CPET data (i.e., preprocessed heart-lung exercise test stage data) are obtained.

[0074] The four stages of rest, increasing load, peak, and recovery of the CPET process are identified using a phased hidden semi-Markov model (HSMM). This includes HSMM model construction, defining one-way transitions (S1, S2, S3, S4) of the four stages, and using 8-dimensional features of "HR-VO2-VE-Power" as observation sequences. These features include HR instantaneous value, VO2 instantaneous value, VE instantaneous value, Power instantaneous value, HR 5-second sliding mean value, VO2 5-second sliding mean value, VE 1-second change rate, and Power 1-second change rate. The Power parameter has dual functions: it is used as an observation sequence feature for model training, and it is used for secondary verification of stage boundaries, such as the increasing load stage (S2) requiring Power 1-second change rate ≥ 0 and the peak stage (S3) requiring Power instantaneous value ≥ 90% peak Power. Based on 1000 normal data, the duration distribution of each stage is fitted (S1 ~ N(300, 60), S2 ~ Gamma(10, 60), S3 ~ Poisson(30), S4 ~ Weibull(2, 900), units: seconds). HSMM model training and verification include selecting 500 stage boundary labeled data, such as doctor labeling, with CPET data error ≤ 5 seconds, of which 350 are used for training and 150 for verification. The observation sequences are standardized and the first 10-second parameter slope and the first 5-second parameter variance are extracted to expand the 8-dimensional features. The model parameters are estimated using the expectation maximization (EM) algorithm, requiring the stage recognition error of the verification set to be ≤ 10 seconds and the accuracy to be ≥ 95%. If the requirements are not met, retraining is performed. Phase inference and boundary correction include decoding the hidden state sequence of new data using an improved Viterbi algorithm (i.e., the Viterbi algorithm), with state transition time being the stage boundary. If the stage duration exceeds the distribution 99% confidence interval, combined with Power changes, such as S2 requiring continuous upward trend, secondary correction is performed to ensure that the stage division conforms to the test logic.

[0075] Specifically, different rejection strategies are applied in each stage, namely, skew t-distribution model is used in the incremental load stage, VO2 platform tolerance zone is set in the peak stage, and Ornstein-Uhlenbeck rollback model is used in the recovery stage. Finally, stable, continuous and clear structure CPET data is obtained. For the resting stage (S1), the mean (μ0) and standard deviation (σ0) of HR, VO2 and VE are calculated based on the previous 2 minutes of data, and the threshold is set, that is, HR is [μ0±3σ0], VO2 is [μ0±2σ0], and VE is [μ0±4σ0]. Point-by-point screening, data exceeding the threshold and lasting ≥2 seconds are rejected; if HR appears mutation, that is, single change amount is greater than 20 bpm, but VE fluctuation amount is less than 10% / second, that is, VE has no abnormal fluctuation, it is indirectly judged as tension caused by the subject, which is not a physiological abnormality, then the median of adjacent 5 seconds data is corrected to avoid data breakage. For the incremental load stage (S2), the HR and VO2 data of S2 are extracted to fit the skew t-distribution, and the 95% confidence interval is taken as the normal range; data exceeding the interval and HR lower than the mean of the previous 10 seconds-20 bpm or VO2 lower than the mean of the previous 10 seconds-300 mL / min are rejected, and the missing data is filled by linear interpolation. For the peak stage (S3), the change amount of VO2 every 20 seconds in S3 is calculated, and when the change amount of VO2 in two consecutive windows is ≤100 mL / min, it is judged as the platform period, and the tolerance zone [μ_plat±150 mL / min] (μ_plat is the mean of the platform period) is set; data exceeding the tolerance zone and HR less than 0.9×(220-age) are rejected, and the extreme value is verified in combination with RER (≥1.1) and device state. For the recovery stage (S4), the data of the previous 3 minutes of S4 are taken to fit the Ornstein-Uhlenbeck (O-U) model (dX t =κ(μ-X t )dt+σdW t ), the expected value of each time point is predicted, and [E[X t ]±2σ] is taken as the normal range; data exceeding the range and lasting ≥3 seconds are rejected, and if HR falls slowly but SpO2 is less than 90%, it is retained, otherwise it is judged as a delayed signal rejection of the device. Data quality verification includes effective data rate ≥90%, missing rate ≤5%, stage boundary and power change matching error ≤10 seconds, that is, S2 requires continuous power increase, S4 requires power=0, and HR-VO2 correlation coefficient ≥0.8. After verification, three types are output, one is the stage information table, that is, containing the start and end time, duration, and effective data amount of each stage, and S2=incremental load stage, S4=recovery stage are labeled, which is used for subsequent slope calculation; the second is CSV time series data, that is, containing Time, HR, VO2, VE, Power and stage label, that is, 1-resting, 2-incremental load, 3-peak, 4-recovery; the third is the preprocessing report, which is used as the basis for subsequent analysis.

[0076] S2, calculate an oxygen pulse parameter based on the preprocessed cardiopulmonary exercise test data, reconstruct the oxygen pulse parameter by using a multi-scale fitting method to obtain a reconstructed oxygen pulse parameter, and calculate a ratio of a post-slope to a pre-slope slope based on the reconstructed oxygen pulse parameter.

[0077] Specifically, the method comprises the following steps:

[0078] S21, calculate an oxygen pulse parameter based on the preprocessed cardiopulmonary exercise test data, perform wavelet decomposition on the oxygen pulse parameter by using a wavelet function to obtain a stationary sequence, and extract a dominant trend signal based on the stationary sequence.

[0079] Specifically, the method comprises the following steps:

[0080] S211, calculate an oxygen pulse parameter based on the preprocessed cardiopulmonary exercise test data, perform wavelet decomposition on the oxygen pulse parameter by using a wavelet function to obtain an oxygen pulse approximate component, perform stationarity determination on the oxygen pulse approximate component by using an enhanced Dickey-Fuller unit root test to obtain a determination parameter.

[0081] S212, perform a significance test on the determination parameter, perform sequence screening on the oxygen pulse approximate component based on a result of the significance test to obtain a stationary sequence.

[0082] S213, perform frequency spectrum analysis on the stationary sequence by using a fast Fourier transform, and extract a dominant trend based on a result of the frequency spectrum analysis to obtain a dominant trend signal.

[0083] S22, perform linear fitting on the dominant trend signal by using a linear fitting model, and perform global optimization on a result of the linear fitting by using a simulated annealing algorithm with a minimized regression loss function as a target to obtain a trend function.

