Medical task prediction model construction method

Abstract

The invention provides a medical task prediction model construction method, which comprises the steps of S1, acquiring a training data set which comprises time sequence blood pressure data of a plurality of preset time periods; s2, constructing an initial model, wherein the initial model comprises a feature extraction module, a multi-channel prototype network module and a full connection layer; and S3, taking the time sequence blood pressure data as input, taking a disease classification result of the time sequence blood pressure data as prediction output, training the initial model according to a preset training rule until convergence, and obtaining a medical task prediction model. According to the medical task prediction model provided by the invention, feature extraction and similarity calculation can be carried out on data of different channels in the time sequence blood pressure data, so that feature information of the data of the different channels is fully obtained, and the model prediction precision is further improved; and the time sequence blood pressure data can be matched with the global prototype, so that the decision-making process of the model can be intuitively understood.

Description

Technical Field

[0001] The present invention relates to the field of deep learning, specifically to a model building technology in the field of deep learning, and more specifically to a method for building a medical task prediction model. Background Art

[0002] As the leading cause of death worldwide, cardiovascular disease poses a severe challenge to public health with its high morbidity and mortality. In the evaluation system of cardiovascular health, multi-channel blood pressure monitoring occupies a pivotal position. Blood pressure, as a comprehensive physiological indicator, usually covers multiple dimensions such as systolic pressure, diastolic pressure and mean arterial pressure, each of which carries unique physiological meanings and clinical values. These blood pressure data from different channels play an irreplaceable role in the diagnosis, treatment and prognosis of various cardiovascular diseases. Especially during surgery, continuous monitoring of multi-channel blood pressure is particularly important. Surgical operations are often accompanied by drastic changes in the patient's physiological state. The sharp fluctuations in blood pressure may directly induce serious cardiovascular events such as stroke and myocardial infarction, thereby significantly increasing the risk of postoperative complications and death. Therefore, by real-time monitoring and analyzing the changing trends of blood pressure data from each channel during surgery, abnormal conditions can be quickly identified, intervention measures can be taken in a timely manner, and the deterioration of the disease can be effectively curbed to ensure patient safety.

[0003] In recent years, with the rapid development of artificial intelligence technology, especially the widespread application of deep learning technology, the medical field has ushered in unprecedented changes. In terms of cardiovascular disease prediction, deep learning models have significantly improved the accuracy of disease prediction with their powerful data processing and pattern recognition capabilities, providing more reliable data support for clinical decision-making. However, the high complexity and non-intuitiveness of deep learning models have also brought new problems - the lack of interpretability of model decisions. Due to the extremely complex internal structure and decision-making logic of deep learning models, their output results are often difficult for medical professionals to intuitively understand and accept, which largely limits the in-depth application of deep learning technology in clinical practice.

[0004] In addition, in the process of processing multi-channel blood pressure data, existing deep learning models often ignore the possible independence between different channels. In fact, although systolic blood pressure, diastolic blood pressure and mean arterial pressure belong to the same category of blood pressure, they have significant differences in physiological mechanisms, influencing factors and clinical significance. Therefore, when building a prediction model, if these differences are not fully considered, it may lead to a decline in model performance.

[0005] In summary, although deep learning models have shown great potential in the prediction of cardiovascular diseases, deep learning models built using existing technologies still face two challenges: first, the model's interpretability is insufficient. Due to the extremely complex internal structure and decision-making logic of deep learning models, it is difficult for medical professionals to intuitively understand the model's decision-making process; second, the independence between multi-channel blood pressure data is not fully utilized, resulting in low prediction accuracy of deep learning models. Summary of the invention

[0006] Therefore, the purpose of the present invention is to overcome the defects of the above-mentioned prior art and provide a method for constructing a medical task prediction model.

[0007] The objectives of the present invention are achieved through the following technical solutions.

[0008] According to a first aspect of the present invention, a method for constructing a medical task prediction model is provided, the method comprising: step S1, obtaining a training data set, the training data set comprising time-series blood pressure data of multiple preset time periods, wherein each time-series blood pressure data comprises data of multiple channels, and each time-series blood pressure data is provided with a corresponding disease label; step S2, constructing an initial model, the initial model comprising a feature extraction module, a multi-channel prototype network module and a fully connected layer, wherein: the feature extraction module is used to perform feature extraction on the time-series blood pressure data to obtain a feature matrix of the time-series blood pressure data, wherein the feature matrix of the time-series blood pressure data comprises feature sub-matrices of multiple channels; the multi-channel prototype network module comprises multiple ProtoPNet units, wherein each ProtoPNet unit is configured with multiple global prototypes, and one global prototype corresponds to one disease label The target similarity matrix is ​​a matrix of the feature matrix of the time series blood pressure data and each global prototype configured on the ProtoPNet unit according to a preset processing method to obtain a target similarity matrix. An element in the target similarity matrix represents the similarity score between the feature data in a channel of the time series blood pressure data that is most similar to a global prototype and the global prototype; the fully connected layer is used to perform predictive processing on multiple target similarity matrices obtained by the multi-channel prototype network module to obtain the disease classification results of the time series blood pressure data; step S3, taking the time series blood pressure data as input and the disease classification results of the time series blood pressure data as the predicted output, train the initial model according to the preset training rules until convergence to obtain the medical task prediction model.

