Processing method and device for predicting ejection fraction retention type heart failure

Abstract

The embodiment of the invention relates to a processing method and device for predicting ejection fraction retention type heart failure. The method comprises the steps that a deep learning model used for conducting dichotomy prediction on the ejection fraction retention type heart failure is constructed and recorded as a first prediction model; constructing a model data set and recording the model data set as a first data set; training the first prediction model based on the first data set; after model training is finished, heart structure feature analysis, heart hemodynamic feature analysis and left ventricle vortex time sequence feature analysis are conducted on four-dimensional blood flow heart magnetic resonance imaging data input by a user to obtain a corresponding structure feature set, and the hemodynamic feature set and the vortex time sequence feature form a corresponding first feature data set; and inputting the first feature data set into a first prediction model for prediction to obtain a corresponding first classification vector which is fed back to the current user. According to the invention, the prediction accuracy and the prediction efficiency can be improved.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and in particular to a processing method and device for predicting heart failure with preserved ejection fraction. Background Art

[0002] Heart failure (HF) is a disease caused by an inability of the heart to pump blood. Heart failure with preserved ejection fraction (HFpEF), also known as HF with preserved ejection fraction, is a type of HF characterized by a normal left ventricular ejection fraction (LVEF) but with the presence of HF symptoms such as dyspnea and fatigue. The diagnostic criteria for HFpEF given by the European Society of Cardiology (ESC) are: 1) the presence of symptoms and / or signs of heart failure; 2) LVEF ≥ 50%; 3) brain natriuretic peptide (BNP) ≥ 35 pg / ml and / or N-terminal pro-brain natriuretic peptide (NT-ProBNP) ≥ 125 pg / ml; and 4) at least one additional criterion, namely related structural heart disease: left atrial enlargement, left ventricular hypertrophy, diastolic dysfunction, etc.

[0003] Under normal circumstances, further confirmation of HFpEF classification in suspected heart failure patients requires a series of tests, such as biomarker testing, chest X-rays, electrocardiograms, and echocardiography. Among them, echocardiography plays a key role in the assessment of cardiac function and LVEF. However, echocardiography also has some shortcomings: 1) The clarity of two-dimensional imaging is limited and is easily affected by factors such as the patient's body shape (such as obesity), chest wall thickness, and emphysema. The accuracy of using analysis tools to directly assess cardiac function based on images is insufficient, and further analysis must be based on the doctor's manual experience; 2) The pixel features of traditional echocardiography do not provide hemodynamic characteristics such as velocity, so the accuracy of measuring quantitative parameters such as blood flow velocity and flow is limited, making it difficult to improve the accuracy of LVEF assessment and meet the needs of hemodynamic characteristic analysis.

[0004] Four-Dimensional Flow Cardiac Magnetic Resonance (4D Flow CMR) is an emerging technology that can comprehensively assess cardiac function and cardiac hemodynamic characteristics. The voxel characteristics of the imaging data acquired through 4D Flow CMR include time and three-dimensional velocity components in addition to signal intensity. Compared to traditional echocardiography, 4D Flow CMR data can obtain high-resolution three-dimensional cardiac anatomy and is not affected by factors such as patient body shape (such as obesity), chest wall thickness, and emphysema. It can also capture three-dimensional hemodynamic information of the heart throughout the cardiac cycle, providing strong support for comprehensive assessment of cardiac function. Therefore, using 4D Flow CMR technology instead of traditional echocardiography in the evaluation process of HFpEF classification can help improve the accuracy of HFpEF classification.

[0005] Currently, there is a series of mature data analysis software for four-dimensional blood flow cardiac magnetic resonance imaging data (such as cvi42, CAAS MR Solutions 5.0, Medis Suite MR, Segment 3.0, FlowTool, VortexCoreTracker, etc.), which can be used to analyze cardiac structural characteristics (such as left atrial maximum volume index, left atrial minimum volume index, left atrial longitudinal peak strain, left ventricular mass index, left ventricular global longitudinal strain, etc.), cardiac hemodynamic characteristics (such as left ventricular filling pressure, left ventricular kinetic energy, left ventricular hemodynamics, left ventricular myocardial work efficiency, left ventricular energy loss, early diastolic mitral blood flow peak velocity to mitral annular motion velocity ratio, tricuspid regurgitation velocity, left ventricular ejection fraction, etc.), and cardiac vortex timing characteristics (such as the timing characteristics of the left ventricular vortex area). Research has found that these cardiac structural characteristics, cardiac hemodynamic characteristics, and cardiac vortex timing characteristics are all associated with heart failure with preserved ejection fraction. If a deep learning model could be used to learn the relationship between these features and heart failure with preserved ejection fraction and perform binary classification predictions based on the learning results, it would not only help improve the accuracy of HFpEF classification assessments, but also increase assessment efficiency. However, how to construct this binary classification prediction model is the technical problem that this invention aims to solve. Summary of the Invention

[0006] The purpose of the present invention is to address the shortcomings of the existing technology and provide a processing method, device, electronic device and computer-readable storage medium for predicting heart failure with preserved ejection fraction. The present invention constructs a first prediction model for binary prediction of heart failure with preserved ejection fraction based on a first feature data set input by the model (consisting of left atrial maximum volume index, left atrial minimum volume index, left atrial longitudinal peak strain, left ventricular mass index, left ventricular global longitudinal strain, left ventricular filling pressure, left ventricular kinetic energy, left ventricular hemodynamics, left ventricular myocardial work efficiency, left ventricular energy loss, early diastolic mitral blood flow peak velocity to mitral annular motion velocity ratio, tricuspid regurgitation velocity, left ventricular ejection fraction, and eddy current timing characteristics); and constructs a model data set by collecting volunteer data to train the first prediction model; and after the model training is completed, the four-dimensional blood flow cardiac magnetic resonance imaging data of the subject is first analyzed for cardiac structural characteristics, cardiac hemodynamic characteristics and left ventricular eddy current timing characteristics to obtain a first feature data set corresponding to the corresponding structural feature group, hemodynamic feature group and eddy current timing feature, and then the first prediction model is used to perform binary prediction of heart failure with preserved ejection fraction based on the first feature data set. On the one hand, the present invention improves the accuracy of original / analyzed data by introducing four-dimensional blood flow cardiac magnetic resonance imaging technology, thereby achieving the purpose of improving evaluation accuracy; on the other hand, it achieves the purpose of improving evaluation / prediction efficiency by introducing deep learning models.

[0007] To achieve the above objectives, a first aspect of an embodiment of the present invention provides a method for predicting heart failure with preserved ejection fraction, the method comprising:

[0008] A deep learning model for binary classification prediction of heart failure with preserved ejection fraction is constructed and recorded as the first prediction model; the first prediction model is used to perform binary classification prediction of heart failure with preserved ejection fraction based on the first feature data set input by the model and output the corresponding first classification vector; the first feature data set includes a structural feature group, a hemodynamic feature group and an eddy current timing feature; the structural feature group includes the left atrial maximum volume index, the left atrial minimum volume index, the left atrial longitudinal peak strain, the left ventricular mass index, and the left ventricular overall longitudinal strain; the hemodynamic feature group includes left ventricular filling pressure, left ventricular kinetic energy, left ventricular hemodynamics, left ventricular myocardial work efficiency, left ventricular energy loss, the ratio of early diastolic mitral blood flow peak velocity to mitral annulus motion velocity, tricuspid regurgitation velocity, and left ventricular ejection fraction; the eddy current timing feature is a left ventricular vortex area A time series sequence of features, wherein the time span of the eddy current time series features covers a complete cardiac cycle; the eddy current time series features are formed by sequentially sorting the left ventricular eddy current zone features at multiple sampling time points, and the time interval between each two adjacent time points is the same; each left ventricular eddy current zone feature is composed of one or more first eddy current zone features, and each first eddy current zone feature is used to reflect the characteristics of an eddy current zone in the left ventricle at the current time point; each first eddy current zone feature includes the center position of the eddy current zone, the maximum blood flow of the eddy current zone, the transverse radius of the eddy current zone, the longitudinal radius of the eddy current zone, and the vortex strength of the eddy current zone; the first classification vector includes a first classification probability ρ1 and a second classification probability ρ2, the first classification probability ρ1 is the predicted probability of heart failure with preserved ejection fraction, and the second classification probability ρ2 is the predicted probability of heart failure without preserved ejection fraction, ρ1+ρ2=1;

[0009] The constructed model data set is recorded as a first data set; and the first prediction model is trained based on the first data set;

[0010] After the model training is completed, the four-dimensional blood flow cardiac magnetic resonance imaging data input by the user is analyzed for cardiac structural characteristics, cardiac hemodynamic characteristics and left ventricular eddy current timing characteristics to obtain the corresponding first feature data set composed of the structural feature group, the hemodynamic feature group and the eddy current timing characteristics; and the first feature data set is input into the first prediction model for prediction to obtain the corresponding first classification vector and feedback it to the current user.

[0011] Preferably, the four-dimensional blood flow cardiac magnetic resonance imaging data is a four-dimensional data object obtained based on four-dimensional blood flow cardiac magnetic resonance imaging technology, and is composed of multiple basic data elements; the data element information of each basic data element includes time information, signal intensity information, three-dimensional voxel coordinate information, three-dimensional blood flow velocity information, and three-dimensional blood flow vorticity interest; the time span of the four-dimensional blood flow cardiac magnetic resonance imaging data is a complete cardiac cycle;

[0012] The first data set includes a plurality of first data records; the first data record includes a first training feature data set and a first label classification vector; the first training feature data set includes the structural feature group, the hemodynamic feature group and the eddy current time series feature; the first label classification vector includes a first label probability and the second label probability The first label probability is the label probability of heart failure with preserved ejection fraction, the second label probability is the label probability of non-ejection fraction preserved heart failure, And one of the two label probabilities is 1 and the other is 0.