[0084] S23, reconstruct the oxygen pulse parameter based on the trend function to obtain a reconstructed oxygen pulse parameter, define a time range of an exercise stage based on the preprocessed cardiopulmonary exercise test data, and calculate a ratio of a post-slope to a pre-slope slope based on the reconstructed oxygen pulse parameter and the time range of the exercise stage.

[0085] Specifically, based on preprocessed cardiopulmonary exercise test stage data, an oxygen pulse (O2-pulse=VO2 / HR) parameter as shown in Figure 2 reflects an oxygen uptake efficiency per heartbeat. The present application introduces a multi-scale fitting method combining wavelet transform and Huber regression on the basis of traditional processing, as shown in Figure 3As shown, the ascending slope Slope1 and descending slope Slope2 are extracted, and the ratio of the slope of the back slope to the slope of the front slope, Slope Ratio=Slope2 / Slope1, is calculated to construct a quantifiable dynamic reserve capacity index.

[0086] like Figure 3 As shown in the diagram, the upward slope Slope1 is extracted from the start of the test up to the second-to-last minute, representing the cardiopulmonary system's load response capacity; the downward slope Slope2 is extracted in the last two minutes, measuring its recovery speed. By calculating the ratio of the backward slope to the forward slope, Slope Ratio = Slope2 / Slope1, a quantifiable dynamic reserve capacity index is constructed.

[0087] Specifically, based on the two key parameters of heart rate (HR) and oxygen uptake (VO2) from the preprocessed cardiopulmonary exercise test data, the oxygen pulse (O2-pulse) parameter was calculated, i.e., O2-pulse = VO2 / HR. Furthermore, the slope of the upward trend of the oxygen pulse was extracted from the time interval from the start of the exercise test to the second-to-last minute, denoted as Slope1. This phase corresponds to the gradual acceptance of exercise load by the cardiopulmonary system and the continuous increase in metabolic level, reflecting the functional response level of the cardiopulmonary system to exercise stimulation and providing a key reference for identifying cardiopulmonary adaptability. In the recovery phase after the exercise test, the last two minutes were selected as the analysis time window, and the slope of the downward trend of the oxygen pulse over time was extracted, denoted as Slope2. This slope reflects the speed and efficiency of the individual's cardiopulmonary system's return from a hypermetabolic state to a resting level after the external load is removed, reflecting its recovery and regulatory capacity and autonomic nervous system function. After obtaining the upward slope Slope1 during the exercise load phase and the downward slope Slope2 during the recovery phase, the ratio of the backward slope to the forward slope, i.e., Slope Ratio = Slope2 / Slope1, is further calculated as a comprehensive characteristic indicator for measuring the dynamic reserve capacity of the cardiopulmonary system.

[0088] Specifically, to extract the dominant trend information (i.e., dominant trend signal) from the oxygen pulse signal (i.e., oxygen pulse parameters), the fourth-order (db4) De Bech wavelet function is used to perform a 5-level decomposition on the oxygen pulse (O2-pulse=VO2 / HR) time series signal. The fifth-level approximate component is extracted as a low-frequency trend candidate. The original oxygen pulse signal is decomposed layer by layer into approximate components and detail components. Its mathematical expression is as follows:

[0089] ;

[0090] In the formula, denoted by j, which represents the approximation coefficient of the j-th layer, corresponding to the low-frequency part; k represents the position index after the decomposition of this layer, that is, the position number after the signal length of each layer is halved; denotes the jth layer detail coefficient corresponding to the high frequency part; denotes the sequence of the last layer approximation component; m denotes the discrete time index, which takes all integer values; denotes the low-pass filter coefficient corresponding to the 4th order wavelet; denotes the high-pass filter coefficient; j denotes the number of wavelet decomposition layers, which is 1 to 5 here.

[0091] Subsequently, the oxygen pulse approximation component is judged for stationarity by enhanced Dickey-Fuller unit root test, and the sequence to be tested is z t , and a regression equation is constructed:

[0092] ;

[0093] In the formula, denotes the first-order difference of the sequence z t , which is used to eliminate the trend; z t denotes the value of the time series at time t, which is the trend component obtained after wavelet decomposition; a denotes the coefficient to be estimated, and whether it is 0 is tested; if not, it is indicated that the sequence is stationary; denotes the regression coefficient of the rth-order lagged difference; denotes the white noise disturbance term.

[0094] The ADF statistic is obtained through the significance test of the parameter a, and if the significance probability is less than 5%, it is identified as a stationary sequence. The frequency spectrum of the oxygen pulse approximation component is analyzed by combining fast Fourier transform, and it is confirmed that the dominant frequency is lower than 0.05 Hz, that is, it meets the slow regulation rhythm of the cardiopulmonary system; further, the proportion of the component in the total energy of the oxygen pulse is calculated, and if it is more than 60%, it is identified as the dominant trend, which can be used for accurate extraction of the oxygen pulse slope in the subsequent process.

[0095] After the trend signal is extracted, the Huber regression method is used for robust fitting of the 5th layer trend signal. By defining the Huber loss function, combining the first-order linear fitting model, and enhancing the robustness to outliers, the global optimization is performed by using the simulated annealing algorithm, the temperature is gradually reduced to avoid local optimization, and the convergence to the global optimal solution is ensured. That is, the minimum Huber loss function is taken as the target, the parameters to be optimized are the slope and intercept of the first-order linear model, the parameter search interval is set according to the statistical physiological range in the initial stage, and the initial temperature and iteration number are determined based on the characteristics of the loss function; in the iteration, the neighborhood solution is generated by disturbance, and the solution is accepted or rejected according to the Metropolis criterion, and the temperature is reduced by using the geometric attenuation method until the temperature is reduced to a minimum value or the loss value is stable; the key parameters are set according to the clinical CPET data and statistical standards, and the optimization accuracy and efficiency are balanced. Finally, the regression fitting curve with global trend, local smoothness and anti-interference ability is output, which provides a reliable basis for subsequent trend reconstruction.

[0096] ;

[0097] where r represents the fitting residual; δ represents the switching threshold, controlling the outlier identification range; represents the Huber loss function, which is used to evaluate the deviation between the regression curve and the oxygen pulse trend signal.

[0098] After the Huber regression is completed, a regression curve is obtained, which is fitted by minimizing the Huber loss function. The regression curve is actually a robust approximation of the 5th layer wavelet approximation component, and its expression can be written as:

[0099] ;

[0100] where represents the fitted trend function, which has both global trend and local robustness, represents the smoothed value at time t, represents the Huber regression intercept, represents the Huber regression slope.

