Stroke prognosis warning system, method and storage medium based on deep learning

Through a deep learning-based stroke prognosis warning system, identifying and calculating factors affecting disease prognosis, the inaccurate risk assessment caused by relying on the subjective experience of doctors in the existing technology is solved, and more efficient prognosis assessment and risk warning are achieved.

CN117976218BActive Publication Date: 2024-11-15ANHUI PROVINCIAL HOSPITAL
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410202892.7
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-02-23
Publication Date
2024-11-15
Estimated Expiration
2044-02-23

AI Technical Summary

Technical Problem

At this stage, the prognosis assessment of stroke patients mainly relies on the subjective experience of doctors, and the lack of accurate risk warning models has led to the inability to promptly early warning assessment of the patient's prognosis risk.

Method used

A stroke prognosis warning system based on deep learning is adopted to obtain medical source data, preprocess and construct clinical data service models, establish and train prognosis prediction models, identify factors affecting disease prognosis and calculate impact values, and provide scientific decision-making basis.

Benefits of technology

It improves the accuracy and accuracy of prognostic evaluation, reduces errors, can promptly warn and evaluate patients' risks, and provides more scientific decision-making support.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN117976218B_ABST
    Figure CN117976218B_ABST
Patent Text Reader

Abstract

The present invention discloses a stroke prognosis warning system, method and storage medium based on deep learning, which solves the problem that the current prognosis assessment of stroke patients mainly relies on the subjective experience of doctors and lacks an accurate risk warning model for prognosis assessment, resulting in the inability to timely warn and assess the risk of patient prognosis. The method includes: acquiring medical source data, preprocessing the medical source data based on a pre-built metadata model, extracting the medical source data with a clinical data service model and constructing a disease database for a single disease; the prognosis prediction model analyzes and identifies the medical data source, identifies the factors affecting the disease prognosis, and calculates the influence values ​​of the influencing factors; the present invention can analyze and identify the medical data source by constructing a prognosis prediction model, identifies the factors affecting the disease prognosis, and calculates the influence values ​​of the influencing factors, thereby helping medical personnel understand and predict the trends and results of patient prognosis based on a large amount of medical data sources, and providing a scientific basis for decision-making.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the field of data processing technology, and specifically relates to a stroke prognosis warning system, method and storage medium based on deep learning. Background Art

[0002] With the rapid development of medical information technology and biomedical technology, the types and scale of medical data are growing at an unprecedented rate. The continuous development of precision medicine has made medicine truly enter the era of big data. While posing a huge challenge to traditional data processing technology, it has also created conditions for the development of related big data technologies.

[0003] Stroke is an important disease that endangers human health. In my country, it is the first leading cause of disability and the second leading cause of death. Every year, there are 3 million new stroke patients and 1.2 million deaths. With the aging of the population and the improvement of people's living standards, it has gradually shown a trend of high morbidity, high mortality, high disability rate and high economic burden. It seriously affects the quality of life of patients and imposes a huge economic burden on families and society. At present, the most ideal treatment for acute ischemic stroke is to reopen the occluded blood vessels as soon as possible and strive to reverse more ischemic penumbras, thereby saving lives and restoring lost neurological functions. Although endovascular treatment is considered to be an effective method for treating acute ischemic stroke, some patients still do not obtain good functional prognosis. Therefore, it is extremely important to find out which patients can benefit from endovascular treatment. At present, the prognosis assessment of stroke patients mainly relies on the subjective experience of doctors, and there is a lack of accurate risk warning models for prognosis assessment, which leads to the inability to timely warn and assess the risk of patient prognosis. In order to solve the above problems, we propose a stroke prognosis warning system, method and storage medium based on deep learning. Summary of the invention

[0004] The purpose of the present invention is to address the shortcomings of the existing technology and provide a stroke prognosis warning system, method and storage medium based on deep learning, which solves the problem that the current prognosis assessment of stroke patients mainly relies on the subjective experience of doctors and lacks an accurate risk warning model for prognosis assessment, resulting in the inability to timely warn and assess the risk of patient prognosis.