[0013] Preferably, the model input end of the first prediction model is used to receive the first feature data set, and the model output end is used to output the first classification vector;

[0014] The first prediction model includes a feature diversion module, a first feature extraction network, a second feature extraction network, a feature fusion module, a low-dimensional feature mapping module and a Softmax function module;

[0015] The input end of the feature diversion module is connected to the input end of the model, and the first and second output ends are connected to the input ends of the first and second feature extraction networks respectively; the output ends of the first and second feature extraction networks are connected to the first and second input ends of the feature fusion module respectively; the output end of the feature fusion module is connected to the input end of the low-dimensional feature mapping module; the output end of the low-dimensional feature mapping module is connected to the input end of the Softmax function module; the output end of the Softmax function module is connected to the output end of the model;

[0016] The feature diversion module is used to combine the thirteen types of feature information of the structural feature group and the hemodynamic feature group of the first feature data set into corresponding feature sequences Send to the first feature extraction network; and use the eddy current time series features of the first feature data set as the corresponding left ventricular eddy current area feature sequence Send to the second feature extraction network; wherein the feature sequence Thirteen characteristics Composition, characteristics is the jth type of feature information among the thirteen types of feature information, 1≤indexj≤13; the left ventricular vortex area feature sequence Characterized by multiple left ventricular vortex zones Arranged in chronological order, the characteristics of the left ventricular vortex area is the left ventricular vortex area feature at the kth sampling time point, 1≤indexk≤N1, N1 is the left ventricular vortex area feature sequence Total number of time points;

[0017] The first feature extraction network is composed of a feature normalization module, a high-dimensional feature mapping module, a nonlinear activation module and a feature alignment module connected in sequence;

[0018] The feature standardization module is used to standardize the feature sequence according to the preset standardization rules. Each of the above features Perform standardization to obtain the corresponding features And all the features obtained The corresponding eigenvector A 1 Send to the high-dimensional feature mapping module; wherein the normalization rule includes at least the z-score normalization rule; the feature vector A 1 The shape is C A1 ×1,C A1 is the feature vector A 1 The feature dimension, C A1 =13;

[0019] The high-dimensional feature mapping module is implemented based on a fully connected network; the high-dimensional feature mapping module is used to map the feature vector A through the fully connected network. 1 Perform high-dimensional feature encoding to obtain the corresponding feature vector A 2 Sending to the nonlinear activation module;

[0020] Among them, A 2 =W1A 1 +b1; W1 and b1 are the weight matrix and offset vector of the fully connected network; the shape of the weight matrix W1 is C A2 ×C A1 , the shape of the offset vector b1 is C A2 ×1, the eigenvector A 2 The shape is C A2 ×1, C A2 is the preset high-dimensional feature dimension, C A2 >C A1 ;

[0021] The nonlinear activation module performs a nonlinear activation on the feature vector A based on a preset first nonlinear activation function. 2 Perform nonlinear activation processing to obtain the corresponding feature vector A 3 Send to the feature alignment module; wherein the first nonlinear activation function includes at least a ReLU activation function; the feature vector A 3 The shape is CA2 ×1;

[0022] The feature alignment module is composed of a fully connected network and a preset second nonlinear activation function; the feature alignment module is used to transform the feature vector A through the fully connected network. 3 The feature dimension of the second feature extraction network is aligned with the output feature dimension of the second feature extraction network to obtain the corresponding feature vector A 4 ; and based on the second nonlinear activation function, the feature vector A 4 Perform nonlinear activation processing to obtain the corresponding feature vector A 5 Sending to the feature fusion module;

[0023] Among them, A 4 =W2A 3 +b2; W2, b2 are the weight matrix and offset vector of the fully connected network; the shape of the weight matrix W2 is C A3 ×C A2 , the shape of the offset vector b1 is C A3 ×1, the eigenvector A 4 、A 5 The shape is C A3 ×1, C A3 is a preset feature dimension, C A3 >C A2 , and the characteristic dimension C A3 The output feature dimension of the second feature extraction network is consistent; the second nonlinear activation function includes at least a ReLU activation function;

[0024] The second feature extraction network is composed of a spatial feature encoding module, a spatiotemporal feature aggregation module, a temporal feature encoding module and a layer normalization module connected in sequence;

[0025] The spatial feature encoding module has a built-in fully connected network and a preset third nonlinear activation function; the spatial feature encoding module is used to convert the left ventricular eddy current area feature sequence Characteristics of each left ventricular vortex area Each of the first eddy current zone characteristics is used as the corresponding eddy current characteristics 1≤indexq(k)≤N k , N k is the kth left ventricular vortex area characteristic The total number of the first eddy current zone features; and each of the eddy current features through a fully connected network Perform feature encoding processing to obtain the corresponding encoding features And based on the third nonlinear activation function, the encoding feature Perform nonlinear activation processing to obtain the corresponding activation features And according to the characteristics of each left ventricular vortex area The corresponding N k The activation feature The corresponding eddy zone feature group And the N1 eddy current zone feature groups are obtained The corresponding eddy zone feature group sequence Sending to the spatiotemporal feature aggregation module;

[0026] in, W3, b3 are the weight matrix and offset vector of the fully connected network; the eddy current characteristics The shape is C e0 ×1, C e0 is the characteristic dimension of the first eddy current zone feature; the shape of the weight matrix W3 is C e1 ×C e0 , the shape of the offset vector b1 is C e1 ×1, the coding feature The activation feature The shape is C e1 ×1, C e1 is the preset spatial encoding feature dimension, C e1 >C e0 ; The third nonlinear activation function includes at least a ReLU activation function;

[0027] The spatiotemporal feature aggregation module is used to aggregate the eddy current zone feature group sequence Each of the eddy current zone characteristic groups N k The activation feature Perform spatial feature attention weighted aggregation processing to obtain the corresponding eddy current aggregation feature And the N1 eddy current aggregation characteristics are obtained The corresponding eddy current aggregation characteristic sequence Sending to the temporal feature encoding module;

[0028] in,

[0029] α q(k) For the activation feature Spatial feature attention weight, 1≤indexq(k) ’ ≤N k , tanh() is the hyperbolic tangent activation function, W4 and V are two preset weight matrices; the shape of the weight matrix W4 is C e1 ×C e1 , the shape of the weight matrix V is 1×Ce1 ,feature The shape is C e1 ×1, The shape of the function output feature is C e1 ×1, the spatial feature attention weight α q(k) is a scalar, the eddy current aggregation characteristic The shape is C e1 ×1;

[0030] The temporal feature encoding module is implemented based on the BiLSTM model; the temporal feature encoding module is used to encode the eddy current aggregation feature sequence Perform bidirectional temporal feature encoding to obtain the corresponding left-to-right encoding feature sequence and right-to-left encoding feature sequences And the N1th encoding features from left to right and the first right-to-left encoding feature Spliced ​​into a corresponding bidirectional encoding feature vector recorded as feature vector D 3 Send to the layer normalization module; wherein the left-to-right encoding feature sequence The features are encoded from left to right by N1 Composition, the encoding feature sequence from right to left The features are encoded from right to left by N1 Composition, the encoding features from left to right and the right-to-left encoding feature The shape is C EN ×1, C EN is the preset one-way coding feature dimension; the feature vector D 3 The shape is C Bi-EN ×1, C Bi-EN is the preset bidirectional encoding feature dimension, C Bi-EN =2×C EN , the bidirectional encoding feature dimension C Bi-EN It is also the output feature dimension of the second feature extraction network;

[0031] The layer normalization module is used to normalize the feature vector D 3 Perform layer normalization to obtain the corresponding feature vector D 4 Send to the feature fusion module; the feature vector D 4 Shape C Bi-EN ×1;

[0032] The feature fusion module is used to 5 and the eigenvector D 4Perform cross-modal correlation matrix calculation to obtain the corresponding correlation matrix R; and use the correlation matrix R and the eigenvector D 4 Multiply to get the corresponding eigenvector D 5 ; and by the feature vector A 5 and the eigenvector D 5 Splice into a corresponding fusion feature vector G 1 To the low-dimensional feature mapping module;

[0033] in,

[0034] softmax() is the Softmax activation function, W5 is a preset weight matrix; the eigenvector A 5 The shape is C A3 ×1, the eigenvector D 4 The shape is C Bi-EN ×1, C A3 =C Bi-EN , the shape of the weight matrix W5 is C A3 ×C A3 =C Bi-EN ×C Bi-EN , feature W5A 5 The shape is C A3 ×1, feature W5D 4 The shape is C Bi-EN ×1=C A3 ×1, feature W5D 4 Transpose (W5D 4 ) T The shape is 1×C A3 , Features (W5A 5 )(W5D 4 ) T The shape is C A3 ×C A3 =C Bi-EN ×C Bi-EN , the shape of the incidence matrix R is C A3 ×C A3 =C Bi-EN ×C Bi-EN ; The feature vector D 5 The shape is C Bi-EN ×1; the fused feature vector G 1 The shape is C G1 ×1, C G1 is the preset fusion feature dimension, C G1 =2×C Bi-EN ;

[0035] The low-dimensional feature mapping module is formed by sequentially connecting a first linear layer, a preset fourth nonlinear activation function, and a second linear layer; the first linear layer is used to transform the fused feature vector G 1 Perform feature dimensionality reduction processing to obtain the corresponding reduced dimensionality feature vector G 2 ; The fourth nonlinear activation function is used to reduce the dimension of the feature vector G 2 Perform nonlinear activation processing to obtain the corresponding activation feature vector G 3 ; The second linear layer is used to activate the feature vector G 3 Perform two-dimensional feature extraction to obtain the corresponding two-dimensional feature vector G 4 Send to the Softmax function module;

[0036] Among them, G 2 =W6G 1 +b4, G 4 =W7G 3 +b5;

[0037] W6 and b4 are the weight matrix and offset vector of the first linear layer, W7 and b5 are the weight matrix and offset vector of the second linear layer; the shape of the weight matrix W6 is C G2 ×C G1 , the shape of the offset vector b4 is C G2 ×1, C G2 is the preset dimension reduction feature dimension, the dimension reduction feature vector G 2 and the activation feature vector G 3 The shape is C G2 ×1; the shape of the weight matrix W7 is C G3 ×C G2 , the shape of the offset vector b5 is C G3 ×1, C G3 is the preset two-dimensional feature dimension, C G3 =2, the two-dimensional feature vector G 4 The shape is C G3 ×1; the fourth nonlinear activation function includes at least a ReLU activation function;

[0038] The Softmax function module is used to calculate the two-dimensional feature vector G 4 Perform binary classification probability calculation to obtain the corresponding first classification probability ρ1 and the second classification probability ρ2; and form the corresponding first classification vector from the first classification probability ρ1 and the second classification probability ρ2 and output it.