[0101] Finally, the fitted trend function is used as the core expression to reconstruct the smoothed oxygen pulse time series. By using the reconstructed VO2 and HR data for calculation, we obtain the oxygen pulse parameter value corresponding to each time point, i.e., O2-pulse = VO2 / HR.

[0102] On the basis of the obtained oxygen pulse parameters, first, the exercise phase time range is defined based on the Power parameter, i.e., the exercise start time = the time point at which Power first becomes greater than 0, and the exercise end time = the time point at which Power first drops to 0 and lasts for 10 seconds; the time interval from the exercise start time to 2 minutes before the exercise end time is selected, and the slope of the oxygen pulse rising trend is extracted, denoted as Slope1. This stage corresponds to the process of the cardiopulmonary system gradually accepting exercise load and continuously increasing metabolic level, reflecting the functional response level of the cardiopulmonary system to exercise stimulation, providing a key reference for identifying cardiopulmonary adaptation ability.

[0103] In the recovery stage after the end of the exercise test, the time range of the recovery stage is defined based on the power parameter, i.e. the recovery start time = the end time of exercise, and the recovery end time = 10 minutes after the recovery start time, i.e. the length of the clinical routine recovery observation; the analysis time window is selected from the recovery start time to 2 minutes after the recovery start time, and the descending trend slope of oxygen pulse over time in this stage is extracted, which is denoted as Slope2. The slope reflects the speed and efficiency of the individual's cardiopulmonary system falling back from a high metabolic state to a resting level after the external load is removed, and embodies the recovery regulation ability and autonomic nervous function state. After the ascending slope Slope1 in the exercise load stage and the descending slope Slope2 in the recovery stage are obtained, the ratio of the latter slope to the former slope, i.e. Slope Ratio = Slope2 / Slope1, is further calculated as a comprehensive feature index for measuring the dynamic reserve capacity of the cardiopulmonary system.

[0104] S3, constructing a regression fitting equation using a logarithmic regression model according to the preprocessed cardiopulmonary exercise test data, and extracting the slope coefficient of the regression fitting equation as the oxygen uptake efficiency slope, and performing multi-level enhancement processing on the oxygen uptake efficiency slope to obtain an oxygen uptake efficiency slope comprehensive feature.

[0105] Specifically, a regression fitting equation is constructed using a logarithmic regression model according to the preprocessed cardiopulmonary exercise test data, and the slope coefficient of the regression fitting equation is extracted as the oxygen uptake efficiency slope, and multi-level enhancement processing is performed on the oxygen uptake efficiency slope to obtain an oxygen uptake efficiency slope comprehensive feature, which includes:

[0106] S31, constructing a regression fitting equation using a logarithmic regression model according to the preprocessed cardiopulmonary exercise test data, and extracting the slope coefficient of the regression fitting equation as the oxygen uptake efficiency slope;

[0107] S32, setting a ventilation threshold based on the preprocessed cardiopulmonary exercise test data using a ratio inflection point method, dividing the metabolic stage according to the ventilation threshold and generating a segmented weight, and performing weighted fusion processing on the oxygen uptake efficiency slope based on the segmented weight to obtain a weighted oxygen uptake efficiency slope;

[0108] S33, performing time sequence extraction on the preprocessed cardiopulmonary exercise test data based on recovery period dynamics analysis to obtain a nonlinear invariant, and performing multi-modal fusion on the nonlinear invariant to obtain a nonlinear enhanced oxygen uptake efficiency slope.

[0109] Specifically, the oxygen uptake efficiency slope, the weighted oxygen uptake efficiency slope and the nonlinear enhanced oxygen uptake efficiency slope are integrated using a normalized weighted fusion method to obtain an oxygen uptake efficiency slope comprehensive feature, which includes:

[0110] S331, time period demarcation is performed on the pretreated cardiopulmonary exercise test data based on the recovery period kinetic analysis, and time sequence extraction is performed on the pretreated cardiopulmonary exercise test data according to the time period demarcation result, to obtain nonlinear invariants;

[0111] S332, the nonlinear invariants are denoised to obtain denoised nonlinear invariants, and the denoised nonlinear invariants are standardized by Z-score method to obtain standard nonlinear invariants;

[0112] S333, the phase space reconstruction parameters are optimized based on the standard nonlinear invariants and the pretreated cardiopulmonary exercise test data, and kinetic invariants are extracted according to the optimization result, and the kinetic invariants are fused in multiple modes to obtain a nonlinear enhanced oxygen uptake efficiency slope.

[0113] Specifically, the phase space reconstruction parameters are optimized based on the standard nonlinear invariants and the pretreated cardiopulmonary exercise test data, and kinetic invariants are extracted according to the optimization result, and the kinetic invariants are fused in multiple modes to obtain a nonlinear enhanced oxygen uptake efficiency slope, including:

[0114] S3331, time delay and embedding dimension are determined based on the standard nonlinear invariants and the pretreated cardiopulmonary exercise test data, and the phase space reconstruction parameters are optimized by the time delay and the embedding dimension to obtain trajectory optimization results;

[0115] S3332, maximum Lyapunov exponent, correlation dimension and cross-sectional offset rate are extracted according to the trajectory optimization results, and the maximum Lyapunov exponent, correlation dimension and cross-sectional offset rate are integrated to obtain kinetic invariants;

[0116] S3333, dynamic time warping is performed on the trajectory optimization results, and a tensor is constructed according to the dynamic time warping results, and multi-modal phase space fusion is performed based on the tensor and the kinetic invariants to obtain a nonlinear enhanced oxygen uptake efficiency slope.

[0117] S34, the oxygen uptake efficiency slope, the weighted oxygen uptake efficiency slope and the nonlinear enhanced oxygen uptake efficiency slope are integrated by using a normalized weighted fusion method to obtain an oxygen uptake efficiency slope comprehensive feature.