[0009] In some embodiments of the present invention, each time-series blood pressure data includes data of three channels: diastolic pressure, systolic pressure and mean arterial pressure.

[0010] In some embodiments of the present invention, the feature extraction module is a gated Transtormer network.

[0011] In some embodiments of the present invention, the preset processing method is: calculating the similarity between the feature submatrix of each channel in the feature matrix of the time series blood pressure data and each global prototype configured on the ProtoPNet unit to obtain a similarity matrix corresponding to the number of global prototypes, each similarity matrix includes the feature submatrix of each channel of the time series blood pressure data and the similarity submatrix obtained after the corresponding global prototype is calculated, wherein the similarity submatrix is ​​calculated as follows:

[0012]

[0013] in, Indicates the time series blood pressure data The characteristic sub-matrix of each channel and the global prototype The calculated similarity sub-matrix; Indicates the first The feature sub-matrix of each channel; Represents the feature submatrix A feature point in Indicates The first of the ProtoPNet units A global prototype; Represents a decimal; Representing feature points With global prototype The square of the Euclidean distance between them; perform maximum pooling on each similarity matrix, and concatenate all similarity matrices after pooling to obtain the target similarity matrix.

[0014] In some embodiments of the present invention, the preset training rule is to perform two-stage training on the initial model in the following manner: the first stage training uses the time series blood pressure data as input, the disease classification results of the time series blood pressure data as the prediction output, performs multiple rounds of iterative training and updates the initial model parameters with the first objective function to obtain an intermediate model; the second stage training uses the time series blood pressure data as input, the disease classification results of the time series blood pressure data as the prediction output, performs multiple rounds of iterative training and updates the intermediate model parameters with the second objective function until convergence to obtain a medical task prediction model; wherein, in each round of iterative training, each global prototype of each ProtoPNet unit in the multi-channel prototype network module of the intermediate model is updated according to a preset matching method.

[0015] In some embodiments of the present invention, the first objective function is:

[0016]

[0017] in, represents the cross entropy loss function, Represents the number of time-series blood pressure data, represents the disease labels of time-series blood pressure data, Indicates Disease labels for time series blood pressure data, represents the predicted output of time series blood pressure data, Indicates Prediction output of time series blood pressure data;

[0018] The second objective function is:

[0019]

[0020] in,

[0021]

[0022] in,

[0023]

[0024]

[0025]

[0026] in, represents the second objective function, represents the multi-channel loss, represents the number of ProtoPNet units, Indicates The weight parameters of the ProtoPNet units, Indicates ProtoPNet units, Indicates The first loss of the ProtoPNet unit, Indicates The second loss of the ProtoPNet unit, Indicates The third loss of the ProtoPNet unit, Indicated in ProtoPNet units are selected from A global prototype with consistent disease label categories for time-series blood pressure data , Represents the feature submatrix Belong to The feature matrix of time series blood pressure data, Represents the first The feature matrix of time series blood pressure data, Indicated in ProtoPNet units are selected from A global prototype for inconsistent disease label categories in time-series blood pressure data , Represents the feature submatrix A feature point in Representing feature points With global prototype The square of the Euclidean distance between Global prototype With the global prototype The square of the Euclidean distance between Indicates The first of the ProtoPNet units global prototype, global prototype With the global prototype Corresponding to different disease label categories.

[0027] In some embodiments of the present invention, the preset matching method is:

[0028]

[0029] in, Representing feature points The first The feature point set in the feature sub-matrix of the channel , Represents the first The first time series of blood pressure data The feature sub-matrix of the channels, Indicates The disease label categories of the time series blood pressure data are classified as follows: , Indicates The disease classification in ProtoPNet units is The global prototype collection of .

[0030] According to the second aspect of the present invention, a medical task prediction method is provided, the method comprising: step T1, obtaining time-series blood pressure data to be processed; step T2, using the medical task prediction model constructed by the method described in the first aspect of the present invention to perform prediction processing on the time-series blood pressure data to be processed to obtain a disease classification result, wherein in the prediction processing process, multiple target similarity matrices obtained by the medical task prediction model are used as a benchmark, each target similarity matrix is ​​matched with a feature matrix of the time-series blood pressure data to be processed to obtain a local data segment of a channel in the time-series blood pressure data corresponding to each element in each target similarity matrix, and one or more local data segments that meet preset requirements are selected from all the obtained local data segments as prediction indicators to explain the disease classification result obtained by the prediction processing.