[0039] Preferably, the model building dataset is recorded as the first dataset, which specifically includes:

[0040] Recruiting multiple volunteers with heart failure with preserved ejection fraction and multiple volunteers without heart failure with preserved ejection fraction to form corresponding volunteer groups;

[0041] And each of the volunteers in the volunteer set is taken as the corresponding current volunteer; and based on the four-dimensional blood flow cardiac magnetic resonance imaging technology, four-dimensional blood flow cardiac magnetic resonance imaging data of one cardiac cycle is collected for the current volunteer; and the cardiac structural characteristics, cardiac hemodynamic characteristics and left ventricular eddy current timing characteristics of the four-dimensional blood flow cardiac magnetic resonance imaging data collected this time are analyzed to obtain the corresponding structural feature group, the hemodynamic feature group and the eddy current timing feature to form a corresponding first training feature data set; and whether the current volunteer suffers from heart failure with preserved ejection fraction is identified; if so, the corresponding first label probability is set =1, the second label probability If it is 0, then set the corresponding first label probability is 0, the second label probability is 1; and the first label probability obtained at that time and the second label probability Forming a corresponding first label classification vector; and forming a corresponding first data record by the first training feature data set corresponding to the current volunteer and the first label classification vector;

[0042] All the obtained first data records form the corresponding first data set.

[0043] Preferably, the training of the first prediction model based on the first data set specifically includes:

[0044] Step 51: randomly split the first data set into two sub-data sets based on a preset first split ratio as a corresponding first training set and a first evaluation set;

[0045] Wherein, both the first training set and the first evaluation set are composed of a plurality of the first data records; the ratio of the total number of records in the first training set and the first evaluation set satisfies the first segmentation ratio;

[0046] Step 52: Perform a round of traversal on all the first data records in the first training set; during this round of traversal, use the currently traversed first data record as the corresponding current training record; input the first training feature data set of the current training record as the corresponding first feature data set into the first prediction model for prediction, and use the first classification vector output by this prediction as the corresponding first prediction classification vector; and form a corresponding first prediction-label pair from the first prediction classification vector and the first label classification vector of the current training record; and at the end of this round of traversal, bring all the obtained first prediction-label pairs into the preset first model loss function to calculate the corresponding first loss value;

[0047] Wherein, the first model loss function is implemented based on a binary cross entropy loss function, a weighted binary cross entropy loss function or a focal loss function;

[0048] Step 53: Identify whether the first loss value satisfies a preset first loss value range. If so, proceed to step 54. If not, perform a round of modulation on the model parameters of the first prediction model in a direction that minimizes the first model loss function based on a preset first model optimizer, and return to step 52 at the end of this round of modulation.

[0049] Wherein, the first model optimizer includes at least an Adam optimizer and an SFD optimizer;

[0050] Step 54: Perform a round of traversal on all the first data records of the first evaluation set; and during this round of traversal, use the first data record currently traversed as the corresponding current evaluation record; and input the first training feature data set of the current evaluation record as the corresponding first feature data set into the first prediction model for prediction, and use the first classification vector output by this prediction as the corresponding second prediction classification vector; and form a corresponding second prediction-label pair from the second prediction classification vector and the first label classification vector of the current evaluation record; and at the end of this round of traversal, bring all the obtained second prediction-label pairs into the preset first model evaluation function to calculate and obtain the corresponding first evaluation value;

[0051] Wherein, the first model evaluation function is implemented based on MAE function, MSE function or RMSE function;

[0052] Step 55 , identifying whether the first evaluation value satisfies a preset first evaluation value range; if not, returning to step 52 ; if so, confirming that the model training is completed.

[0053] A second aspect of an embodiment of the present invention provides a device for implementing the processing method for predicting heart failure with preserved ejection fraction described in the first aspect, the device comprising: a model building module, a model training module, and a model application module;

[0054] The model building module is used to construct a deep learning model for binary prediction of heart failure with preserved ejection fraction, which is recorded as the first prediction model; the first prediction model is used to perform binary prediction of heart failure with preserved ejection fraction based on the first feature data set input by the model and output the corresponding first classification vector; the first feature data set includes a structural feature group, a hemodynamic feature group and an eddy current timing feature; the structural feature group includes the left atrial maximum volume index, the left atrial minimum volume index, the left atrial longitudinal peak strain, the left ventricular mass index, and the left ventricular overall longitudinal strain; the hemodynamic feature group includes left ventricular filling pressure, left ventricular kinetic energy, left ventricular hemodynamics, left ventricular myocardial work efficiency, left ventricular energy loss, the ratio of early diastolic mitral blood flow peak velocity to mitral annulus motion velocity, tricuspid regurgitation velocity, and left ventricular ejection fraction; the eddy current timing feature is a left A time series sequence of ventricular vortex zone features, wherein the time span of the vortex time series features covers a complete cardiac cycle; the vortex time series features are formed by sequentially sorting the left ventricular vortex zone features at multiple sampling time points, and the time interval between each two adjacent time points is the same; each left ventricular vortex zone feature is composed of one or more first vortex zone features, and each first vortex zone feature is used to reflect the characteristics of a vortex zone in the left ventricle at the current time point; each first vortex zone feature includes the center position of the vortex zone, the maximum blood flow of the vortex zone, the transverse radius of the vortex zone, the longitudinal radius of the vortex zone, and the vortex strength of the vortex zone; the first classification vector includes a first classification probability ρ1 and a second classification probability ρ2, the first classification probability ρ1 is the predicted probability of heart failure with preserved ejection fraction, and the second classification probability ρ2 is the predicted probability of heart failure without preserved ejection fraction, ρ1+ρ2=1;

[0055] The model training module is used to construct a model data set recorded as a first data set; and train the first prediction model based on the first data set;

[0056] The model application module is used to analyze the cardiac structural characteristics, cardiac hemodynamic characteristics and left ventricular eddy current timing characteristics of the four-dimensional blood flow cardiac magnetic resonance imaging data input by the user after the model training is completed, to obtain the corresponding first feature data set composed of the structural feature group, the hemodynamic feature group and the eddy current timing feature; and to input the first feature data set into the first prediction model for prediction to obtain the corresponding first classification vector and feedback it to the current user.

[0057] A third aspect of an embodiment of the present invention provides an electronic device, including: a memory, a processor, and a transceiver;

[0058] The processor is configured to be coupled to the memory, read and execute instructions in the memory, so as to implement the method steps described in the first aspect above;

[0059] The transceiver is coupled to the processor, and the processor controls the transceiver to send and receive messages.

[0060] A fourth aspect of an embodiment of the present invention provides a computer-readable storage medium, which stores computer instructions. When the computer instructions are executed by a computer, the computer executes the instructions of the method described in the first aspect above.

[0061] Embodiments of the present invention provide a processing method, device, electronic device, and computer-readable storage medium for predicting heart failure with preserved ejection fraction. From the above content, it can be seen that the embodiment of the present invention constructs a first prediction model for binary prediction of heart failure with preserved ejection fraction based on a first feature data set input by the model (consisting of left atrial maximum volume index, left atrial minimum volume index, left atrial longitudinal peak strain, left ventricular mass index, left ventricular global longitudinal strain, left ventricular filling pressure, left ventricular kinetic energy, left ventricular hemodynamics, left ventricular myocardial work efficiency, left ventricular energy loss, early diastolic mitral blood flow peak velocity to mitral annular motion velocity ratio, tricuspid regurgitation velocity, left ventricular ejection fraction, and eddy current timing characteristics); and constructs a model data set to train the first prediction model through volunteer data collection; and after the model training is completed, the four-dimensional blood flow cardiac magnetic resonance imaging data of the subject is first analyzed for cardiac structural characteristics, cardiac hemodynamic characteristics and left ventricular eddy current timing characteristics to obtain a first feature data set corresponding to the corresponding structural feature group, hemodynamic feature group and eddy current timing feature composition, and then the first prediction model is used to perform binary prediction of heart failure with preserved ejection fraction based on the first feature data set. The embodiments of the present invention improve the accuracy of original / analyzed data and the accuracy of evaluation by introducing four-dimensional blood flow cardiac magnetic resonance imaging data; and improve the evaluation / prediction efficiency and the generalization performance of evaluation / prediction by introducing a deep learning model. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 A schematic diagram of a method for predicting heart failure with preserved ejection fraction provided in Example 1 of the present invention;

[0063] Figure 2 A schematic diagram of the modules of the first prediction model provided in the first embodiment of the present invention;

[0064] Figure 3This is a module structure diagram of a processing device for predicting heart failure with preserved ejection fraction provided in the second embodiment of the present invention;

[0065] Figure 4 This is a structural diagram of an electronic device provided in Example 3 of the present invention. DETAILED DESCRIPTION

[0066] To make the objectives, technical solutions, and advantages of the present invention more apparent, the present invention will be further described in detail below with reference to the accompanying drawings. It should be understood that the embodiments described herein are merely some, rather than all, of the present invention. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present invention without inventive effort are intended to fall within the scope of protection of the present invention.

[0067] The first embodiment of the present invention provides a method for predicting heart failure with preserved ejection fraction, such as Figure 1 A schematic diagram of a method for predicting heart failure with preserved ejection fraction provided in Example 1 of the present invention is shown. The method mainly includes the following steps:

[0068] Step 1: Build a deep learning model for binary classification prediction of heart failure with preserved ejection fraction, which is recorded as the first prediction model.