[0118] Specifically, in the same dynamic time interval as the oxygen pulse parameter calculation, the nonlinear relationship between oxygen uptake VO2 and ventilation VE is further analyzed, and a logarithmic regression model VO2=a×log 10(V E) + b, and the regression slope a was extracted as the oxygen uptake efficiency slope (OUES) value. To improve the correlation between OUES and clinical risk, a ventilation threshold (VT) was introduced to divide the metabolic stages and give a higher weight to the high load stage based on the traditional calculation. Meanwhile, the recovery kinetics analysis was combined to extract the nonlinear invariants and develop multi-modal fusion. Finally, the basic OUES, weighted OUES, and nonlinear enhanced OUES were unified to form the OUES comprehensive index, which combined the static efficiency, dynamic regulation, and nonlinear characteristics, providing a robust input for subsequent model construction. According to the oxygen pulse parameter calculation, the nonlinear relationship between oxygen uptake VO2 and ventilation VE within the same dynamic process was further analyzed. The logarithmic regression model VO2 = a x log 10 (V E) + b, and the regression slope a was extracted as the oxygen uptake efficiency slope (OUES) value. To improve the correlation between OUES and clinical risk, a ventilation threshold (VT) was introduced to divide the metabolic stages and give a higher weight to the high load stage based on the traditional calculation. Meanwhile, the recovery kinetics analysis was combined to extract the nonlinear invariants and develop multi-modal fusion. Finally, the basic OUES, weighted OUES, and nonlinear enhanced OUES were unified to form the OUES comprehensive index, which combined the static efficiency, dynamic regulation, and nonlinear characteristics, providing a robust input for subsequent model construction. According to the oxygen pulse parameter calculation, the nonlinear relationship between oxygen uptake VO2 and ventilation VE within the same dynamic process was further analyzed. The logarithmic regression model VO2 = a x log

[0119] From the pre-processed data, VO2 and VE parameters were selected to ensure consistency with the dynamic interval of the previous stage, ensuring the comparability and continuity of the calculation results. Based on VO2 and VE, a logarithmic regression equation was established, and the regression slope a was extracted to obtain the basic OUES value, which describes the oxygen supply efficiency under ventilation driving. Considering the differences in different metabolic stages, a ventilation threshold (VT) was set based on cardiopulmonary exercise test data to divide the aerobic and anaerobic stages, and according to the clinical verification results, a higher weight was given to the high load stage to form the weighted OUES, strengthening the clinical relevance of the index. Further, the recovery kinetics analysis was introduced to reconstruct the phase space and multi-modal fusion of VO2, VE, and O2-pulse signals, and extract nonlinear invariants such as Lyapunov exponent and correlation dimension to obtain nonlinear enhanced OUES that can characterize the stability and complexity of cardiopulmonary regulation. The basic OUES, weighted OUES, and nonlinear enhanced OUES were unified to obtain the OUES comprehensive characteristics. This comprehensive characteristic combines static efficiency, dynamic regulation ability, and nonlinear recovery characteristics, providing a more robust and comprehensive input for subsequent modeling.

[0120] From the pre-processed cardiopulmonary exercise test data, the two key parameters VE and VO2 were extracted, representing respiratory ventilation intensity and body oxygen uptake capacity, respectively. Based on the extracted VE and VO2 data, a logarithmic regression model between the two was established, and the fitting equation was VO2 = a x log 10 (V E) + b. The regression slope a is the basic OUES value, reflecting the oxygen supply efficiency under ventilation driving, and b is the intercept, which is the theoretical VO2 value when VE = 1 L / min. This model is used to describe the response trend of oxygen uptake to ventilation changes, reflecting the metabolic adaptation ability of individuals under different ventilation intensities. By analyzing the VO2 and log10 The slope coefficient a in the regression equation is extracted, and the larger the OUES value is, the higher the oxygen supply efficiency of the cardiopulmonary system under respiratory drive is, which is usually related to better cardiopulmonary fitness and metabolic reserve, and has strong clinical discriminant value. In order to enhance the pathological relevance of OUES to the risk of myocardial infarction, first, the ventilation threshold (VT) is set, and then the metabolic stage (aerobic / anaerobic) is divided, and the contribution of the high load stage is strengthened through a double weight system.

[0121] Among them, the ventilation threshold (VT) setting method is based on cardiopulmonary exercise test data, and the VE / VO2 ratio inflection point method is used to determine VT, that is, VCO2 is not required; the VE and VO2 data of the incremental load stage (S2) are smoothed by a 3-second sliding window, and the VE / VO2 ratio time sequence is calculated; when the VE / VO2 ratio suddenly rises from the stable period (fluctuation less than 5%) and the rising amplitude is ≥20% within 5 seconds, the corresponding time point is the VT time, that is, it marks the entry into the anaerobic metabolism stage.

[0122] Based on the determined VT time, the duration of the metabolic stage is quantified. Before VT (aerobic stage), the duration is as follows:

[0123] ;

[0124] In the formula, t0 is the start of exercise, t VT is the ventilation threshold time. After VT (anaerobic stage), the duration is as follows:

[0125] ;

[0126] In the formula, is the ventilation threshold time. Then the total duration is defined as follows:

[0127] ;

[0128] If the measured , it is marked as not reaching effective load and needs to be measured again. The double weight system is constructed. The duration ratios before and after VT are defined as , , as follows:

[0129] ;

[0130] ;

[0131] Based on the verification, that is, the prediction force in the high load stage is 1.5 times that of the low load, the weight coefficient k=1.5 is set, and the final OUES is the weighted fusion of the segmented slope, defined as is the efficient conversion slope of VE to VO2 under aerobic metabolism, is the efficiency decay slope under anaerobic metabolism. The specific fusion formula is shown in the following formula:

[0132] ;

[0133] For the nonlinear characteristics of the recovery period of cardiopulmonary regulation, the kinetic analysis makes up for the static limitations of OUES, solves the problems of difficult quantification of nonlinear coupling disorder, unexplored correlation of multi-modal signals, and missed detection of subclinical risks, and realizes accurate analysis of dynamic stability and early risk capture of the whole cycle. The recovery period data is defined and preprocessed. First, the time period is defined, and the peak value of exercise is automatically identified, that is, , the maximum value point of VO2, and the end point of the recovery period, that is, , VE≤1.2×resting VE and lasts for 3 minutes. The specific method is as follows: the peak value of exercise ( ) refers to the time point when the oxygen uptake (VO2) reaches the maximum value during exercise, and the automatic identification process is as follows: first, the VO2 time series data (the sampling frequency is usually 1 / s) of the exercise stage is smoothed by a 3-second sliding window to eliminate short-term measurement noise, and the smoothed VO2 sequence is obtained; then in the exercise load stage, that is, from the start of exercise to the time when the subject stops exercise , the local maximum point of is screened, and the VO2 values before and after the point are required to be lower than the point value to exclude false peaks; then the time point corresponding to the maximum value is selected from the candidate peak value as , if there are multiple equal maximum values (such as a plateau), the starting time of the plateau is taken; finally, verify , if it exceeds, it is determined that the exercise does not reach the effective intensity, and needs to be retested.