[0005] At present, the prognosis assessment of stroke patients mainly relies on the subjective experience of doctors. There is a lack of accurate risk warning models for prognosis assessment, and the risk of death and disability of patients cannot be clearly determined. In order to solve the above problems, we propose a stroke prognosis warning system, method and storage medium based on deep learning. In short, the warning method includes obtaining medical source data, preprocessing the medical source data based on a pre-built metadata model, building a clinical data service model based on the pre-processed medical source data, extracting the medical source data and building a disease database for a single disease, building and training a prognosis prediction model based on the disease database, extracting medical data sources from the disease database, and analyzing and identifying the prognosis factors of the disease by the prognosis prediction model. The embodiment of the present invention can analyze and identify the medical data source by constructing a prognosis prediction model, identify the prognosis factors of the disease, and calculate the influence value of the influencing factors, thereby helping medical personnel understand and predict the prognosis trend and results of patients based on a large number of medical data sources, and providing a scientific basis for decision-making. Moreover, the evaluation and prediction results of the prognosis prediction model constructed in this embodiment have higher accuracy and smaller errors than the existing technology, which can solve the problem that the current prognosis evaluation of stroke patients mainly relies on the subjective experience of doctors and lacks an accurate risk warning model for prognosis evaluation, resulting in the inability to timely warn and evaluate the risk of patient prognosis.

[0006] The present invention is implemented in this way: a stroke prognosis warning method based on deep learning, the stroke prognosis warning method based on deep learning comprises:

[0007] Acquire medical source data and pre-process the medical source data based on a pre-built metadata model, wherein the medical source data includes basic patient information, patient prognosis information, patient follow-up information, and multi-source heterogeneous medical system data;

[0008] Load medical source data and build a clinical data service model based on the preprocessed medical source data. The clinical data service model extracts the medical source data and builds a disease database for a single disease.

[0009] A prognosis prediction model is established and trained based on the disease database, and medical data sources are extracted from the disease database. The prognosis prediction model analyzes and identifies the medical data sources, identifies factors affecting disease prognosis, and calculates the impact values ​​of the factors.

[0010] Preferably, the stroke prognosis warning method based on deep learning further includes:

[0011] Obtain the impact value of the influencing factor and determine whether the impact value of the influencing factor exceeds the preset impact threshold. If it exceeds the impact threshold, trigger the early warning protocol, generate an early warning instruction, and upload the early warning instruction to the disease database.

[0012] Preferably, the method for constructing the metadata model specifically includes:

[0013] Traverse the medical source data, take the medical source data as input, build a corpus, divide the medical source data into blocks based on the data block algorithm to form multiple groups of source data blocks, and perform hash calculation on the multiple groups of source data blocks;

[0014] A sampling algorithm is used to extract multiple hash element values ​​from the hash calculation value as the characteristic values ​​of the medical source data;

[0015] Integrate the feature values ​​and input them into the corpus to update the corpus;

[0016] A mixup initial model is constructed based on deep learning, wherein the mixup initial model is a neural network model including at least one group of neural network regressors.

[0017] Preferably, the method for constructing the metadata model further includes:

[0018] The mixup initial model sets hyperparameters and extracts training and validation sets from the corpus based on the hyperparameters.

[0019] Take the training set as input, execute the mixup initial model, and obtain the mel feature spectrum of the training set respectively;

[0020] The training set and mel feature spectrum are respectively input into the neural network regressor of the mixup initial model. The neural network regressor calculates the predicted value of the training set based on the LOSS function and determines whether the predicted value exceeds the preset prediction threshold. If it exceeds, the LOSS function is used to reversely update the hyperparameters of the mixup initial model.

[0021] Load the validation set, use the validation set as input, execute the mixup initial model after adjusting the parameters, determine the recognition accuracy of the output value of the mixup initial model, and output the mixup initial model as the metadata model if the recognition accuracy exceeds the preset accuracy threshold.