[0069] Here, the first prediction model of the embodiment of the present invention is used to perform binary classification prediction of heart failure with preserved ejection fraction based on the first feature data set input by the model and output a corresponding first classification vector.

[0070] The first feature dataset includes a structural feature group, a hemodynamic feature group, and an eddy current time series feature. 1) The structural feature group includes the Left Atrial Volume Index (LAVI) max ), Left Atrial Volume Index (min), LAVI min), Peak Atrial Longitudinal Strain (PALS), Left Ventricular Mass Index (LVMI), and Left Ventricular Global Longitudinal Strain (LVGLS). 2) Hemodynamic characteristics group includes left ventricular filling pressure (LVFP), left ventricular kinetic energy (LVKE), left ventricular hemodynamic force (LVHDF), left ventricular myocardial work efficiency (LVMWE), left ventricular energy loss (LVEL), ratio of early diastolic mitral blood flow peak velocity to mitral annular motion velocity E / e' (E is the early diastolic mitral blood flow peak velocity, e' is the early diastolic mitral annular motion velocity), tricuspid regurgitation velocity (TRV), left ventricular ejection fraction (LVEF), left ventricular myocardial work efficiency (LVMWE), left ventricular energy loss (LVEL), early diastolic mitral blood flow peak velocity to mitral annular motion velocity ratio E / e' (E is the early diastolic mitral blood flow peak velocity, e' is the early diastolic mitral annular motion velocity), tricuspid regurgitation velocity (TRV), left ventricular ejection fraction (LVEF), left ventricular myocardial work efficiency (LVMWE), left ventricular energy loss (LVEL), left ventricular energy loss (LVEL), left ventricular energy loss (LVEF), left ventricular myocardial work efficiency (LVMWE), left ventricular energy loss (LVEL), left ventricular myocardial work efficiency (LVMWE ... myocardial work efficiency ( Ejection Fraction (LVEF). 3) The vortex time series feature is a time series sequence of left ventricular vortex zone features, and the time span of the vortex time series feature covers a complete cardiac cycle. The vortex time series feature is specifically formed by sequentially sorting the left ventricular vortex zone features at multiple sampling time points, with the time interval between every two adjacent time points being the same. Each left ventricular vortex zone feature is composed of one or more first vortex zone features. Each first vortex zone feature is used to reflect the characteristics of a vortex zone in the left ventricle at the current time point. The characteristic information of each first vortex zone feature includes the center position of the vortex zone, the maximum blood flow in the vortex zone, the transverse radius of the vortex zone, the longitudinal radius of the vortex zone, and the vortex strength of the vortex zone.

[0071] The first classification vector includes a first classification probability ρ1 and a second classification probability ρ2; wherein the first classification probability ρ1 is the predicted probability of heart failure with preserved ejection fraction, the second classification probability ρ2 is the predicted probability of heart failure without preserved ejection fraction, and ρ1+ρ2=1.

[0072] The model input end of the first prediction model of the embodiment of the present invention is used to receive the first feature data set, and the model output end is used to output the first classification vector; Figure 2This is a module diagram of the first prediction model provided in Example 1 of the present invention.

[0073] The model components of the first prediction model include: a feature diversion module, a first-path feature extraction network, a second-path feature extraction network, a feature fusion module, a low-dimensional feature mapping module and a Softmax function module.

[0074] The connection relationship of the model components of the first prediction model is: the input end of the feature diversion module is connected to the model input end, and the first and second output ends are respectively connected to the input ends of the first and second feature extraction networks; the output ends of the first and second feature extraction networks are respectively connected to the first and second input ends of the feature fusion module; the output end of the feature fusion module is connected to the input end of the low-dimensional feature mapping module; the output end of the low-dimensional feature mapping module is connected to the input end of the Softmax function module; and the output end of the Softmax function module is connected to the model output end.

[0075] The model component functions of the first prediction model are shown below.

[0076] 1) Feature diversion module:

[0077] The feature splitting module of the embodiment of the present invention is used to combine the thirteen types of feature information of the structural feature group and the hemodynamic feature group of the first feature data set into corresponding feature sequences Send to the first feature extraction network; and use the eddy current time series features of the first feature data set as the corresponding left ventricular eddy current area feature sequence Send to the second feature extraction network;

[0078] Among them, the feature sequence Thirteen characteristics Composition, characteristics is the jth type of feature information among the thirteen types of feature information, 1≤indexj≤13; left ventricular vortex area feature sequence Characterized by multiple left ventricular vortex zones Arranged in chronological order, the characteristics of the left ventricular vortex area is the left ventricular vortex area feature at the kth sampling time point, 1≤indexk≤N1, N1 is the left ventricular vortex area feature sequence The total number of time points.

[0079] 2) The first feature extraction network:

[0080] The first feature extraction network of the embodiment of the present invention is composed of a feature normalization module, a high-dimensional feature mapping module, a nonlinear activation module and a feature alignment module connected in sequence; Figure 2 As shown. Among them:

[0081] a. Feature standardization module is used to standardize feature sequences according to preset standardization rules. The various characteristics of Perform standardization to obtain the corresponding features And all the features obtained The corresponding eigenvector A 1 Send to the high-dimensional feature mapping module;

[0082] Here, the normalization rule of the embodiment of the present invention includes at least the z-score normalization rule; the feature vector A 1 The shape is C A1 ×1,C A1 is the eigenvector A 1 The feature dimension, C A1 =13;

[0083] b. The high-dimensional feature mapping module is implemented based on a fully connected network; the high-dimensional feature mapping module is used to map the feature vector A through the fully connected network. 1 Perform high-dimensional feature encoding to obtain the corresponding feature vector A 2 Send to the nonlinear activation module;

[0084] Among them, A 2 =W1A 1 +b1;

[0085] W1 and b1 are the weight matrix and bias vector of the fully connected network; the shape of the weight matrix W1 is C A2 ×C A1 , the shape of the offset vector b1 is C A2 ×1, eigenvector A 2 The shape is C A2 ×1, C A2 is the preset high-dimensional feature dimension, C A2 >C A1 ;

[0086] c. The nonlinear activation module performs the first nonlinear activation function on the feature vector A. 2 Perform nonlinear activation processing to obtain the corresponding feature vector A 3 Send to the feature alignment module;

[0087] The first nonlinear activation function at least includes the ReLU activation function; the feature vector A 3 The shape is C A2 ×1;

[0088] d. The feature alignment module consists of a fully connected network and a preset second nonlinear activation function; the feature alignment module is used to transform the feature vector A through the fully connected network. 3The feature dimension of the second feature extraction network is aligned with the output feature dimension to obtain the corresponding feature vector A 4 ; And based on the second nonlinear activation function, the feature vector A 4 Perform nonlinear activation processing to obtain the corresponding feature vector A 5 Send to the feature fusion module;

[0089] Among them, A 4 =W2A 3 +b2;

[0090] W2 and b2 are the weight matrix and offset vector of the fully connected network; the shape of the weight matrix W2 is C A3 ×C A2 , the shape of the offset vector b1 is C A3 ×1, eigenvector A 4 、A 5 The shape is C A3 ×1, C A3 is a preset feature dimension, C A3 >C A2 , and the feature dimension C A3 The output feature dimension is consistent with that of the second feature extraction network; the second nonlinear activation function includes at least a ReLU activation function.

[0091] 3) Second feature extraction network:

[0092] The second feature extraction network of the embodiment of the present invention is composed of a spatial feature encoding module, a spatiotemporal feature aggregation module, a temporal feature encoding module, and a layer normalization module connected in sequence.

[0093] a. The spatial feature encoding module has a built-in fully connected network and a preset third nonlinear activation function; the spatial feature encoding module is used to convert the left ventricular eddy current area feature sequence Characteristics of each left ventricular vortex area The characteristics of each first eddy zone are used as the corresponding eddy characteristics 1≤indexq(k)≤N k , N k is the characteristic of the kth left ventricular vortex area The total number of the first eddy zone features; and the eddy characteristics of each eddy Perform feature encoding processing to obtain the corresponding encoding features And based on the third nonlinear activation function, the encoding features Perform nonlinear activation processing to obtain the corresponding activation features The characteristics of each left ventricular vortex area The corresponding N k Active Features The corresponding eddy zone feature group And the N1 eddy current zone feature groups are obtained The corresponding eddy zone feature group sequence Send to the spatiotemporal feature aggregation module;

[0094] in,

[0095] W3 and b3 are the weight matrix and offset vector of the fully connected network; eddy current characteristics The shape is C e0 ×1, C e0 is the characteristic dimension of the first eddy current zone; the shape of the weight matrix W3 is C e1 ×C e0 , the shape of the offset vector b1 is C e1 ×1, encoding features Activate Features The shape is C e1 ×1, C e1 is the preset spatial encoding feature dimension, C e1 >C e0 ;

[0096] The third nonlinear activation function includes at least a ReLU activation function;

[0097] b. The spatiotemporal feature aggregation module is used to aggregate the feature group sequences of the eddy current area. Each eddy zone characteristic group N k Activate Features Perform spatial feature attention weighted aggregation processing to obtain the corresponding eddy current aggregation feature And the N1 eddy current aggregation characteristics are obtained The corresponding eddy current aggregation characteristic sequence Send to the temporal feature encoding module;

[0098] in,

[0099] α q(k) To activate the feature Spatial feature attention weight, 1≤indexq(k) ’ ≤N k , tanh() is the hyperbolic tangent activation function, W4 and V are two preset weight matrices; the shape of the weight matrix W4 is C e1 ×C e1 , the shape of the weight matrix V is 1×C e1 ,feature The shape is C e1 ×1, The shape of the function output feature is C e1 ×1, spatial feature attention weight α q(k) is a scalar, eddy current aggregation characteristics The shape is C e1 ×1;

[0100] c. The temporal feature encoding module is implemented based on the BiLSTM model; the temporal feature encoding module is used to encode the eddy current aggregation feature sequence. Perform bidirectional temporal feature encoding to obtain the corresponding left-to-right encoding feature sequence and right-to-left encoding feature sequences And the N1th encoding features from left to right and the first right-to-left encoding feature Spliced ​​into a corresponding bidirectional encoding feature vector recorded as feature vector D 3 Send to the layer normalization module;

[0101] Among them, the feature sequence is encoded from left to right N1 features are encoded from left to right Composition, encoding feature sequence from right to left N1 features encoded from right to left Composition, encoding features from left to right and right-to-left encoding features The shape is C EN ×1, C EN is the preset one-way encoding feature dimension; the feature vector D 3 The shape is C Bi -EN ×1, C Bi-EN is the preset bidirectional encoding feature dimension, C Bi-EN =2×C EN , bidirectional encoding feature dimension C Bi -EN is also the output feature dimension of the second feature extraction network;

[0102] d. Layer normalization module is used to normalize the feature vector D 3 Perform layer normalization to obtain the corresponding feature vector D 4 Send to the feature fusion module;

[0103] Among them, the eigenvector D 4 Shape C Bi-EN ×1.