[0031] Compared with the prior art, the advantages of the present invention are: (1) considering the independence between different channels of time-series blood pressure data, the medical task prediction model can perform feature extraction and similarity calculation on the data of different channels in the time-series blood pressure data, so as to fully obtain the feature information of the data of different channels, thereby improving the prediction accuracy of the model; (2) matching the time-series blood pressure data with the global prototype, using the local data fragments in the matched time-series blood pressure data as the prediction index to explain the classification basis of the medical task prediction model, thereby intuitively understanding the decision-making process of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] The embodiments of the present invention are further described below with reference to the accompanying drawings, in which:

[0033] Figure 1 A schematic diagram of a process for constructing a medical task prediction model according to an embodiment of the present invention;

[0034] Figure 2 A schematic diagram of the composition of an initial model according to an embodiment of the present invention;

[0035] Figure 3 It is a schematic diagram showing an example of matching results between multiple test time series blood pressure data and a global prototype according to an embodiment of the present invention;

[0036] Figure 4 The figure is a schematic diagram showing an example of matching results between test time series blood pressure data and a global prototype according to an embodiment of the present invention. DETAILED DESCRIPTION

[0037] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below through specific embodiments in conjunction with the accompanying drawings. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0038] As mentioned in the background technology section, although deep learning models have shown great potential in the prediction of cardiovascular diseases, deep learning models built using existing technologies still face two challenges: First, the model's interpretability is insufficient. Since the internal structure and decision-making logic of the deep learning model are extremely complex, it is difficult for medical professionals to intuitively understand the model's decision-making process; second, the independence between multi-channel blood pressure data is not fully utilized, resulting in low prediction accuracy of the deep learning model.

[0039] In order to solve the above problems, the inventors studied the prior art and found that the prior art usually uses ProtoPNet to solve the problem of model interpretability. In the prior art, ProtoPNet in the field of image recognition is introduced when building a deep learning model (an interpretable method used in the field of fine-grained image recognition, which classifies images by obtaining the similarity information between each area in the image sample and the learned prototype, making the decision-making process of the model more transparent and interpretable). However, since image data and time series data have different characteristics, there is a strong correlation in multiple directions such as up, down, left, and right of a data point in the image data, while time series data pays more attention to the information in the front and back directions of a data point. Therefore, even if ProtoPNet is introduced, the problem of model interpretability cannot be solved well. Especially for multi-channel time series such as time series blood pressure data, researchers did not take into account the independence between different channels, resulting in the loss of unique feature information between channels, and even affecting the interpretability effect. For example, in the complication prediction task based on time-series blood pressure data, because the multi-channel situation is not considered, the constructed deep learning model automatically uses the systolic pressure prototype to match the diastolic pressure channel data, which directly leads to a decrease in the prototype similarity and affects the overall explanatory effect of the deep learning model.

[0040] Based on the above analysis, the inventors proposed a medical task model construction scheme for multi-channel time-series blood pressure data, in which multiple ProtoPNet units are set to address the model interpretability problem, wherein each ProtoPNet unit includes multiple global prototypes, which are used to perform similarity comparisons with each channel in the time-series blood pressure data to obtain a target similarity matrix corresponding to each ProtoPNet unit, and each target similarity matrix is ​​used to determine the local data segments of each channel in the time-series blood pressure data that have the highest similarity with each global prototype; the disease category of the time-series blood pressure data is predicted based on multiple target similarity matrices, and at the same time, the local data segments obtained are used as interpretable support for the prediction of the medical task model, so that medical staff can better understand the decision-making process of the medical task model.

[0041] In summary, if Figure 1As shown, the present invention proposes a method for constructing a medical task prediction model, the method comprising: step S1, obtaining a training data set, the training data set comprising time-series blood pressure data of multiple preset time periods, wherein each time-series blood pressure data comprises data of multiple channels, and each time-series blood pressure data is provided with a corresponding disease label; step S2, constructing an initial model, the initial model comprising a feature extraction module, a multi-channel prototype network module and a fully connected layer, wherein: the feature extraction module is used to perform feature extraction on the time-series blood pressure data to obtain a feature matrix of the time-series blood pressure data, wherein the feature matrix of the time-series blood pressure data comprises feature sub-matrices of multiple channels; the multi-channel prototype network module comprises multiple ProtoPNet units, wherein each ProtoPNet unit is configured with multiple global prototypes, one global prototype corresponds to one disease label corresponding to The target similarity matrix is ​​a characteristic fragment of a channel in the time series blood pressure data, and different global prototypes correspond to different disease categories. Each ProtoPNet unit is used to calculate the similarity between the characteristic matrix of the time series blood pressure data and each global prototype configured on the ProtoPNet unit according to a preset processing method to obtain a target similarity matrix, and an element in the target similarity matrix represents the similarity score between the characteristic data in a channel of the time series blood pressure data that is most similar to a global prototype and the global prototype; the fully connected layer is used to perform predictive processing on multiple target similarity matrices obtained by the multi-channel prototype network module to obtain the disease classification result of the time series blood pressure data; step S3, taking the time series blood pressure data as input and the disease classification result of the time series blood pressure data as the prediction output, the initial model is trained according to the preset training rules until convergence to obtain the medical task prediction model.