[0022] Preferably, the method for constructing a clinical data service model based on preprocessed medical source data specifically includes:

[0023] Obtain pre-processed medical source data;

[0024] Retrieve the pre-established corpus and mixup initial model, and use the gradient descent method to update the mixup initial model;

[0025] Load the medical source data and divide the preprocessed medical source data into a training set and a validation set, where the ratio of the training set to the validation set is 3:1;

[0026] Take a certain characteristic attribute value of the data group in the training set as input, and continuously input characteristic Gaussian noise into the mixup initial model in the multi-step forward diffusion to form multiple groups of noise characteristic values;

[0027] The neural network regressor in the mixup initial model is used to embed the noise eigenvalues ​​to obtain the embedding transformation vector;

[0028] The gradient descent method is used to adjust the neural network parameters of the mixup initial model until convergence;

[0029] Obtain the validation set, use the validation set as input, perform n rounds of iterative training on the mixup initial model, and determine if the classification accuracy and association accuracy of the clinical data service model are greater than the preset classification threshold and association threshold. The training of the mixup initial model is completed, and the clinical data service model is output.

[0030] Preferably, the method of extracting medical source data and constructing a disease database for a single disease in the clinical data service model specifically includes:

[0031] The medical source data is traversed, and the clinical data service model identifies the medical source data to obtain the first classification degree of the medical source data;

[0032] A data stream classification algorithm based on frequent closed itemsets is used to calculate a second classification degree of the first classification degree;

[0033] Obtain the first classification degree and the second classification degree, and calculate the association weight of the medical source data based on the high-dimensional data stream clustering algorithm of Top-K frequent patterns;

[0034] The clinical data service model randomly generates unique indexes and index links that are greater than the associated weight based on the empi technology, integrates the unique indexes and index links, and adds the unique indexes and index links to the disease database to complete the update and construction of the disease database.

[0035] Preferably, the method for establishing and training a prognosis prediction model based on the disease database specifically includes:

[0036] Traverse the disease database, generate a prognosis data set, and process the prognosis data set;

[0037] The prognosis dataset was divided into a training set and a validation set with a ratio of 7:3, where the training set was used to train the model and the validation set was used to evaluate the model;

[0038] The model is constructed using the MLP neural network module, which can adjust the number of network layers and the number of neurons in each layer;

[0039] Load the completed prognosis prediction model and conduct model evaluation on the prognosis prediction model;

[0040] The validation set was used to verify the model performance, monitor the loss and accuracy of the prognostic prediction model, and the hyperparameters of the prognostic prediction model were adjusted based on the validation set verification results.

[0041] Preferably, the prognosis prediction model analyzes and identifies the medical data source, identifies the factors affecting the disease prognosis, and calculates the impact value of the factors, specifically including:

[0042] A medical data source is obtained, and information is extracted from the obtained medical data source, wherein information is extracted from the data source based on an information extraction function of a prognosis prediction model, and the information extraction function is:

[0043]

[0044] Among them, J x is the information extraction value of the information extraction function, λ is the extraction function curvature of the information extraction function, x is the summation upper limit, and is the variable of the information extraction function, f(x n ) to enter a value for the data source, and Enter the variance value for the data source;

[0045]

[0046] Among them, x n is a data set containing n sets of data sources, and λ is the extraction function curvature of the information extraction function;

[0047]

[0048] Performing dimension reduction feature decomposition on the information extraction value of the information extraction function to obtain a feature decomposition value;

[0049]

[0050] Among them, J τ x is the eigendecomposition value, and x n is a data set containing n groups of data sources;

[0051] Based on the eigendecomposition value, the largest k groups of disease prognosis influencing factors are obtained, and the influence values ​​of the influencing factors are calculated.

[0052]

[0053] Among them, J τ xis the eigendecomposition value, α is the pre-coefficient of the influencing factor extraction function, β is the post-coefficient of the influencing factor extraction function, and the pre-coefficient and post-coefficient are both constants. x Output values ​​for influencing factors;

[0054]

[0055] Among them, Y x Output value for influencing factors, is the influence value of the influencing factor, α is the pre-coefficient of the influencing factor extraction function, and β is the post-coefficient of the influencing factor extraction function.

[0056] Preferably, the method of loading the prognosis prediction model after modeling and evaluating the prognosis prediction model specifically includes:

[0057] Calculate evaluation indicators through confusion matrix;

[0058] Evaluate whether the prognostic prediction model correctly classifies the data Accuracy;

[0059] Accuracy=(TP+TN) / (TP+TN+FP+FN)

[0060] Evaluate the accuracy of the prediction model, where the accuracy is calculated by the ratio of data evaluated as positive to real positive data.