[0104] 4) Feature fusion module:

[0105] The feature fusion module of the embodiment of the present invention is used to 5 and the eigenvector D 4The cross-modal correlation matrix is ​​calculated to obtain the corresponding correlation matrix R; and the correlation matrix R and the eigenvector D 4 Multiply to get the corresponding eigenvector D 5 ; and by the eigenvector A 5 and the eigenvector D 5 Splice into a corresponding fusion feature vector G 1 To low-dimensional feature mapping module;

[0106] in,

[0107] softmax() is the Softmax activation function, W5 is a preset weight matrix; the eigenvector A 5 The shape is C A3 ×1, eigenvector D 4 The shape is C Bi-EN ×1, C A3 =C Bi-EN , the shape of the weight matrix W5 is C A3 ×C A3 =C Bi-EN ×C Bi-EN , feature W5A 5 The shape is C A3 ×1, feature W5D 4 The shape is C Bi-EN ×1=C A3 ×1, feature W5D 4 Transpose (W5D 4 ) T The shape is 1×C A3 , Features (W5A 5 )(W5D 4 ) T The shape is C A3 ×C A3 =C Bi-EN ×C Bi-EN , the shape of the incidence matrix R is C A3 ×C A3 =C Bi-EN ×C Bi-EN ; eigenvector D 5 The shape is C Bi-EN ×1; fused feature vector G 1 The shape is C G1 ×1, C G1 is the preset fusion feature dimension, C G1 =2×C Bi-EN .

[0108] 5) Low-dimensional feature mapping module:

[0109] The low-dimensional feature mapping module of the embodiment of the present invention is composed of a first linear layer, a preset fourth nonlinear activation function and a second linear layer connected in sequence; the first linear layer is used to fusion feature vector G 1 Perform feature dimensionality reduction processing to obtain the corresponding reduced dimensionality feature vector G 2 ; The fourth nonlinear activation function is used to reduce the dimension of the feature vector G 2 Perform nonlinear activation processing to obtain the corresponding activation feature vector G 3 ; The second linear layer is used to activate the feature vector G 3 Perform two-dimensional feature extraction to obtain the corresponding two-dimensional feature vector G 4 Send to the Softmax function module;

[0110] Among them, G 2 =W6G 1 +b4, G 4 =W7G 3 +b5;

[0111] W6 and b4 are the weight matrix and offset vector of the first linear layer, and W7 and b5 are the weight matrix and offset vector of the second linear layer. The shape of the weight matrix W6 is C G2 ×C G1 , the shape of the offset vector b4 is C G2 ×1, C G2 is the preset dimensionality reduction feature dimension, the dimensionality reduction feature vector G 2 and activation feature vector G 3 The shape is C G2 ×1; the shape of the weight matrix W7 is C G3 ×C G2 , the shape of the offset vector b5 is C G3 ×1, C G3 is the preset two-dimensional feature dimension, C G3 =2, two-dimensional feature vector G 4 The shape is C G3 ×1;

[0112] The fourth nonlinear activation function includes at least a ReLU activation function.

[0113] 6) Softmax function module:

[0114] The Softmax function module of the embodiment of the present invention is used to calculate the eigenvector G according to the two-dimensional feature vector G. 4 Perform binary classification probability calculation to obtain the corresponding first classification probability ρ1 and second classification probability ρ2; and form the corresponding first classification vector from the first classification probability ρ1 and the second classification probability ρ2 and output it.

[0115] Step 2: construct a model data set and record it as the first data set; and train a first prediction model based on the first data set;

[0116] Specifically comprising: step 21, constructing a model data set and recording it as the first data set;

[0117] The first data set includes a plurality of first data records; the first data record includes a first training feature data set and a first label classification vector; the first training feature data set includes a structural feature group, a hemodynamic feature group and an eddy current time series feature; the first label classification vector includes a first label probability and the second label probability First label probability is the label probability of heart failure with preserved ejection fraction, the second label probability is the label probability of non-ejection fraction preserved heart failure, And one of the two label probabilities is 1 and the other is 0;

[0118] Specifically, the method includes: step 211, recruiting a plurality of volunteers suffering from heart failure with preserved ejection fraction and a plurality of volunteers not suffering from heart failure with preserved ejection fraction to form a corresponding volunteer set;

[0119] Step 212, and each volunteer in the volunteer set is taken as the corresponding current volunteer; and based on the four-dimensional blood flow cardiac magnetic resonance imaging technology, four-dimensional blood flow cardiac magnetic resonance imaging data of one cardiac cycle is collected for the current volunteer; and the cardiac structural characteristics, cardiac hemodynamic characteristics and left ventricular eddy current timing characteristics of the four-dimensional blood flow cardiac magnetic resonance imaging data collected this time are analyzed to obtain a corresponding structural feature group, hemodynamic feature group and eddy current timing feature to form a corresponding first training feature data set; and identify whether the current volunteer suffers from heart failure with preserved ejection fraction; if so, set the corresponding first label probability 1, the second label probability If it is 0, then set the corresponding first label probability 0, the second label probability is 1; and the probability of the first label obtained at that time and the second label probability A corresponding first label classification vector is formed; and a corresponding first data record is formed by the first training feature data set corresponding to the current volunteer and the first label classification vector;

[0120] Here, the four-dimensional blood flow cardiac magnetic resonance imaging data of the embodiment of the present invention is a four-dimensional data object obtained based on the four-dimensional blood flow cardiac magnetic resonance imaging technology, and is composed of multiple basic data elements; the data element information of each basic data element includes time information, signal intensity information, three-dimensional voxel coordinate information, three-dimensional blood flow velocity information, and three-dimensional blood flow vorticity interest; the time span of the four-dimensional blood flow cardiac magnetic resonance imaging data is a complete cardiac cycle;

[0121] It should be noted that, when analyzing cardiac structural features, cardiac hemodynamic features and left ventricular eddy current time series features of four-dimensional blood flow cardiac magnetic resonance imaging data, the embodiment of the present invention can automatically analyze the four-dimensional blood flow cardiac magnetic resonance imaging data analysis interface provided by a series of mature data analysis software, such as cvi42, CAAS MRSolutions 5.0, Medis Suite MR, Segment 3.0, FlowTool, VortexCoreTracker, etc.; specifically, the five types of structural feature information (LAVI) of the embodiment of the present invention are analyzed based on these analysis interfaces. max 、LAVI min , PALS, LVMI, LVGLS), eight types of hemodynamic information (LVFP, LVKE, LVHDF, LVMWE, LVEL, E / e ’ , TRV, LVEF), and five types of eddy current characteristic information (the center position of the eddy current area, the maximum blood flow in the eddy current area, the transverse radius of the eddy current area, the longitudinal radius of the eddy current area, and the eddy current intensity) are automatically analyzed;

[0122] Step 213: All the first data records obtained are combined into a corresponding first data set;

[0123] Step 22: training a first prediction model based on the first data set;

[0124] Specifically comprising: step 221, randomly dividing the first data set into two sub-data sets as a corresponding first training set and a first evaluation set based on a preset first division ratio;

[0125] The first split ratio is a preset ratio parameter, such as 8:2; the first training set and the first evaluation set are both composed of a plurality of first data records; the ratio of the total number of records in the first training set to the total number of records in the first evaluation set satisfies the first split ratio;

[0126] Step 222: Perform a round of traversal on all first data records of the first training set; during this round of traversal, use the currently traversed first data record as the corresponding current training record; input the first training feature data set of the current training record as the corresponding first feature data set into the first prediction model for prediction, and use the first classification vector output by this prediction as the corresponding first prediction classification vector; and form a corresponding first prediction-label pair from the first prediction classification vector and the first label classification vector of the current training record; and at the end of this round of traversal, bring all the obtained first prediction-label pairs into the preset first model loss function to calculate and obtain the corresponding first loss value;

[0127] The first model loss function is implemented based on a binary cross entropy loss function, a weighted binary cross entropy loss function, or a focal loss function;

[0128] Step 223: Identify whether the first loss value satisfies a preset first loss value range. If so, proceed to step 224. If not, perform a round of modulation on the model parameters of the first prediction model in a direction that minimizes the first model loss function based on a preset first model optimizer, and return to step 222 at the end of this round of modulation.

[0129] The first loss value range is a preset value range; the first model optimizer includes at least an Adam optimizer and an SFD optimizer;

[0130] Step 224: Perform a round of traversal on all first data records of the first evaluation set; during this round of traversal, use the currently traversed first data record as the corresponding current evaluation record; input the first training feature data set of the current evaluation record as the corresponding first feature data set into the first prediction model for prediction; and use the first classification vector output by this prediction as the corresponding second prediction classification vector; and form a corresponding second prediction-label pair from the second prediction classification vector and the first label classification vector of the current evaluation record; and at the end of this round of traversal, bring all the obtained second prediction-label pairs into the preset first model evaluation function to calculate and obtain the corresponding first evaluation value;

[0131] Wherein, the first model evaluation function is implemented based on the MAE function, the MSE function or the RMSE function;

[0132] Step 225 , identifying whether the first evaluation value meets the preset first evaluation value range; if not, returning to step 222 ; if so, confirming that the model training is completed.