[0042] In order to better understand the present invention, each step is described in detail below in conjunction with specific embodiments.

[0043] 1. Step S1

[0044] In step S1, a training data set is obtained. The training data set includes time-series blood pressure data of multiple preset time periods, wherein each time-series blood pressure data includes data of multiple channels, and each time-series blood pressure data is provided with a corresponding disease label. According to one embodiment of the present invention, each time-series blood pressure data includes data of three channels: diastolic pressure, systolic pressure and mean arterial pressure. Take a specific example to introduce the time-series blood pressure data: the time-series blood pressure data includes data of three channels: systolic pressure, diastolic pressure and mean arterial pressure, wherein the data acquisition frequency of each channel is set to once every 0.5 minutes, and the data acquisition is continuously collected for 350 minutes. In this case, the data length of each channel is 700 (700 time points).

[0045] Step S2

[0046] In step S2, an initial model is constructed. Figure 2 As shown, the initial model includes a feature extraction module, a multi-channel prototype network module and a fully connected layer. The following is a detailed description of each component of the initial model.

[0047] The feature extraction module is used to extract features from the time series blood pressure data to obtain a feature matrix of the time series blood pressure data, wherein the feature matrix of the time series blood pressure data includes feature sub-matrices of multiple channels, namely, a feature sub-matrix of diastolic pressure, a feature sub-matrix of systolic pressure, and a feature sub-matrix of mean arterial pressure. According to one embodiment of the present invention, the feature extraction module is a gated Transtormer network.

[0048] It should be noted that, since the time-series blood pressure data includes data from three channels: diastolic pressure, systolic pressure, and mean arterial pressure, the feature extraction module needs to perform feature extraction on each channel of the time-series blood pressure data to obtain a feature submatrix for each channel of the time-series blood pressure data. Among them, the feature submatrix of each channel obtained by feature extraction includes multiple feature values, and each feature value represents the data of a portion of the time-series segment of its corresponding channel. Taking the time-series blood pressure data with a data length of 700 for each channel as described in the above example as an example for feature extraction, after the time-series blood pressure data is processed by the feature extraction module, the feature submatrix of each channel includes 23 feature points, which means that the 700 time points in each channel are divided into 23 regions and a feature point is extracted from each region as the representative of the region.

[0049] Depend on Figure 2 It can be seen that the multi-channel prototype network module includes multiple ProtoPNet units, wherein each ProtoPNet unit is configured with multiple global prototypes, a global prototype corresponds to a feature segment of a channel in the time series blood pressure data corresponding to a disease label, and different global prototypes correspond to different disease categories, and each ProtoPNet unit is used to calculate the similarity between the feature matrix of the time series blood pressure data and each global prototype configured on the ProtoPNet unit according to a preset processing method to obtain a target similarity matrix, wherein an element in the target similarity matrix represents the similarity score between the feature data in a channel of the time series blood pressure data that is most similar to a global prototype and the global prototype.

[0050] According to one embodiment of the present invention, the preset processing method is: calculating the similarity between the feature submatrix of each channel in the feature matrix of the time series blood pressure data and each global prototype configured on the ProtoPNet unit to obtain a similarity matrix corresponding to the number of global prototypes, each similarity matrix includes the feature submatrix of each channel of the time series blood pressure data and the similarity submatrix obtained after the corresponding global prototype is calculated, wherein the similarity submatrix is ​​calculated as follows:

[0051]

[0052] in, Indicates the first The characteristic sub-matrix of each channel and the global prototype The calculated similarity sub-matrix; Indicates the first The feature sub-matrix of each channel; Represents the feature submatrix A feature point in Indicates The first of the ProtoPNet units A global prototype; Represents a decimal; Representing feature points With the global prototype The square of the Euclidean distance between them; perform maximum pooling on each similarity matrix, and concatenate all similarity matrices after pooling to obtain the target similarity matrix.

[0053] It should be noted that each ProtoPNet unit includes a prototype network and a maximum pooling layer. Among them, multiple global prototypes are configured in the prototype network, and each global prototype calculates the similarity with the feature matrix of the time-series blood pressure data to obtain the corresponding similarity matrix. The maximum pooling layer is used to perform maximum pooling processing on the similarity matrix corresponding to each global prototype in the ProtoPNet unit in which it is located. After pooling, each element in the similarity matrix represents the similarity score between the feature data (feature point) most similar to the global prototype in the corresponding channel of the time-series blood pressure data and the global prototype. All similarity matrices after pooling are spliced ​​to obtain the target similarity matrix.

[0054] The fully connected layer is used to perform prediction processing on multiple target similarity matrices obtained by the multi-channel prototype network module to obtain disease classification results of time-series blood pressure data.

[0055] Step S3

[0056] In step S3, the time series blood pressure data is used as input, and the disease classification results of the time series blood pressure data are used as prediction outputs. The initial model is trained according to the preset training rules until convergence, and a medical task prediction model is obtained. According to one embodiment of the present invention, the preset training rules are to perform two-stage training on the initial model. The training process of the two-stage training is introduced below.