[0134] The end point of the recovery period ( ) refers to the time point when the cardiopulmonary function basically recovers to the resting level, and the automatic identification process is as follows: first, determine the resting VE reference value , that is, select the ventilation (VE) data of the subject in the resting state before exercise for at least 5 minutes without obvious activity, calculate the 5-minute average value, if there is fluctuation in the resting stage, the abnormal value exceeding needs to be removed and recalculated; then take the exercise stop time as the starting point of the recovery period ( ); then monitor the VE dynamic change from , when , that is, the ventilation decreases to within 1.2 times of the resting level, and this state lasts for at least 3 minutes, mark this time as the preliminary target time ; finally, define as seconds (i.e. the end of the sustained target attainment), if VE cannot meet the target within 1 hour in the recovery period , then the 60 minutes after the exercise is stopped is taken as , i.e. the default maximum recovery period in the clinic.

[0135] Based on the above steps, the time series of VO2, VE, and O2-pulse (VO2 / HR) are extracted. The signals are then purified by wavelet soft threshold denoising, db4 wavelet, and threshold set , N is the sequence length, and the sensor noise can be suppressed. Z-score standardization is then performed, as shown in the following formula:

[0136] ;

[0137] In the formula, represents the value obtained after standardizing the time series of VO2, VE, and O2-pulse, is the mean of the recovery period sequence, is the standard deviation of the recovery period sequence. The influence of the dimension can be eliminated.

[0138] The parameters of phase space reconstruction (PSR) are optimized, i.e. the time delay τ is first determined by calculating the mutual information between the VO2 sequence in the recovery phase (S4, Power=0) and its time delay sequence. When the mutual information first drops to 80% of the minimum value, the corresponding τ is the optimal time delay; then the embedding dimension is determined by calculating the correlation integral as the embedding dimension and radius change. When , is the optimal embedding dimension; finally, the VO2 and VE sequences in the recovery phase (S4) are constructed into a high-dimensional trajectory matrix according to τ and , where N is the length of the VO2 / VE sequence in the recovery phase.

[0139] The kinetic invariants are extracted. First, the maximum Lyapunov exponent is extracted. In the VO2-VE combined phase space, the separation rate of the reference point and the nearest neighbor point is calculated as shown in the following formula:

[0140] ;

[0141] In the formula, is the distance between the neighbor points, and T is the total time. If , it indicates that there is a chaotic characteristic in the recovery process, indicating poor stability.

[0142] The correlation dimension is then extracted, and is fitted. The slope is the correlation dimension . The larger the , the higher the system complexity and the more disordered the recovery regulation.

[0143] Finally, the cross-section offset rate p is extracted, and p is selected as the cross-section As the cross-section, the coordinate offset standard deviation of the trajectory intersection is calculated. The larger p is, the weaker the periodicity of the recovery process is, and the higher the disorder degree is. Multimodal phase space fusion is performed. First, dynamic time warping is performed on the VO2-VE trajectory (x1) ) and the O2-pulse trajectory (x2) ). By aligning the time axis through DTW, the individual differences in the length of the recovery period can be solved, and equal-length trajectory matrices X1 and X2 are obtained. .

[0144] Then, the tensor T is reconstructed , where n is the length of the aligned sequence, is the embedding dimension, and VV and OP corresponding to channel 2 are extracted. Through Tucker decomposition, the core tensor K and the factor matrices A, B, and C are extracted. Finally, the is calculated, as shown in the following formula:

[0145]

[0146] ;

[0147] In the formula, K is the Frobenius norm of the core tensor, the fusion trajectory complexity is associated with the multimodal, , are the correlation dimensions of the two types of trajectories, respectively.

[0148] The basic OUES value, the weighted OUES value, and the recovery period nonlinear enhancement index are obtained. To avoid information fragmentation caused by isolated calculation of multiple source features, the above results are fused by a normalized weighted fusion method to form the OUES comprehensive feature.

[0149] S4, according to the ratio of the slope of the back slope to the slope of the front slope and the oxygen uptake efficiency slope comprehensive feature, an expanded joint feature vector is constructed, the expanded joint feature vector is input into a myocardial infarction risk prediction model to predict the risk of myocardial infarction, and the risk level of the myocardial infarction risk prediction result is divided to obtain the risk level of myocardial infarction.

[0150] Specifically, according to the ratio of the slope of the back slope to the slope of the front slope and the oxygen uptake efficiency slope comprehensive feature, an expanded joint feature vector is constructed, the expanded joint feature vector is input into a myocardial infarction risk prediction model to predict the risk of myocardial infarction, and the risk level of the myocardial infarction risk prediction result is divided to obtain the risk level of myocardial infarction, including:

[0151] ​​​​S41, square expansion is performed on the ratio of the rear slope to the front slope and the oxygen uptake efficiency slope comprehensive feature, and an expanded joint feature vector is constructed based on the square expansion result;

[0152] S42, a multilayer perception model is constructed based on the expanded joint feature vector, the expanded joint feature vector is divided into a training set and a test set, the multilayer perception model is trained using the training set, and a heart attack risk prediction model is obtained.

[0153] Specifically, based on the expanded joint feature vector, a multilayer perception model is constructed, the expanded joint feature vector is divided into a training set and a test set, the multilayer perception model is trained using the training set, and a heart attack risk prediction model is obtained, which includes:

[0154] S421, a multilayer perception model is constructed based on the expanded joint feature vector, and the multilayer perception model is parameterized, to obtain an initial multilayer perception model;

[0155] S422, the expanded joint feature vector is divided into a training set and a test set, the initial multilayer perception model is trained using the training set, and a preliminary heart attack risk prediction result is obtained;

[0156] S423, the prediction error is calculated using the cross-entropy loss function according to the preliminary heart attack risk prediction result, and the initial multilayer perception model is optimized using back propagation according to the error calculation result, to obtain a heart attack risk prediction model.

[0157] S43, the test set is predicted for heart attack risk using the heart attack risk prediction model, and the heart attack risk prediction result is classified into a risk level according to a preset risk level classification rule, to obtain a heart attack risk level.

[0158] Specifically, as shown in Figure 4 The joint feature vector is constructed using the obtained ratio of the rear slope to the front slope and the obtained OUES comprehensive feature, and is nonlinearly enhanced through square expansion, to finally form a four-dimensional expanded feature vector. The feature vector is used as input to build a multilayer perception (MLP) model, the model includes an input layer, a hidden layer and an output layer, the hidden layer uses a ReLU activation function to enhance the nonlinear expression ability, and the output layer uses a Sigmoid activation function to output a heart attack risk probability.