[0061] Precision=TP / (TP+FP)

[0062] Calculate the recall rate of the model. The recall rate of the model represents the proportion of all positive data that are evaluated as positive.

[0063] Recall=TP / (TP+FN)

[0064] Calculate the comprehensive evaluation index F1 of the classification effect of the prognostic prediction model;

[0065] F1=2TP / (2TP+FP+FN)

[0066] Among them, the receiver operating characteristic curve is an indicator for evaluating the predictive value, with the X-axis representing the false positive rate FPR and the Y-axis representing the true positive rate TPR;

[0067] FPR = FP / (FP + TN)

[0068] TPR = TP / (TP+FN).

[0069] On the other hand, an embodiment of the present invention further provides a stroke prognosis warning system based on deep learning, wherein the stroke prognosis warning system based on deep learning specifically comprises:

[0070] A source data acquisition module, the source data acquisition module is used to acquire medical source data and pre-process the medical source data based on a pre-built metadata model, wherein the medical source data includes basic patient information, patient prognosis information, patient follow-up information, and multi-source heterogeneous medical system data;

[0071] A database construction module, wherein the database construction module is used to load medical source data, and to construct a clinical data service model based on the preprocessed medical source data, and the clinical data service model extracts the medical source data and constructs a disease database for a single disease;

[0072] An influence value calculation module, wherein the influence value calculation module establishes and trains a prognosis prediction model based on the disease database, extracts medical data sources from the disease database, and the prognosis prediction model analyzes and identifies the medical data sources, identifies factors affecting disease prognosis, and calculates the influence values ​​of the factors;

[0073] The impact warning module is used to obtain the impact value of the influencing factor and determine whether the impact value of the influencing factor exceeds the preset impact threshold. If it exceeds the impact threshold, the warning protocol is triggered, a warning instruction is generated, and the warning instruction is uploaded to the disease database.

[0074] On the other hand, an embodiment of the present invention further provides a computer-readable storage medium storing computer program instructions, which, when executed, implement the stroke prognosis warning method based on deep learning.

[0075] Compared with the prior art, the embodiments of the present application have the following beneficial effects:

[0076] The embodiment of the present invention can analyze and identify the medical data source by constructing a prognosis prediction model, identify the factors affecting the prognosis of the disease, and calculate the influence value of the influencing factors, thereby helping medical personnel understand and predict the trend and results of the patient's prognosis based on a large number of medical data sources, and providing a scientific basis for decision-making. In addition, the evaluation prediction results of the prognosis prediction model constructed in this embodiment have higher accuracy and smaller errors than the existing technology, which can solve the problem that the current prognosis evaluation of stroke patients mainly relies on the subjective experience of doctors and lacks an accurate risk warning model for prognosis evaluation, resulting in the problem that the risk of patient prognosis cannot be timely warned and evaluated. The prognosis prediction model in the embodiment of the present invention establishes the neural network input, output, network connection mode, number of network layers and number of nodes in each layer through the forward propagation algorithm during construction and training, and calls the loss function to evaluate the difference between the predicted value and the true value, and uses the validation set to verify the model performance of the prognosis prediction model, monitor the loss and accuracy, so that the constructed prognosis prediction model has higher accuracy and smaller errors than the existing technology.

[0077] The embodiment of the present invention can pre-process the collected medical source data through the pre-built metadata model, thereby realizing automatic cleaning and data labeling of the medical source data, and normalizing and extracting a large amount of semi-structured and unstructured data in the medical data source, facilitating the construction of a disease database, thereby providing services for the prognosis prediction model to evaluate the factors affecting disease prognosis and calculate the influence values ​​of the influencing factors. When constructing the metadata model, a mixup initial model is constructed based on deep learning, and the LOSS function is used to reversely update the hyperparameters of the mixup initial model, thereby obtaining a metadata model for efficiently processing medical source data, providing auxiliary support for the disease database and the factors affecting disease prognosis, and calculating the influence values ​​of the influencing factors, effectively reducing the operating load of the prognosis prediction model. BRIEF DESCRIPTION OF THE DRAWINGS

[0078] Figure 1 It is a schematic diagram of the implementation flow of the stroke prognosis warning method based on deep learning provided by the present invention.