[0057] The first stage of training uses the time series blood pressure data as input and the disease classification results of the time series blood pressure data as the prediction output. Multiple rounds of iterative training are performed and the initial model parameters are updated with the first objective function to obtain an intermediate model. The first objective function is:

[0058]

[0059] in, represents the cross entropy loss function, Represents the number of time-series blood pressure data, represents the disease labels of time-series blood pressure data, Indicates Disease labels for time series blood pressure data, represents the predicted output of time series blood pressure data, Indicates The predicted output of time series blood pressure data.

[0060] It should be noted that the iterative rounds of the first stage training can be specified as 10 rounds, 15 rounds or 20 rounds, and each round of iterative training can use 32 time-series blood pressure data. The specific rounds and the number of time-series blood pressure data used in each round are determined by actual needs, and the present invention does not impose any special restrictions.

[0061] The second stage of training uses the time series blood pressure data as input and the disease classification results of the time series blood pressure data as the prediction output. Multiple rounds of iterative training are performed and the intermediate model parameters are updated with the second objective function until convergence to obtain the medical task prediction model; wherein, in each round of iterative training, each global prototype of each ProtoPNet unit in the multi-channel prototype network module of the intermediate model is updated according to the preset matching method. The second objective function is:

[0062]

[0063] in,

[0064]

[0065] in,

[0066]

[0067]

[0068]

[0069] in, represents the second objective function, represents the multi-channel loss, represents the number of ProtoPNet units, Indicates The weight parameters of the ProtoPNet units, Indicates ProtoPNet units, Indicates The first loss of the ProtoPNet unit, Indicates The second loss of the ProtoPNet unit, Indicates The third loss of the ProtoPNet unit, Indicated in ProtoPNet units are selected from A global prototype with consistent disease label categories for time-series blood pressure data , Represents the feature submatrix Belong to The feature matrix of time series blood pressure data, Represents the first The feature matrix of time series blood pressure data, Indicated in ProtoPNet units are selected from A global prototype for inconsistent disease label categories in time-series blood pressure data , Represents the feature submatrix A feature point in Representing feature points With global prototype The square of the Euclidean distance between Global prototype With global prototype The square of the Euclidean distance between Indicates The first of the ProtoPNet units global prototype, global prototype With global prototype Corresponding to different disease label categories. It should be noted that in the second stage of training, in addition to the cross entropy loss, the second objective function also introduces a multi-channel loss, which calculates the distance between each global prototype and the time-series blood pressure data of the same classification and the time-series blood pressure data of different classifications in each round, as well as the distance between each global prototype on the same ProtoPNet unit and other global prototypes, so as to increase the prediction accuracy of the time-series blood pressure data and the distance between the global prototypes of different disease classifications.

[0070] In the second stage of training, not only the parameters of the intermediate model need to be updated according to the second objective function, but also each global prototype of each ProtoPNet unit in the multi-channel prototype network module of the intermediate model needs to be updated according to a preset matching method. According to one embodiment of the present invention, the preset matching method is:

[0071]

[0072] in, Representing feature points The first The feature point set in the feature sub-matrix of the channel , Represents the first The first time series of blood pressure data The feature sub-matrix of the channels, Indicates The disease label categories of the time series blood pressure data are classified as follows: , Indicates The disease classification in ProtoPNet units is A set of global prototypes. It should be noted that each global prototype in each ProtoPNet unit is a feature data (cluster center) obtained by iteratively calculating and clustering the feature matrix of the same channel of multiple time-series blood pressure data with the same disease label category in units of channels. The feature data represents the most similar and representative data fragments in the same channel of multiple time-series blood pressure data with the same disease label category. For example, if a global prototype corresponds to the diastolic pressure channel in disease category A, then during the second stage of training, the global prototype will continuously learn the diastolic pressure channel data in the time-series blood pressure data with disease category A, that is, continuously reduce the square of the Euclidean distance between the global prototype and the same channel of the time-series blood pressure with the same disease label category, and then obtain the cluster center of the diastolic pressure channel data in multiple time-series blood pressure data with disease category A, and the cluster center represents the most similar part of the diastolic pressure channel data in multiple time-series blood pressure data with disease category A.

[0073] Based on the medical task prediction model constructed by the above embodiment, the present invention also provides a medical task prediction method, which includes: step T1, obtaining the time series blood pressure data to be processed; step T2, using the medical task prediction model constructed by the method described in the above embodiment to perform predictive processing on the time series blood pressure data to be processed to obtain a disease classification result, wherein in the prediction processing process, multiple target similarity matrices obtained by the medical task prediction model are used as a benchmark, and each target similarity matrix is ​​matched with a feature matrix of the time series blood pressure data to be processed to obtain a local data segment of a channel in the time series blood pressure data corresponding to each element in each target similarity matrix, and one or more local data segments that meet preset requirements are selected from all the obtained local data segments as prediction indicators to explain the disease classification results obtained by the prediction processing.