[0159] On the basis of the constructed MLP model, the joint feature vector of the input training sample is inputted, and the prediction performance is improved by optimizing the parameters. After the model training is completed, the feature vector of the test sample is inputted, and the corresponding risk probability value of each subject is outputted. According to the risk probability of myocardial infarction outputted by the model, the subjects are divided into risk levels, that is, when the probability is lower than 30%, it is determined as a low-risk group, when the probability is higher than 80%, it is determined as a high-risk individual, and when the probability is between 30% and 80%, it is determined as a medium-risk group. For the medium-risk group, it is recommended to combine other clinical indicators for comprehensive evaluation to assist doctors to realize more accurate early warning and individualized intervention decision-making.

[0160] Based on the obtained ratio of the post-slope to the pre-slope and the formed OUES comprehensive feature, square expansion is performed and a joint feature vector is constructed. The feature vector not only contains the original dynamic adjustment capability index and the comprehensive efficiency index, but also introduces the square term to explicitly depict the nonlinear relationship, thereby enhancing the sensitivity to the marginal change of the risk level while maintaining the biological interpretability. The finally formed four-dimensional expanded feature vector is used as an input variable for subsequent modeling. On the basis of the constructed feature vector, an MLP model is designed and built. The model includes an input layer, three hidden layers and an output layer. The hidden layer uses the ReLU activation function to enhance the nonlinear mapping ability of the model, and the output layer uses the Sigmoid function to output the risk probability value between 0 and 1. Based on the constructed MLP model, the model parameters are initialized and the loss function and optimization algorithm are defined, and the configuration of the MLP model is completed, so that it has the function of accepting feature input and outputting myocardial infarction risk discrimination results.

[0161] First, the ratio of the post-slope to the pre-slope and the OUES comprehensive feature are square expanded, and are combined with the original features to form a four-dimensional feature vector. Among them, the square term is used to explicitly introduce a nonlinear relationship to enhance the sensitivity of the model to the abnormal gradient of cardiopulmonary function, while maintaining the biological interpretability of the original features.

[0162] On the basis of the constructed expanded feature vector, a multilayer perceptron (MLP) model is designed and built. The model includes an input layer of 4 neurons corresponding to the Slope Ratio, the OUES comprehensive feature, the Slope Ratio square, and the OUES comprehensive feature square, respectively, corresponding to the four-dimensional expanded feature vector, three hidden layers, and an output layer. The hidden layer uses the ReLU activation function to enhance the nonlinear mapping ability, and the output layer uses the Sigmoid activation function to output the myocardial infarction risk probability value between 0 and 1.

[0163] Input layer to the first hidden layer is:

[0164] ;

[0165] From the first hidden layer to the second hidden layer for:

[0166] ;

[0167] From the second hidden layer to the third hidden layer for:

[0168] ;

[0169] The third hidden layer to the output layer for:

[0170] ;

[0171] Finally, the predicted probability of the output layer is calculated using the Sigmoid activation function. :

[0172] ;

[0173] In the formula, Indicates the number of input layers. The first neuron is connected to the hidden layer. The weights of each neuron determine the degree to which each input signal affects the output of that neuron. Indicates the first The activation value of each neuron; Indicates the first The bias of each hidden layer neuron is used to adjust the neuron's output, helping the model to better fit the data and avoid the neuron's output being zero. Indicates the hidden layer number 1 The weighted sum of inputs to each neuron; This represents the linear input to the output layer neurons; This indicates starting from the 3rd hidden layer. The weights of each neuron to the output layer; This represents the model output, indicating the probability of myocardial infarction.

[0174] Based on the constructed MLP model, the model parameters are initialized and the loss function and Adam optimization algorithm are defined to complete the construction and configuration of the MLP model, enabling it to accept feature input and output myocardial infarction risk discrimination results.

[0175] The MLP model parameters are initialized by setting the hidden layer structure to three layers with 240, 150, and 20 neurons respectively. The maximum number of iterations is set to 2000, and an early convergence criterion is implemented: training automatically stops when the cross-entropy loss on the validation set shows no decrease for 100 consecutive iterations (i.e., the change is ≤10⁻⁸), to avoid overfitting. A fixed random seed value of 42 is set to ensure the stability and repeatability of the model training results. The model uses the cross-entropy loss function to measure the difference between the predicted probability and the true label. The loss function is defined as follows:

[0176] ;

[0177] in, This represents the actual probability value. This is the predicted probability value of the model. The model calculates the error term for each neuron in each layer using the backpropagation algorithm and updates the weights layer by layer:

[0178] For the gradient of the first hidden layer, the error term The calculation is as follows:

[0179] ;

[0180] For the gradient of the second hidden layer, the error term The calculation is as follows:

[0181] ;

[0182] For the gradient of the third hidden layer, the error term The calculation is as follows:

[0183] ;

[0184] The loss function weights the input to the output layer. The gradient is:

[0185] ;

[0186] The gradient of the loss function with respect to the weights and biases of each layer is calculated using the chain rule:

[0187] ;

[0188] In the formula, Indicates the first The error term of each neuron in the output layer is used for backpropagation to calculate the weights; This represents the error term of the k-th neuron from the n-th hidden layer to the output layer, used to calculate the error of the previous hidden layer. denotes the weights connecting two neural layers, determining how the error term is passed through; denotes the derivative of the activation function with respect to the weighted input of the th neuron, reflecting the sensitivity of the activation function of this neuron to the change of the input signal; denotes the weighted input of the th neuron in the first hidden layer, which is the sum of the input signal calculated through weights and biases, input to the activation function for nonlinear transformation.

[0189] The Adam optimization algorithm is adopted to improve the training speed and avoid local optimal solution.

[0190] The parameter update formula of Adam is as follows:

[0191] ;

[0192] In the formula, denotes the momentum of the gradient; is a hyperparameter controlling the momentum update, denotes the influence degree of the current time gradient.

[0193] ;

[0194] In the formula, denotes the second moment of the gradient; is a hyperparameter controlling the second moment update, denotes the influence degree of the past gradient square; denotes the second moment of the gradient at the last time.

[0195] ;

[0196] In the formula, denotes the updated parameter; denotes the parameter before updating; is the learning rate, controlling the step size of each update; is the momentum of the gradient after bias correction; is the second moment of the gradient after bias correction; is a small constant added to avoid division by zero, ensuring that each parameter is adaptively adjusted according to its gradient history, thereby improving the training efficiency of the model.