[0079] Figure 2 It is a schematic diagram of the implementation process of the metadata model construction method provided by the present invention.

[0080] Figure 3 It is a schematic diagram of the flow of the artificial intelligence follow-up speech for stroke recurrence in patients provided by the present invention.

[0081] Figure 4 It is a schematic diagram of the mRS scoring rules for stroke recurrence in patients through intelligent follow-up provided by the present invention.

[0082] Figure 5 It is a schematic diagram of the implementation flow of the method for constructing a clinical data service model based on preprocessed medical source data provided by the present invention.

[0083] Figure 6 It is a schematic diagram of the implementation process of the clinical data service model provided by the present invention for extracting medical source data and constructing a disease database for a single disease.

[0084] Figure 7 It is a schematic diagram of the implementation flow of the method for establishing and training a prognosis prediction model based on the disease database provided by the present invention.

[0085] Figure 8 This is a diagram of an operation interface for processing a prognosis data set provided by the present invention.

[0086] Fig. 9 It is a structural schematic diagram of the MLP neural network module provided by the present invention.

[0087] Fig.10 It is a schematic diagram of the functional interface of the disease database provided by the present invention.

[0088] Fig.11 It is a schematic diagram of the implementation process of building a model using the MLP neural network module provided by the present invention.

[0089] Fig.12 It is a structural schematic diagram of the stroke prognosis warning system based on deep learning provided by the present invention.

[0090] Fig.13 It is a schematic diagram of the implementation flow of the prognosis prediction model provided by the present invention for analyzing and identifying medical data sources, identifying factors affecting disease prognosis, and calculating the influence values ​​of the influencing factors. DETAILED DESCRIPTION

[0091] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by technicians in the technical field of this application; the terms used in the specification of the application herein are only for the purpose of describing specific embodiments and are not intended to limit this application; the terms "including" and "having" and any variations thereof in the specification and claims of this application and the above-mentioned drawings are intended to cover non-exclusive inclusions. The terms "first", "second", etc. in the specification and claims of this application or the above-mentioned drawings are used to distinguish different objects, not to describe a specific order.

[0092] Reference to "embodiments" herein means that a particular feature, structure, or characteristic described in conjunction with the embodiments may be included in at least one embodiment of the present application. The appearance of the phrase in various locations in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment that is mutually exclusive with other embodiments. It is explicitly and implicitly understood by those skilled in the art that the embodiments described herein may be combined with other embodiments.

[0093] At present, the prognosis assessment of stroke patients mainly relies on the subjective experience of doctors. There is a lack of accurate risk warning models for prognosis assessment, and the risk of death and disability of patients cannot be clearly determined. In order to solve the above problems, we propose a stroke prognosis warning system, method and storage medium based on deep learning. In short, the warning method includes obtaining medical source data, preprocessing the medical source data based on a pre-built metadata model, building a clinical data service model based on the pre-processed medical source data, extracting the medical source data and building a disease database for a single disease, building and training a prognosis prediction model based on the disease database, extracting medical data sources from the disease database, and analyzing and identifying the prognosis factors of the disease by the prognosis prediction model. The embodiment of the present invention can analyze and identify the medical data source by constructing a prognosis prediction model, identify the prognosis factors of the disease, and calculate the influence value of the influencing factors, thereby helping medical personnel understand and predict the prognosis trend and results of patients based on a large number of medical data sources, and providing a scientific basis for decision-making. Moreover, the evaluation and prediction results of the prognosis prediction model constructed in this embodiment have higher accuracy and smaller errors than the existing technology, which can solve the problem that the current prognosis evaluation of stroke patients mainly relies on the subjective experience of doctors and lacks an accurate risk warning model for prognosis evaluation, resulting in the inability to timely warn and evaluate the risk of patient prognosis.

[0094] The embodiment of the present invention provides a stroke prognosis warning method based on deep learning. Figure 1 The figure shows a schematic diagram of the implementation process of the stroke prognosis warning method based on deep learning, and the stroke prognosis warning method based on deep learning specifically includes:

[0095] Step S10, obtaining medical source data, and preprocessing the medical source data based on a pre-built metadata model, wherein the medical source data includes patient basic information, patient prognosis information, patient follow-up information, and multi-source heterogeneous medical system data.