[0074] In order to better understand the execution process of medical task prediction, the following is a detailed description of the medical task prediction method described in the above embodiment. The specific execution process is: first, obtain the time series blood pressure data to be processed; then, use the medical task prediction model constructed by the method described in the above embodiment to predict the time series blood pressure data to be processed to obtain the disease classification result. Among them, the specific processing process of the medical task prediction model is: first, the feature extraction module extracts features from the time series blood pressure data to be processed to obtain a feature matrix; then, the multi-channel prototype network module calculates the similarity of the feature matrix: the feature matrix and each global prototype in each ProtoPNet unit calculate the similarity to obtain a similarity matrix corresponding to the number of global prototypes in each ProtoPNet unit, and each similarity matrix in each ProtoPNet unit is subjected to maximum pooling processing, and all similarity matrices after pooling processing in each ProtoPNet unit are spliced ​​to obtain the target similarity matrix corresponding to each ProtoPNet unit; finally, the fully connected layer predicts the multiple target similarity matrices transmitted by the multi-channel prototype network module to obtain the disease classification result of the time series blood pressure data to be processed.

[0075] It should be noted that in order to enable medical staff to understand the decision-making process of the medical task prediction model, multiple target similarity matrices obtained by the medical task prediction model during the prediction process can be used as a benchmark, and each target similarity matrix can be matched with the feature matrix of the time series blood pressure data to be processed to obtain the local data fragments of the channel in the time series blood pressure data corresponding to each element in each target similarity matrix, and one or more local data fragments that meet the preset requirements are selected from all the local data fragments obtained as prediction indicators to explain the disease classification results obtained by the prediction processing. For example, suppose there is a time series blood pressure data with 3 channels. After the time series blood pressure data is processed by the medical task prediction model, a target similarity matrix consistent with the number of ProtoPNet units is obtained; wherein each target similarity matrix contains 3 sub-matrices, each sub-matrix includes n similarity scores (n is consistent with the number of global prototypes in the ProtoPNet unit), and each similarity score represents the similarity score between the feature data of the corresponding channel in the time series blood pressure data and the corresponding global prototype in the ProtoPNet unit, and the global prototype. The higher the similarity score, the more similar the local data fragment matched by the feature data of the corresponding channel in the time series blood pressure data is to the corresponding global prototype. Therefore, the reason why the medical task prediction model makes the corresponding disease classification result can be explained by obtaining the local data fragment of the channel in the time series blood pressure data corresponding to each element (similarity score) in each target similarity matrix, and selecting one or more local data fragments that meet the preset requirements (for example, the similarity score is greater than or equal to 0.8) from these local data fragments as prediction indicators.

[0076] It should also be noted that one or more local data fragments used as prediction indicators are usually matched with the global prototype corresponding to the ProtoPNet unit. The specific matching results are as follows: Figure 3 and Figure 4 As shown in the example in . Figure 3It can be seen that the left side, from top to bottom, are the global prototypes for systolic blood pressure (Prototype for Systolic Blood Pressure), diastolic blood pressure (Prototype for Diastolic Blood Pressure) and mean arterial pressure (Prototype for Mean Arterial Pressure); the right side is the data of one channel of test samples 138, 240, 5, 41, 1022 and 1308 respectively; among them, the parts selected by the red vertical lines in test samples 138 and 240 match the parts selected by the red vertical lines in the global prototype of systolic blood pressure on the right; the parts selected by the red vertical lines in test samples 5 and 41 match the parts selected by the red vertical lines in the global prototype of diastolic blood pressure on the right; the parts selected by the red vertical lines in test samples 1022 and 1308 match the parts selected by the red vertical lines in the global prototype of mean arterial pressure on the right. Figure 4 It can be seen that the left side is the test sample, and the right side, from top to bottom, is the global prototype of systolic pressure, the global prototype of diastolic pressure, and the global prototype of mean arterial pressure. Among them, the parts selected by red vertical lines in each channel of the test sample match the parts selected by red vertical lines in the global prototype of each channel on the right.

[0077] In order to verify the working performance of the medical task prediction model constructed by the present invention, the medical task prediction model and the existing deep learning model were classified and tested on the same data set, and the experimental results shown in Table 1 were obtained. Existing deep learning models include GTN, LSTM, FCN, ResNet, TapNet, LSTMFCN and InceptionTime; among them, GTN: a multi-channel time series classification model that captures channel-specific and step-by-step dependencies by merging two Transformer towers with gating mechanisms; LSTM: by introducing three gates (input gate, forget gate, output gate) and unit states, this deep learning model is designed to better handle long-term dependencies in sequential data; FCN: a deep learning model that uses CNN to automatically discover and extract the internal structure of input time series data and generate deep features for classification; ResNet: a deep convolutional neural network that solves the performance degradation problem caused by increasing the depth of CNN by utilizing residual modules; TapNet: a deep learning model for multi-dimensional time series classification that learns low-dimensional feature representations and interaction features by paying attention to prototype networks and multi-layer convolutional networks; LSTMFCN: a multi-channel time series classification model achieved by expanding the full convolution block to a squeeze-excitation block; InceptionTime: a deep learning model for time series classification inspired by the Inception network structure.