[0133] Here, the first evaluation value range is a preset numerical range.

[0134] Step 3: After the model training is completed, the four-dimensional blood flow cardiac magnetic resonance imaging data input by the user is analyzed for cardiac structural characteristics, cardiac hemodynamic characteristics and left ventricular eddy current timing characteristics to obtain the corresponding first feature data set consisting of the corresponding structural feature group, hemodynamic feature group and eddy current timing characteristics; and the first feature data set is input into the first prediction model for prediction to obtain the corresponding first classification vector and feedback it to the current user.

[0135] Figure 3 This is a module structure diagram of a processing device for predicting heart failure with preserved ejection fraction provided in the second embodiment of the present invention. The device is a terminal device or server that implements the aforementioned method embodiment, or can be a device that enables the aforementioned terminal device or server to implement the aforementioned method embodiment. For example, the device can be a device or chip system of the aforementioned terminal device or server. Figure 3 As shown, the device includes: a model construction module 201, a model training module 202 and a model application module 203.

[0136] The model building module 201 is used to construct a deep learning model for binary classification prediction of heart failure with preserved ejection fraction, which is recorded as the first prediction model; the first prediction model is used to perform binary classification prediction of heart failure with preserved ejection fraction based on the first feature data set input by the model and output the corresponding first classification vector; the first feature data set includes a structural feature group, a hemodynamic feature group and an eddy current timing feature; the structural feature group includes the maximum volume index of the left atrium, the minimum volume index of the left atrium, the longitudinal peak strain of the left atrium, the left ventricular mass index, and the overall longitudinal strain of the left ventricle; the hemodynamic feature group includes the left ventricular filling pressure, the left ventricular kinetic energy, the left ventricular hemodynamics, the left ventricular myocardial work efficiency, the left ventricular energy loss, the ratio of the early diastolic mitral blood flow peak velocity to the mitral annulus motion velocity, the tricuspid regurgitation velocity, and the left ventricular ejection fraction; the eddy current timing feature is a time series sequence of left ventricular vortex zone features, and the time span of the vortex time series features covers a complete cardiac cycle; the vortex time series features are formed by sequentially sorting the left ventricular vortex zone features at multiple sampling time points, and the time interval between each two adjacent time points is the same; each left ventricular vortex zone feature is composed of one or more first vortex zone features, and each first vortex zone feature is used to reflect the characteristics of a vortex zone in the left ventricle at the current time point; each first vortex zone feature includes the center position of the vortex zone, the maximum blood flow in the vortex zone, the transverse radius of the vortex zone, the longitudinal radius of the vortex zone, and the vortex strength of the vortex zone; the first classification vector includes the first classification probability ρ1 and the second classification probability ρ2, the first classification probability ρ1 is the predicted probability of heart failure with preserved ejection fraction, and the second classification probability ρ2 is the predicted probability of heart failure without preserved ejection fraction, ρ1+ρ2=1.

[0137] The model training module 202 is used to construct a model data set recorded as a first data set; and train a first prediction model based on the first data set.

[0138] The model application module 203 is used to analyze the cardiac structural characteristics, cardiac hemodynamic characteristics and left ventricular eddy current timing characteristics of the four-dimensional blood flow cardiac magnetic resonance imaging data input by the user after the model training is completed, to obtain a corresponding first feature data set consisting of a corresponding structural feature group, a hemodynamic feature group and an eddy current timing feature group; and to input the first feature data set into the first prediction model for prediction to obtain a corresponding first classification vector and feedback it to the current user.

[0139] An embodiment of the present invention provides a processing device for predicting heart failure with preserved ejection fraction, which can execute the method steps in the above method embodiment. Its implementation principles and technical effects are similar and will not be repeated here.

[0140] It should be noted that the division of the modules of the above devices is merely a division of logical functions. In actual implementation, they can be fully or partially integrated into a physical entity or physically separated. Furthermore, these modules can all be implemented in the form of software called by a processing element; or all be implemented in the form of hardware; or some modules can be implemented in the form of software called by a processing element, and some modules can be implemented in the form of hardware. For example, the model building module can be a separate processing element, or it can be integrated into a chip of the above device. In addition, it can be stored in the form of program code in the memory of the above device, and called by a processing element of the above device to perform the functions of the above-mentioned module. The implementation of other modules is similar. In addition, these modules can all or partly be integrated together, or they can be implemented independently. The processing element described here can be an integrated circuit with signal processing capabilities. During implementation, each step of the above method or each of the above modules can be completed by hardware integrated logic circuits in the processor element or instructions in the form of software.

[0141] For example, the above modules may be one or more integrated circuits configured to implement the above methods, such as one or more application-specific integrated circuits (ASICs), one or more digital signal processors (DSPs), or one or more field programmable gate arrays (FPGAs). For another example, when a module is implemented by scheduling program code through a processing element, the processing element may be a general-purpose processor, such as a central processing unit (CPU) or other processor that can call program code. For another example, these modules may be integrated together and implemented in the form of a system-on-a-chip (SOC).

[0142] In the above embodiments, all or part of the embodiments may be implemented using software, hardware, firmware, or any combination thereof. When implemented using software, all or part of the embodiments may be implemented in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the processes or functions described in the above method embodiments are generated in whole or in part. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The above-mentioned computer instructions can be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the above-mentioned computer instructions can be transmitted from one website, computer, server or data center to another website, computer, server or data center via wired (such as coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (such as infrared, wireless, Bluetooth, microwave, etc.) means. The above-mentioned computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server or data center that includes one or more available media. The above-mentioned available medium can be a magnetic medium (such as a floppy disk, hard disk, tape), an optical medium (such as a DVD), or a semiconductor medium (such as a solid state disk (SSD)), etc.

[0143] Figure 4 This is a schematic diagram of the structure of an electronic device provided in the third embodiment of the present invention. The electronic device can be a terminal device or server that implements the method of the aforementioned embodiment, or it can be a terminal device or server that implements the method of the aforementioned embodiment connected to the aforementioned terminal device or server. Figure 4As shown, the electronic device may include: a processor 301 (such as a CPU), a memory 302, and a transceiver 303; the transceiver 303 is coupled to the processor 301, and the processor 301 controls the transceiver 303's transceiver actions. Various instructions may be stored in the memory 302 for completing various processing functions and implementing the processing steps described in the aforementioned embodiment method. Preferably, the electronic device involved in the embodiment of the present invention further includes: a power supply 304, a system bus 305, and a communication port 306. The system bus 305 is used to realize communication connections between components. The above-mentioned communication port 306 is used for connecting and communicating between the electronic device and other peripherals.

[0144] exist Figure 4 The system bus 305 mentioned in the figure can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. The system bus can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, only one thick line is used in the figure, but it does not mean that there is only one bus or one type of bus. The communication interface is used to realize communication between the database access device and other devices (such as clients, read-write libraries, and read-only libraries). The memory may include random access memory (RAM) and may also include non-volatile memory (Non-Volatile Memory), such as at least one disk storage.

[0145] The above-mentioned processors can be general-purpose processors, including central processing units (CPUs), network processors (NPs), graphics processing units (GPUs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0146] It should be noted that an embodiment of the present invention further provides a computer-readable storage medium, which stores instructions. When the computer-readable storage medium is run on a computer, it enables the computer to execute the methods and processing procedures provided in the above embodiments.

[0147] Embodiments of the present invention provide a processing method, device, electronic device, and computer-readable storage medium for predicting heart failure with preserved ejection fraction. From the above content, it can be seen that the embodiment of the present invention constructs a first prediction model for binary prediction of heart failure with preserved ejection fraction based on a first feature data set input by the model (consisting of left atrial maximum volume index, left atrial minimum volume index, left atrial longitudinal peak strain, left ventricular mass index, left ventricular global longitudinal strain, left ventricular filling pressure, left ventricular kinetic energy, left ventricular hemodynamics, left ventricular myocardial work efficiency, left ventricular energy loss, early diastolic mitral blood flow peak velocity to mitral annular motion velocity ratio, tricuspid regurgitation velocity, left ventricular ejection fraction, and eddy current timing characteristics); and constructs a model data set to train the first prediction model through volunteer data collection; and after the model training is completed, the four-dimensional blood flow cardiac magnetic resonance imaging data of the subject is first analyzed for cardiac structural characteristics, cardiac hemodynamic characteristics and left ventricular eddy current timing characteristics to obtain a first feature data set corresponding to the corresponding structural feature group, hemodynamic feature group and eddy current timing feature composition, and then the first prediction model is used to perform binary prediction of heart failure with preserved ejection fraction based on the first feature data set. The embodiments of the present invention improve the accuracy of original / analyzed data and the accuracy of evaluation by introducing four-dimensional blood flow cardiac magnetic resonance imaging data; and improve the evaluation / prediction efficiency and the generalization performance of evaluation / prediction by introducing a deep learning model.

[0148] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein may be implemented using hardware, a software module executed by a processor, or a combination of the two. The software module may be placed in a random access memory (RAM), a memory, a read-only memory (ROM), an electrically programmable ROM, an electrically erasable programmable ROM, a register, a hard disk, a removable disk, a CD-ROM, or any other form of storage medium known in the art.