[0197] On the basis of the constructed MLP model, the joint feature vector of the input training sample is inputted, and the prediction performance is improved by optimizing the parameters. After the model training is completed, the feature vector of the test sample is inputted, and the risk probability value corresponding to each subject is outputted. According to the heart attack probability outputted by the model, the heart attack risk is stratified and judged. When the prediction probability is lower than 30%, it indicates that the subject has a low risk of heart attack, and such individuals can be judged as having no obvious heart attack risk and no further intervention is needed. On the contrary, when the prediction probability is higher than 80%, the model judges that there is a high risk of heart attack.

[0198] To sum up, by means of the above technical solutions of the present application, the present application realizes comprehensive representation of the dynamic adjustment ability and metabolic efficiency level of the cardiopulmonary system in the process of exercise load by fusing the ratio of the slope of the post-slope to the slope of the pre-slope and the comprehensive characteristics of the oxygen uptake efficiency slope, combining the multi-layer perception machine model to model the complex nonlinear relationship, thereby improving the accuracy and stability of risk prediction and promoting the development of cardiopulmonary exercise test data towards data-driven intelligence. At the same time, the model output is presented in the form of risk probability, which enhances the data interpretability and practical application value of the results. The problems of difficulty in nonlinear feature modeling and insufficient accuracy of traditional methods in heart attack risk prediction are avoided, and precise modeling and intelligent discrimination of individual heart attack risk are realized; the present application quantifies the dynamic adjustment ability of the cardiopulmonary system in the process of load and recovery by the expanded joint feature vector. At the same time, the oxygen uptake efficiency slope is enhanced in multiple levels, the correlation is improved by weighted evaluation in the metabolic stage, the nonlinear features are extracted by combining the recovery period dynamics analysis, and finally the comprehensive OUES index is formed by fusion, realizing intelligent prediction and grading of heart attack risk, and providing efficient and accurate decision support for heart attack risk assessment.

[0199] The above only describes the preferred embodiments of the present application and should not be used to limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application should be included in the protection scope of the present application.

Claims

1. An intelligent assessment method of myocardial infarction risk based on cardiopulmonary exercise test data, characterized in that, The method comprises: S1, dividing the acquired cardiopulmonary exercise test data into data stages to obtain cardiopulmonary exercise test stage data, and performing outlier processing on the cardiopulmonary exercise test stage data to obtain preprocessed cardiopulmonary exercise test data; S2, calculating oxygen pulse parameters based on the preprocessed cardiopulmonary exercise test data, reconstructing the oxygen pulse parameters using a multi-scale fitting method to obtain reconstructed oxygen pulse parameters, and calculating the ratio of the post-slope to the pre-slope based on the reconstructed oxygen pulse parameters; S3, constructing a regression fitting equation using a logarithmic regression model based on the preprocessed cardiopulmonary exercise test data, extracting the slope coefficient of the regression fitting equation as the oxygen uptake efficiency slope, performing multi-level enhancement processing on the oxygen uptake efficiency slope to obtain oxygen uptake efficiency slope comprehensive features; S4, constructing an extended joint feature vector based on the ratio of the post-slope to the pre-slope and the oxygen uptake efficiency slope comprehensive features, inputting the extended joint feature vector into a myocardial infarction risk prediction model to predict the myocardial infarction risk, and dividing the myocardial infarction risk prediction results into risk levels to obtain the myocardial infarction risk level.

2. The intelligent assessment of heart attack risk based on cardiopulmonary exercise test data method according to claim 1, characterized in that, The method comprises: S11, acquiring cardiopulmonary exercise test data and performing preliminary outlier processing on the cardiopulmonary exercise test data based on a pre-set physiological parameter interval to obtain initial cardiopulmonary exercise test data; S12, dividing the initial cardiopulmonary exercise test data into data stages using a staged hidden semi-Markov model to obtain cardiopulmonary exercise test stage data; The cardiopulmonary exercise test stage data comprises: resting stage data, incremental load stage data, peak stage data, and recovery stage data; S13, performing outlier processing on the resting stage data, incremental load stage data, peak stage data, and recovery stage data, respectively, and performing quality verification on the outlier processing results to obtain preprocessed cardiopulmonary exercise test stage data.

3. The method of intelligent assessment of heart attack risk based on cardiopulmonary exercise test data as claimed in claim 2 wherein, The method comprises: S121, fitting the duration distribution of the cardiopulmonary exercise test stage based on pre-defined one-way transition and observation sequence, and constructing a staged hidden semi-Markov model; S122, expanding the feature dimension of the observation sequence based on pre-labeled cardiopulmonary exercise test data to obtain an expanded observation sequence, training the staged hidden semi-Markov model based on the expanded observation sequence using an expectation maximization algorithm to obtain a trained staged hidden semi-Markov model; S123, decoding the initial cardiopulmonary exercise test data using an improved Viterbi algorithm to obtain a hidden state sequence, obtaining stage boundaries based on the hidden state sequence, and performing stage reasoning and correction on the initial cardiopulmonary exercise test data based on the stage boundaries to obtain cardiopulmonary exercise test stage data.

4. The intelligent assessment of heart attack risk based on cardiopulmonary exercise test data method of claim 1, wherein, The oxygen pulse parameter is calculated based on the preprocessed cardiopulmonary exercise test data, the oxygen pulse parameter is reconstructed by using a multi-scale fitting method to obtain a reconstructed oxygen pulse parameter, and a ratio of a post-slope to a pre-slope is calculated based on the reconstructed oxygen pulse parameter. S21, the oxygen pulse parameter is calculated based on the preprocessed cardiopulmonary exercise test data, the oxygen pulse parameter is decomposed by using a wavelet function to obtain a stationary sequence, and a dominant trend signal is extracted based on the stationary sequence; S22, the linear fitting model is used for linear fitting of the dominant trend signal, and the simulated annealing algorithm is used for global optimization of the linear fitting result with the minimum regression loss function as the target to obtain a trend function; S23, the oxygen pulse parameter is reconstructed based on the trend function to obtain a reconstructed oxygen pulse parameter, the time range of the exercise stage is defined based on the preprocessed cardiopulmonary exercise test data, and the ratio of the post-slope to the pre-slope is calculated based on the reconstructed oxygen pulse parameter and the time range of the exercise stage.