[0078] As can be seen from Table 1, the medical task prediction model constructed by the present invention shows excellent performance, and its indicators such as Acc, Pre, F1Score and AUC are better than other related models. This shows that the medical task prediction model constructed by the present invention has good disease classification ability, and can maintain efficient discrimination ability when processing highly unbalanced data sets. Among them, Acc (Accuracy): accuracy, used to measure the overall classification performance of the model; Pre (Presicion): precision, which can reflect the accuracy of the model in predicting the positive class; Recall rate: reflects the model's ability to recognize positive samples; F1 (F1 Score): F1 value, taking into account the precision and recall rate of the classification model; AUC (Area Under the ROC Curve): the area under the ROC curve, is a comprehensive indicator used to evaluate the overall discrimination ability of the model.

[0079] Table 1

[0080]

[0081] In order to further analyze the working performance of the medical task prediction model constructed by the present invention, multiple medical task models are trained with different loss functions and the performance of each medical task model is evaluated. The evaluation results are shown in Table 2. Indicates that only the cross entropy loss function is used to train the medical task model; Indicates the use As a loss function to train the medical task model; C-PPT means using As the loss function, the medical task model is trained. As shown in Table 2, The medical task model trained as the loss function outperforms other models in terms of indicators such as Acc, Pre, F1 Score and AUC.

[0082] Table 2

[0083]

[0084] In order to further analyze the working performance of the medical task prediction model constructed by the present invention, As the loss function, the medical task prediction model is trained and the model performance is evaluated in two cases: without adding weights and adding weights. The evaluation results are shown in Table 3. Among them, C-PPT (no λ) means that no weights are added in the loss function, and C-PPT means that weights are added in the loss function. As can be seen from Table 3, the performance of the medical task prediction model trained after adding weights in the loss function is better.

[0085] Table 3

[0086]

[0087] The beneficial effects of the present invention are: (1) considering the independence between different channels of time-series blood pressure data, the medical task prediction model can perform feature extraction and similarity calculation on the data of different channels in the time-series blood pressure data, so as to fully obtain the feature information of the data of different channels, thereby improving the prediction accuracy of the model; (2) matching the time-series blood pressure data with the global prototype, using the local data fragments in the matched time-series blood pressure data as prediction indicators to explain the classification basis of the medical task prediction model, thereby intuitively understanding the decision-making process of the model.

[0088] It should be noted that although the above describes the various steps in a specific order, it does not mean that the various steps must be executed in the above specific order. In fact, some of these steps can be executed concurrently or even in a different order as long as the required functions can be achieved.

[0089] The present invention may be a system, a method and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for causing a processor to implement various aspects of the present invention.

[0090] A computer-readable storage medium may be a tangible device that holds and stores instructions used by an instruction execution device. Computer-readable storage media may include, for example, but are not limited to, an electrical storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination thereof. More specific examples (non-exhaustive list) of computer-readable storage media include: a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), a static random access memory (SRAM), a portable compact disk read-only memory (CD-ROM), a digital versatile disk (DVD), a memory stick, a floppy disk, a mechanical encoding device, such as a punch card or a protruding structure in a groove on which instructions are stored, and any suitable combination thereof.

[0091] The embodiments of the present invention have been described above, and the above description is exemplary, not exhaustive, and is not limited to the disclosed embodiments. Many modifications and variations will be apparent to those of ordinary skill in the art without departing from the scope and spirit of the described embodiments. The selection of terms used herein is intended to best explain the principles of the embodiments, practical applications, or technical improvements in the market, or to enable other persons of ordinary skill in the art to understand the embodiments disclosed herein.

Claims

1. A method for constructing a medical task prediction model, characterized in that: The method comprises: Step S1, obtaining a training data set, wherein the training data set includes time-series blood pressure data of multiple preset time periods, wherein each time-series blood pressure data includes data of multiple channels, and each time-series blood pressure data is provided with a corresponding disease label; Step S2: construct an initial model, which includes a feature extraction module, a multi-channel prototype network module and a fully connected layer, wherein: The feature extraction module is used to extract features from the time series blood pressure data to obtain a feature matrix of the time series blood pressure data, wherein the feature matrix of the time series blood pressure data includes feature sub-matrices of multiple channels; The multi-channel prototype network module includes a plurality of ProtoPNet units, wherein each ProtoPNet unit is configured with a plurality of global prototypes, a global prototype corresponds to a feature segment of a channel in the time series blood pressure data corresponding to a disease label, and different global prototypes correspond to different disease categories, and each ProtoPNet unit is used to calculate the similarity between the feature matrix of the time series blood pressure data and each global prototype configured on the ProtoPNet unit according to a preset processing method to obtain a target similarity matrix, wherein an element in the target similarity matrix represents the similarity score between the feature data in a channel of the time series blood pressure data that is most similar to a global prototype and the global prototype; The fully connected layer is used to perform prediction processing on multiple target similarity matrices obtained by the multi-channel prototype network module to obtain disease classification results of time-series blood pressure data; Step S3: using the time series blood pressure data as input and the disease classification results of the time series blood pressure data as prediction output, the initial model is trained according to the preset training rules until convergence to obtain a medical task prediction model.