[0149] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for predicting heart failure with preserved ejection fraction, characterized in that: The method comprises: A deep learning model for binary classification prediction of heart failure with preserved ejection fraction is constructed and recorded as the first prediction model; the first prediction model is used to perform binary classification prediction of heart failure with preserved ejection fraction based on the first feature data set input by the model and output the corresponding first classification vector; the first feature data set includes a structural feature group, a hemodynamic feature group and an eddy current timing feature; the structural feature group includes the left atrial maximum volume index, the left atrial minimum volume index, the left atrial longitudinal peak strain, the left ventricular mass index, and the left ventricular overall longitudinal strain; the hemodynamic feature group includes left ventricular filling pressure, left ventricular kinetic energy, left ventricular hemodynamics, left ventricular myocardial work efficiency, left ventricular energy loss, the ratio of early diastolic mitral blood flow peak velocity to mitral annulus motion velocity, tricuspid regurgitation velocity, and left ventricular ejection fraction; the eddy current timing feature is a left ventricular vortex area A time series sequence of features, wherein the time span of the eddy current time series features covers a complete cardiac cycle; the eddy current time series features are formed by sequentially sorting the left ventricular eddy current zone features at multiple sampling time points, and the time interval between each two adjacent time points is the same; each left ventricular eddy current zone feature is composed of one or more first eddy current zone features, and each first eddy current zone feature is used to reflect the characteristics of an eddy current zone in the left ventricle at the current time point; each first eddy current zone feature includes the center position of the eddy current zone, the maximum blood flow of the eddy current zone, the transverse radius of the eddy current zone, the longitudinal radius of the eddy current zone, and the vortex strength of the eddy current zone; the first classification vector includes a first classification probability ρ1 and a second classification probability ρ2, the first classification probability ρ1 is the predicted probability of heart failure with preserved ejection fraction, and the second classification probability ρ2 is the predicted probability of heart failure without preserved ejection fraction, ρ1+ρ2=1; The constructed model data set is recorded as a first data set; and the first prediction model is trained based on the first data set; After the model training is completed, the four-dimensional blood flow cardiac magnetic resonance imaging data input by the user is analyzed for cardiac structural characteristics, cardiac hemodynamic characteristics and left ventricular eddy current timing characteristics to obtain the corresponding first feature data set composed of the structural feature group, the hemodynamic feature group and the eddy current timing characteristics; and the first feature data set is input into the first prediction model for prediction to obtain the corresponding first classification vector and feedback it to the current user.

2. The method for predicting heart failure with preserved ejection fraction according to claim 1, characterized in that: The four-dimensional blood flow cardiac magnetic resonance imaging data is a four-dimensional data object obtained based on four-dimensional blood flow cardiac magnetic resonance imaging technology, and is composed of multiple basic data elements; the data element information of each basic data element includes time information, signal intensity information, three-dimensional voxel coordinate information, three-dimensional blood flow velocity information and three-dimensional blood flow vorticity interest; the time span of the four-dimensional blood flow cardiac magnetic resonance imaging data is a complete cardiac cycle; The first data set includes a plurality of first data records; the first data record includes a first training feature data set and a first label classification vector; the first training feature data set includes the structural feature group, the hemodynamic feature group and the eddy current time series feature; the first label classification vector includes a first label probability and the second label probability The first label probability is the label probability of heart failure with preserved ejection fraction, the second label probability is the label probability of non-ejection fraction preserved heart failure, And one of the two label probabilities is 1 and the other is 0.

3. The method for predicting heart failure with preserved ejection fraction according to claim 2, wherein: The model input end of the first prediction model is used to receive the first feature data set, and the model output end is used to output the first classification vector; The first prediction model includes a feature diversion module, a first feature extraction network, a second feature extraction network, a feature fusion module, a low-dimensional feature mapping module and a Softmax function module; The input end of the feature diversion module is connected to the input end of the model, and the first and second output ends are connected to the input ends of the first and second feature extraction networks respectively; the output ends of the first and second feature extraction networks are connected to the first and second input ends of the feature fusion module respectively; the output end of the feature fusion module is connected to the input end of the low-dimensional feature mapping module; the output end of the low-dimensional feature mapping module is connected to the input end of the Softmax function module; the output end of the Softmax function module is connected to the output end of the model; The feature diversion module is used to combine the thirteen types of feature information of the structural feature group and the hemodynamic feature group of the first feature data set into corresponding feature sequences Send to the first feature extraction network; and use the eddy current time series features of the first feature data set as the corresponding left ventricular eddy current area feature sequence Send to the second feature extraction network; wherein the feature sequence Thirteen characteristics Composition, characteristics is the jth type of feature information among the thirteen types of feature information, 1≤indexj≤13; the left ventricular vortex area feature sequence Characterized by multiple left ventricular vortex zones Arranged in chronological order, the characteristics of the left ventricular vortex area is the left ventricular vortex area feature at the kth sampling time point, 1≤indexk≤N1, N1 is the left ventricular vortex area feature sequence Total number of time points; The first feature extraction network is composed of a feature normalization module, a high-dimensional feature mapping module, a nonlinear activation module and a feature alignment module connected in sequence; The feature standardization module is used to standardize the feature sequence according to the preset standardization rules. Each of the above features Perform standardization to obtain the corresponding features And all the features obtained The corresponding eigenvector A 1 Send to the high-dimensional feature mapping module; wherein the normalization rule includes at least the z-score normalization rule; the feature vector A 1 The shape is C A1 ×1,C A1 is the feature vector A 1 The feature dimension, C A1 =13; The high-dimensional feature mapping module is implemented based on a fully connected network; the high-dimensional feature mapping module is used to map the feature vector A through the fully connected network. 1 Perform high-dimensional feature encoding to obtain the corresponding feature vector A 2 Sending to the nonlinear activation module; Among them, A 2 =W1A 1 +b1; W1 and b1 are the weight matrix and offset vector of the fully connected network; the shape of the weight matrix W1 is C A2 ×C A1 , the shape of the offset vector b1 is C A2 ×1, the eigenvector A 2 The shape is C A2 ×1, C A2 is the preset high-dimensional feature dimension, C A2 >C A1 ; The nonlinear activation module performs a nonlinear activation on the feature vector A based on a preset first nonlinear activation function. 2 Perform nonlinear activation processing to obtain the corresponding feature vector A 3 Send to the feature alignment module; wherein the first nonlinear activation function includes at least a ReLU activation function; the feature vector A 3 The shape is C A2 ×1; The feature alignment module is composed of a fully connected network and a preset second nonlinear activation function; the feature alignment module is used to transform the feature vector A through the fully connected network. 3 The feature dimension of the second feature extraction network is aligned with the output feature dimension of the second feature extraction network to obtain the corresponding feature vector A 4 ; and based on the second nonlinear activation function, the feature vector A 4 Perform nonlinear activation processing to obtain the corresponding feature vector A 5 Sending to the feature fusion module; Among them, A 4 =W2A 3 +b2; W2 and b2 are the weight matrix and offset vector of the fully connected network; the shape of the weight matrix W2 is C A3 ×C A2 , the shape of the offset vector b1 is C A3 ×1, the eigenvector A 4 、A 5 The shape is C A3 ×1, C A3 is a preset feature dimension, C A3 >C A2 , and the characteristic dimension C A3 The output feature dimension of the second feature extraction network is consistent; the second nonlinear activation function includes at least a ReLU activation function; The second feature extraction network is composed of a spatial feature encoding module, a spatiotemporal feature aggregation module, a temporal feature encoding module and a layer normalization module connected in sequence; The spatial feature encoding module has a built-in fully connected network and a preset third nonlinear activation function; the spatial feature encoding module is used to convert the left ventricular eddy current area feature sequence Characteristics of each left ventricular vortex area Each of the first eddy current zone characteristics is used as the corresponding eddy current characteristics 1≤indexq(k)≤N k , N k is the kth left ventricular vortex area characteristic The total number of the first eddy current zone features; and each of the eddy current features through a fully connected network Perform feature encoding processing to obtain the corresponding encoding features And based on the third nonlinear activation function, the encoding feature Perform nonlinear activation processing to obtain the corresponding activation features And according to the characteristics of each left ventricular vortex area The corresponding N k The activation feature The corresponding eddy zone feature group And the N1 eddy current zone feature groups are obtained The corresponding eddy zone feature group sequence Sending to the spatiotemporal feature aggregation module; in, W3, b3 are the weight matrix and offset vector of the fully connected network; the eddy current characteristics The shape is C e0 ×1, C e0 is the characteristic dimension of the first eddy current zone feature; the shape of the weight matrix W3 is C e1 ×C e0 , the shape of the offset vector b1 is C e1 ×1, the coding feature The activation feature The shape is C e1 ×1, C e1 is the preset spatial encoding feature dimension, C e1 >C e0 ; The third nonlinear activation function includes at least a ReLU activation function; The spatiotemporal feature aggregation module is used to aggregate the eddy current zone feature group sequence Each of the eddy current zone characteristic groups N k The activation feature Perform spatial feature attention weighted aggregation processing to obtain the corresponding eddy current aggregation feature And the N1 eddy current aggregation characteristics are obtained The corresponding eddy current aggregation characteristic sequence Sending to the temporal feature encoding module; in, α q(k) For the activation feature Spatial feature attention weight, 1≤indexq(k) ’ ≤N k , tanh() is the hyperbolic tangent activation function, W4 and V are two preset weight matrices; the shape of the weight matrix W4 is C e1 ×C e1 , the shape of the weight matrix V is 1×C e1 ,feature The shape is C e1 ×1, The shape of the function output feature is C e1 ×1, the spatial feature attention weight α q(k) is a scalar, the eddy current aggregation characteristic The shape is C e1 ×1; The temporal feature encoding module is implemented based on the BiLSTM model; the temporal feature encoding module is used to encode the eddy current aggregation feature sequence Perform bidirectional temporal feature encoding to obtain the corresponding left-to-right encoding feature sequence and right-to-left encoding feature sequences And the N1th encoding features from left to right and the first right-to-left encoding feature Spliced ​​into a corresponding bidirectional encoding feature vector recorded as feature vector D 3 Send to the layer normalization module; wherein the left-to-right encoding feature sequence The features are encoded from left to right by N1 Composition, the encoding feature sequence from right to left The features are encoded from right to left by N1 Composition, the encoding features from left to right and the right-to-left encoding feature The shape is C EN ×1, C EN is the preset one-way coding feature dimension; the feature vector D 3 The shape is C Bi-EN ×1, C Bi-EN is the preset bidirectional encoding feature dimension, C Bi-EN =2×C EN , the bidirectional encoding feature dimension C Bi-EN It is also the output feature dimension of the second feature extraction network; The layer normalization module is used to normalize the feature vector D 3 Perform layer normalization to obtain the corresponding feature vector D 4 Send to the feature fusion module; the feature vector D 4 Shape C Bi-EN ×1; The feature fusion module is used to 5 and the eigenvector D 4 Perform cross-modal correlation matrix calculation to obtain the corresponding correlation matrix R; and use the correlation matrix R and the eigenvector D 4 Multiply to get the corresponding eigenvector D 5 ; and by the feature vector A 5 and the eigenvector D 5 Splice into a corresponding fusion feature vector G 1 To the low-dimensional feature mapping module; in, softmax() is the Softmax activation function, W5 is a preset weight matrix; the eigenvector A 5 The shape is C A3 ×1, the eigenvector D 4 The shape is C Bi-EN ×1, C A3 =C Bi-EN , the shape of the weight matrix W5 is C A3 ×C A3 =C Bi-EN ×C Bi-EN , feature W5A 5 The shape is C A3 ×1, feature W5D 4 The shape is C Bi-EN ×1=C A3 ×1, feature W5D 4 Transpose (W5D 4 ) T The shape is 1×C A3 , Features (W5A 5 )(W5D 4 ) T The shape is C A3 ×C A3 =C Bi-EN ×C Bi-EN , the shape of the incidence matrix R is C A3 ×C A3 =C Bi-EN ×C Bi-EN ; The feature vector D 5 The shape is C Bi-EN ×1; the fused feature vector G 1 The shape is C G1 ×1, C G1 is the preset fusion feature dimension, C G1 =2×C Bi-EN ; The low-dimensional feature mapping module is formed by sequentially connecting a first linear layer, a preset fourth nonlinear activation function, and a second linear layer; the first linear layer is used to transform the fused feature vector G 1 Perform feature dimensionality reduction processing to obtain the corresponding reduced dimensionality feature vector G 2 ; The fourth nonlinear activation function is used to reduce the dimension of the feature vector G 2 Perform nonlinear activation processing to obtain the corresponding activation feature vector G 3 ; The second linear layer is used to activate the feature vector G 3 Perform two-dimensional feature extraction to obtain the corresponding two-dimensional feature vector G 4 Send to the Softmax function module; Among them, G 2 =W6G 1 +b4, G 4 =W7G 3 +b5; W6 and b4 are the weight matrix and offset vector of the first linear layer, W7 and b5 are the weight matrix and offset vector of the second linear layer; the shape of the weight matrix W6 is C G2 ×C G1 , the shape of the offset vector b4 is C G2 ×1, C G2 is the preset dimension reduction feature dimension, the dimension reduction feature vector G 2 and the activation feature vector G 3 The shape is C G2 ×1; the shape of the weight matrix W7 is C G3 ×C G2 , the shape of the offset vector b5 is C G3 ×1, C G3 is the preset two-dimensional feature dimension, C G3 =2, the two-dimensional feature vector G 4 The shape is C G3 ×1; the fourth nonlinear activation function includes at least a ReLU activation function; The Softmax function module is used to calculate the two-dimensional feature vector G 4 Perform binary classification probability calculation to obtain the corresponding first classification probability ρ1 and the second classification probability ρ2; and form the corresponding first classification vector from the first classification probability ρ1 and the second classification probability ρ2 and output it.