[0096] It should be noted that medical source data includes but is not limited to patient basic information, patient prognosis information, patient follow-up information and multi-source heterogeneous medical system data, and the format of medical source data can be CSV, excel, R, SPSS, state, SAS, CDISC ODM standard data, and can also be database, text file, web-josh, XML, tabular data, image, audio, video and other multi-source heterogeneous data.

[0097] Step S20, loading medical source data, building a clinical data service model based on the pre-processed medical source data, the clinical data service model extracts the medical source data and builds a disease database for a single disease. Fig.10 A schematic diagram of the disease database functional interface is shown.

[0098] In this embodiment, the disease database of a single disease type can use a search tool with customized query conditions to perform full-text search on all clinical information while ensuring patient privacy. The disease database supports simple search, combined / advanced search, and navigation search by ICD-10 disease classification. It can perform secondary search on the search results, export to Excel, and save and archive them to the scientific research project management system. When performing full-text search on the database, the disease database supports a comprehensive fuzzy keyword search function for semantic-based patient medical records, generates a keyword index for the medical record data collected in the disease database, and provides a unified interface. It uses keyword fuzzy search similar to Google or Baidu to conduct a comprehensive search on the data on the data platform, and displays it through the patient comprehensive data browsing module. The structured editor of the disease database supports custom extraction of structured fields.

[0099] Step S30, establishing and training a prognosis prediction model based on the disease database, extracting medical data sources from the disease database, and analyzing and identifying the medical data sources by the prognosis prediction model, identifying factors affecting disease prognosis, and calculating the impact values ​​of the factors.

[0100] Step S40, obtaining the influence value of the influence factor, and determining whether the influence value of the influence factor exceeds a preset influence threshold;

[0101] Step S50: If the impact threshold is exceeded, the early warning protocol is triggered, an early warning instruction is generated, and the early warning instruction is uploaded to the disease database;

[0102] Step S60: If the impact threshold is not exceeded, the early warning protocol is not triggered, and the impact value of the influencing factor is stored in the disease database.

[0103] It should be noted that the impact threshold can be set to 0.8, and the warning protocol supports 5G or DTU communication protocol. Azure Monitor alert rules can be used to select the operation group when the alert is triggered. Once the conditions are met, the system will automatically send an early warning notification to the disease database.

[0104] The embodiment of the present invention can pre-process the collected medical source data through the pre-built metadata model, thereby realizing automatic cleaning and data labeling of the medical source data, and normalizing and extracting a large amount of semi-structured and unstructured data in the medical data source, facilitating the construction of a disease database, thereby providing services for the prognosis prediction model to evaluate the factors affecting disease prognosis and calculate the influence values ​​of the influencing factors. When constructing the metadata model, a mixup initial model is constructed based on deep learning, and the LOSS function is used to reversely update the hyperparameters of the mixup initial model, thereby obtaining a metadata model for efficiently processing medical source data, providing auxiliary support for the disease database and the factors affecting disease prognosis, and calculating the influence values ​​of the influencing factors, effectively reducing the operating load of the prognosis prediction model.

[0105] At the same time, the embodiment of the present invention can analyze and identify medical data sources by constructing a prognosis prediction model, identify factors affecting disease prognosis, and calculate the impact values ​​of the factors, thereby helping medical personnel understand and predict the trends and results of patient prognosis based on a large number of medical data sources, and providing a scientific basis for decision-making. In addition, the evaluation and prediction results of the prognosis prediction model constructed in this embodiment have higher accuracy and smaller errors than the existing technology, which can solve the problem that the current prognosis evaluation of stroke patients mainly relies on the subjective experience of doctors and lacks an accurate risk warning model for prognosis evaluation, resulting in the inability to timely warn and evaluate the risk of patient prognosis.