2. The method according to claim 1, characterized in that Each time series of blood pressure data includes data of three channels: diastolic pressure, systolic pressure and mean arterial pressure.

3. The method according to claim 2, characterized in that The feature extraction module is a gated Transtormer network.

4. The method according to claim 3, characterized in that The preset processing method is: The similarity between the feature submatrix of each channel in the feature matrix of the time series blood pressure data and each global prototype configured on the ProtoPNet unit is calculated to obtain a similarity matrix corresponding to the number of global prototypes. Each similarity matrix includes the feature submatrix of each channel of the time series blood pressure data and the similarity submatrix obtained after the calculation of the corresponding global prototype. The similarity submatrix is ​​calculated as follows: in, Indicates the first The characteristic sub-matrix of each channel and the global prototype The calculated similarity sub-matrix; Indicates the first The feature sub-matrix of each channel; Represents the feature submatrix A feature point in Indicates The first of the ProtoPNet units A global prototype; Represents a decimal; Representing feature points With global prototype The square of the Euclidean distance between them; Perform maximum pooling on each similarity matrix, and concatenate all similarity matrices after pooling to obtain the target similarity matrix.

5. The method according to claim 4, characterized in that The preset training rule is to perform two-stage training on the initial model in the following manner: The first stage of training uses time series blood pressure data as input and the disease classification results of the time series blood pressure data as prediction output. Multiple rounds of iterative training are performed and the initial model parameters are updated with the first objective function to obtain an intermediate model. The second stage of training uses time-series blood pressure data as input and the disease classification results of the time-series blood pressure data as prediction output. Multiple rounds of iterative training are performed and the intermediate model parameters are updated with the second objective function until convergence to obtain a medical task prediction model. In each round of iterative training, each global prototype of each ProtoPNet unit in the multi-channel prototype network module of the intermediate model is updated according to a preset matching method.

6. The method according to claim 5, characterized in that The first objective function is: in, represents the cross entropy loss function, Represents the number of time-series blood pressure data, represents the disease labels of time-series blood pressure data, Indicates Disease labels for time series blood pressure data, represents the predicted output of time series blood pressure data, Indicates Prediction output of time series blood pressure data; The second objective function is: in, in, in, represents the second objective function, represents the multi-channel loss, represents the number of ProtoPNet units, Indicates The weight parameters of the ProtoPNet units, Indicates ProtoPNet units, Indicates The first loss of the ProtoPNet unit, Indicates The second loss of the ProtoPNet unit, Indicates The third loss of the ProtoPNet unit, Indicated in ProtoPNet units are selected from A global prototype with consistent disease label categories for time-series blood pressure data , Represents the feature submatrix Belong to The feature matrix of time series blood pressure data, Represents the first The feature matrix of time series blood pressure data, Indicated in ProtoPNet units are selected from A global prototype for inconsistent disease label categories in time-series blood pressure data , Represents the feature submatrix A feature point in Representing feature points With global prototype The square of the Euclidean distance between Global prototype With global prototype The square of the Euclidean distance between Indicates The first of the ProtoPNet units global prototype, global prototype With global prototype Corresponding to different disease label categories.

7. The method according to claim 6, characterized in that The preset matching method is: in, Representing feature points The first The feature point set in the feature sub-matrix of the channel , Represents the first The first time series of blood pressure data The feature sub-matrix of the channels, Indicates The disease label categories of the time series blood pressure data are classified as follows: , Indicates The disease classification in ProtoPNet units is The global prototype collection of .

8. A medical task prediction method, characterized in that: The method comprises: Step T1, obtaining time series blood pressure data to be processed; Step T2, using the medical task prediction model constructed according to any of the methods described in claims 1-7 to perform prediction processing on the time series blood pressure data to be processed to obtain a disease classification result, wherein during the prediction processing, multiple target similarity matrices obtained by the medical task prediction model are used as a benchmark, and each target similarity matrix is ​​matched with the feature matrix of the time series blood pressure data to be processed to obtain the local data fragments of the channel in the time series blood pressure data corresponding to each element in each target similarity matrix, and one or more local data fragments that meet the preset requirements are selected from all the obtained local data fragments as prediction indicators to explain the disease classification results obtained by the prediction processing.

9. A computer-readable storage medium, characterized in that: A computer program is stored thereon, and the computer program can be executed by a processor to implement the steps of any method described in claims 1-8.

10. An electronic device, characterized in that: include: one or more processors, and A memory, wherein the memory is used to store executable instructions; The one or more processors are configured to implement the steps of any one of the methods of claims 1-8 by executing the executable instructions.

Images