4. The method for predicting heart failure with preserved ejection fraction according to claim 2, wherein: The model building dataset is recorded as the first dataset, which specifically includes: Recruiting multiple volunteers with heart failure with preserved ejection fraction and multiple volunteers without heart failure with preserved ejection fraction to form corresponding volunteer groups; And each of the volunteers in the volunteer set is taken as the corresponding current volunteer; and based on the four-dimensional blood flow cardiac magnetic resonance imaging technology, four-dimensional blood flow cardiac magnetic resonance imaging data of one cardiac cycle is collected for the current volunteer; and the cardiac structural characteristics, cardiac hemodynamic characteristics and left ventricular eddy current timing characteristics of the four-dimensional blood flow cardiac magnetic resonance imaging data collected this time are analyzed to obtain the corresponding structural feature group, the hemodynamic feature group and the eddy current timing feature to form a corresponding first training feature data set; and whether the current volunteer suffers from heart failure with preserved ejection fraction is identified; if so, the corresponding first label probability is set =1, the second label probability If it is 0, then set the corresponding first label probability is 0, the second label probability is 1; and the first label probability obtained at that time and the second label probability Forming a corresponding first label classification vector; and forming a corresponding first data record by the first training feature data set corresponding to the current volunteer and the first label classification vector; All the obtained first data records form the corresponding first data set.

5. The method for predicting heart failure with preserved ejection fraction according to claim 2, wherein: The training of the first prediction model based on the first data set specifically includes: Step 51: randomly split the first data set into two sub-data sets based on a preset first split ratio as a corresponding first training set and a first evaluation set; Wherein, both the first training set and the first evaluation set are composed of a plurality of the first data records; the ratio of the total number of records in the first training set and the first evaluation set satisfies the first segmentation ratio; Step 52: Perform a round of traversal on all the first data records in the first training set; during this round of traversal, use the currently traversed first data record as the corresponding current training record; input the first training feature data set of the current training record as the corresponding first feature data set into the first prediction model for prediction, and use the first classification vector output by this prediction as the corresponding first prediction classification vector; and form a corresponding first prediction-label pair from the first prediction classification vector and the first label classification vector of the current training record; and at the end of this round of traversal, bring all the obtained first prediction-label pairs into the preset first model loss function to calculate the corresponding first loss value; Wherein, the first model loss function is implemented based on a binary cross entropy loss function, a weighted binary cross entropy loss function or a focal loss function; Step 53: Identify whether the first loss value satisfies a preset first loss value range. If so, proceed to step 54. If not, perform a round of modulation on the model parameters of the first prediction model in a direction that minimizes the first model loss function based on a preset first model optimizer, and return to step 52 at the end of this round of modulation. Wherein, the first model optimizer includes at least an Adam optimizer and an SFD optimizer; Step 54: Perform a round of traversal on all the first data records of the first evaluation set; and during this round of traversal, use the first data record currently traversed as the corresponding current evaluation record; and input the first training feature data set of the current evaluation record as the corresponding first feature data set into the first prediction model for prediction, and use the first classification vector output by this prediction as the corresponding second prediction classification vector; and form a corresponding second prediction-label pair from the second prediction classification vector and the first label classification vector of the current evaluation record; and at the end of this round of traversal, bring all the obtained second prediction-label pairs into the preset first model evaluation function to calculate and obtain the corresponding first evaluation value; Wherein, the first model evaluation function is implemented based on MAE function, MSE function or RMSE function; Step 55 , identifying whether the first evaluation value satisfies a preset first evaluation value range; if not, returning to step 52 ; if so, confirming that the model training is completed.

6. A device for executing the method for predicting heart failure with preserved ejection fraction according to any one of claims 1 to 5, characterized in that: The device includes: a model construction module, a model training module and a model application module; The model building module is used to construct a deep learning model for binary prediction of heart failure with preserved ejection fraction, which is recorded as the first prediction model; the first prediction model is used to perform binary prediction of heart failure with preserved ejection fraction based on the first feature data set input by the model and output the corresponding first classification vector; the first feature data set includes a structural feature group, a hemodynamic feature group and an eddy current timing feature; the structural feature group includes the left atrial maximum volume index, the left atrial minimum volume index, the left atrial longitudinal peak strain, the left ventricular mass index, and the left ventricular overall longitudinal strain; the hemodynamic feature group includes left ventricular filling pressure, left ventricular kinetic energy, left ventricular hemodynamics, left ventricular myocardial work efficiency, left ventricular energy loss, the ratio of early diastolic mitral blood flow peak velocity to mitral annulus motion velocity, tricuspid regurgitation velocity, and left ventricular ejection fraction; the eddy current timing feature is a left A time series sequence of ventricular vortex zone features, wherein the time span of the vortex time series features covers a complete cardiac cycle; the vortex time series features are formed by sequentially sorting the left ventricular vortex zone features at multiple sampling time points, and the time interval between each two adjacent time points is the same; each left ventricular vortex zone feature is composed of one or more first vortex zone features, and each first vortex zone feature is used to reflect the characteristics of a vortex zone in the left ventricle at the current time point; each first vortex zone feature includes the center position of the vortex zone, the maximum blood flow of the vortex zone, the transverse radius of the vortex zone, the longitudinal radius of the vortex zone, and the vortex strength of the vortex zone; the first classification vector includes a first classification probability ρ1 and a second classification probability ρ2, the first classification probability ρ1 is the predicted probability of heart failure with preserved ejection fraction, and the second classification probability ρ2 is the predicted probability of heart failure without preserved ejection fraction, ρ1+ρ2=1; The model training module is used to construct a model data set recorded as a first data set; and train the first prediction model based on the first data set; The model application module is used to analyze the cardiac structural characteristics, cardiac hemodynamic characteristics and left ventricular eddy current timing characteristics of the four-dimensional blood flow cardiac magnetic resonance imaging data input by the user after the model training is completed, to obtain the corresponding first feature data set composed of the structural feature group, the hemodynamic feature group and the eddy current timing feature; and to input the first feature data set into the first prediction model for prediction to obtain the corresponding first classification vector and feedback it to the current user.

7. An electronic device, characterized in that: include: memory, processors, and transceivers; The processor is configured to be coupled to the memory, read and execute instructions in the memory, so as to implement the method according to any one of claims 1 to 5; The transceiver is coupled to the processor, and the processor controls the transceiver to send and receive messages.

8. A computer-readable storage medium, characterized in that The computer-readable storage medium stores computer instructions, and when the computer instructions are executed by a computer, the computer is caused to execute the method according to any one of claims 1 to 5.

Images