5. The method of intelligent assessment of heart attack risk based on cardiopulmonary exercise test data as claimed in claim 4, wherein, The oxygen pulse parameter is calculated based on the preprocessed cardiopulmonary exercise test data, the oxygen pulse parameter is decomposed by using a wavelet function to obtain a stationary sequence, and a dominant trend signal is extracted based on the stationary sequence; S211, the oxygen pulse parameter is calculated based on the preprocessed cardiopulmonary exercise test data, the oxygen pulse parameter is decomposed by using a wavelet function to obtain an oxygen pulse approximate component, the stationary sequence is obtained by using an enhanced Dickey-Fuller unit root test to determine the stationary sequence, and a determination parameter is obtained; S212, the determination parameter is subjected to a significance test, and the oxygen pulse approximate component is subjected to sequence screening based on the significance test result to obtain a stationary sequence; S213, the stationary sequence is subjected to frequency spectrum analysis by using a fast Fourier transform, and the dominant trend signal is extracted based on the frequency spectrum analysis result.

6. The intelligent assessment of heart attack risk based on cardiopulmonary exercise test data method of claim 1, wherein, The regression fitting equation is constructed by using a logarithmic regression model based on the preprocessed cardiopulmonary exercise test data, and the slope coefficient of the regression fitting equation is extracted as an oxygen uptake efficiency slope, the oxygen uptake efficiency slope is subjected to multi-level enhancement processing to obtain an oxygen uptake efficiency slope comprehensive feature, and the oxygen uptake efficiency slope comprehensive feature includes: S31, the regression fitting equation is constructed by using a logarithmic regression model based on the preprocessed cardiopulmonary exercise test data, and the slope coefficient of the regression fitting equation is extracted as an oxygen uptake efficiency slope; S32, the ventilation threshold is set by using a ratio inflection point method based on the preprocessed cardiopulmonary exercise test data, the metabolic stage is divided based on the ventilation threshold, and a segmented weight is generated, the oxygen uptake efficiency slope is subjected to weighted fusion processing based on the segmented weight to obtain a weighted oxygen uptake efficiency slope; S33, the preprocessed cardiopulmonary exercise test data is subjected to time sequence extraction based on recovery period dynamics analysis to obtain a nonlinear invariant, and the nonlinear invariant is subjected to multi-modal fusion to obtain a nonlinear enhanced oxygen uptake efficiency slope; S34, the oxygen uptake efficiency slope, the weighted oxygen uptake efficiency slope, and the nonlinear enhanced oxygen uptake efficiency slope are integrated by using a normalized weighted fusion method to obtain an oxygen uptake efficiency slope comprehensive feature.

7. The method of intelligent assessment of heart attack risk based on cardiopulmonary exercise test data as claimed in claim 6 wherein, The oxygen uptake efficiency slope comprehensive feature is obtained by integrating the oxygen uptake efficiency slope, the weighted oxygen uptake efficiency slope and the nonlinear enhanced oxygen uptake efficiency slope using the normalized weighted fusion method, and the oxygen uptake efficiency slope comprehensive feature includes: S331, based on the recovery period kinetics analysis, the preprocessed cardiopulmonary exercise test data is time period defined, and the preprocessed cardiopulmonary exercise test data is time sequence extracted according to the time period definition result, to obtain the nonlinear invariant; S332, the nonlinear invariant is denoised to obtain the denoised nonlinear invariant, and the Z-score method is used to standardize the denoised nonlinear invariant to obtain the standard nonlinear invariant; S333, based on the standard nonlinear invariant and the preprocessed cardiopulmonary exercise test data, the phase space reconstruction parameter is optimized, and the kinetic invariant is extracted according to the optimization result, and the kinetic invariant is fused in multiple modes to obtain the nonlinear enhanced oxygen uptake efficiency slope.

8. The intelligent assessment of heart attack risk based on cardiopulmonary exercise test data method of claim 7, wherein, The phase space reconstruction parameter is optimized based on the standard nonlinear invariant and the preprocessed cardiopulmonary exercise test data, and the kinetic invariant is extracted according to the optimization result, and the kinetic invariant is fused in multiple modes to obtain the nonlinear enhanced oxygen uptake efficiency slope, which includes: S3331, based on the standard nonlinear invariant and the preprocessed cardiopulmonary exercise test data, the time delay and the embedding dimension are determined, the phase space reconstruction parameter is optimized by the time delay and the embedding dimension, and the trajectory optimization result is obtained; S3332, the maximum Lyapunov exponent, the correlation dimension and the cross-section offset rate are extracted according to the trajectory optimization result, and the maximum Lyapunov exponent, the correlation dimension and the cross-section offset rate are integrated to obtain the kinetic invariant; S3333, the trajectory optimization result is dynamically time-warped, and a tensor is constructed according to the dynamic time warping result, and the nonlinear enhanced oxygen uptake efficiency slope is obtained by multi-modal phase space fusion based on the tensor and the kinetic invariant.

9. The intelligent assessment of heart attack risk based on cardiopulmonary exercise test data method of claim 1, wherein, The extended joint feature vector is constructed according to the ratio of the post-slope to the pre-slope slope and the oxygen uptake efficiency slope comprehensive feature, the extended joint feature vector is input into the myocardial infarction risk prediction model to predict the myocardial infarction risk, and the myocardial infarction risk level is obtained by dividing the myocardial infarction risk prediction result into risk levels, which includes: S41, square expansion is performed on the ratio of the post-slope to the pre-slope slope and the oxygen uptake efficiency slope comprehensive feature, and an extended joint feature vector is constructed based on the square expansion result; S42, a multi-layer perception machine model is constructed based on the extended joint feature vector, the extended joint feature vector is divided into a training set and a test set, the multi-layer perception machine model is trained using the training set, and a myocardial infarction risk prediction model is obtained; S43, the myocardial infarction risk prediction model is used to predict the myocardial infarction risk of the test set, and the myocardial infarction risk prediction result is divided into risk levels according to the preset risk level division rule to obtain the myocardial infarction risk level.

10. The intelligent assessment of heart attack risk based on cardiopulmonary exercise test data method of claim 9, wherein, The multi-layer perception machine model is constructed based on the extended joint feature vector, the extended joint feature vector is divided into a training set and a test set, the multi-layer perception machine model is trained using the training set, and a myocardial infarction risk prediction model is obtained, which includes: S421、construct a multilayer perceptron model based on the extended joint feature vector, and initialize parameters of the multilayer perceptron model to obtain an initial multilayer perceptron model; S422、divide the extended joint feature vector into a training set and a test set, train the initial multilayer perceptron model using the training set to obtain a preliminary myocardial infarction risk prediction result; S423、calculate a prediction error using a cross-entropy loss function according to the preliminary myocardial infarction risk prediction result, and optimize the initial multilayer perceptron model using back propagation according to the error calculation result to obtain a myocardial infarction risk prediction model.

Images