[0106] The embodiment of the present invention provides a method for constructing a metadata model. Figure 2 The schematic diagram of the implementation process of the metadata model construction method is shown. The metadata model construction method specifically includes:

[0107] Step S101, traversing medical source data, taking the medical source data as input, constructing a corpus, dividing the medical source data into blocks based on a data block algorithm to form multiple groups of source data blocks, and performing hash calculations on the multiple groups of source data blocks;

[0108] In this embodiment, considering that the existing corpus is insufficient and the quality of the corpus text data is uneven, it is necessary to study the metadata model that minimizes the dependence on the corpus. The metadata model identifies the characteristics of the source data from three perspectives: characters, words, and word boundary information, and accurately and quickly identifies text sentences, voices, and videos, thereby extracting feature vectors into the corpus to provide auxiliary support for the construction of the disease library.

[0109] It should be noted that the medical source data can specifically be patient-related data information, such as age, height, weight, body mass index, past medical history, personal history, and can also be patient postoperative condition data, such as patient postoperative imaging, changes in NIHSS scores 24 hours and 7 days after surgery, changes in venous blood indicators, and other systemic complications.

[0110] In this embodiment, the medical source data mainly collects the demographic information, medical history, important hematological indicators, degree of neurological deficit at admission, stroke etiology classification, National Institute of Heath Stroke Scale (NIHSS) score at admission, modified Rankin scale (mRS) score before illness, occlusion site, collateral circulation status and symptom information of all patients: onset time, disease fluctuation time, arrival time, detailed record of clinical symptoms and changes of patients, etc.

[0111] The imaging data of the patients can be collected for Digital Substraction Angiography (DSA), cranial CT, cranial MRI, head and neck vascular imaging, and head and neck MRI angiography. The Alberta Stroke Program Early CTScore (ASPECTS) score was recorded at admission.

[0112] In this embodiment, when the prognosis data source is obtained, the intelligent follow-up can be used to simulate the interaction between the doctor and the patient to achieve human-computer dialogue, such as Figure 3 As shown, the AI ​​follow-up technique for stroke recurrence in patients was shown. A plan was developed to perform mRS scoring based on 90-day, half-year, 1-year, 2-year, 3-year, and 5-year follow-ups. Figure 4 The following table shows the details of the stroke recurrence mRS score for intelligent follow-up patients to evaluate the prognosis and efficacy of patients. Primary endpoint event: Primary efficacy endpoint event: 90-day mRS score. Primary safety endpoint event: 90-day mortality rate. After the follow-up, data statistical analysis and report generation are performed based on the collected resident intentions and key slot information.

[0113] Step S102, using a sampling algorithm to extract multiple hash element values ​​from the hash calculation value as feature values ​​of the medical source data;

[0114] Step S103, integrating the feature values ​​and inputting the feature values ​​into the corpus to update the corpus;

[0115] Step S104, constructing a mixup initial model based on deep learning, wherein the mixup initial model is a neural network model including at least one group of neural network regressors, and the mixup initial model is an initial model of the metadata model.

[0116] In this embodiment, the mixup initial model is a neural network model including at least one set of neural network regressors. The mixup initial model can be an ASR model or a hidden Markov model based on deep learning. It should be noted that the mixup initial model is the starting point of the metadata model, and the metadata model is the mixup initial model obtained after multiple iterations, adjustments and optimizations. In this process, the mixup initial model undergoes multiple iterations, adjustments and optimizations to ensure that the metadata model can accurately capture the underlying structure and regularity of the data, thereby preprocessing the medical source data.

[0117] Step S105, setting hyperparameters for the mixup initial model, and extracting a training set and a validation set from the corpus based on the hyperparameters;

[0118] In this embodiment, the ratio of the training set to the validation set may be 3:1, and 50 times of 10-fold cross validation are used to test the accuracy of the algorithm and avoid prediction errors when extracting the training set and the validation set.

[0119] Step S106, using the training set as input, executing the mixup initial model, and obtaining the mel feature spectrum of the training set respectively;

[0120] Step S107, respectively input the training set and the mel feature spectrum into the neural network regressor of the mixup initial model, the neural network regressor calculates the predicted value of the training set based on the LOSS function, and determines whether the predicted value exceeds the preset prediction threshold. If it exceeds, the LOSS function is used to reversely update the hyperparameters of the mixup initial model;

[0121] It should be noted that the LOSS function includes but is not limited to CTC loss, CE loss, and LabelSmooth loss functions.

[0122] Step S108, load the validation set, use the validation set as input, execute the mixup initial model after adjusting the parameters, determine the recognition accuracy of the output value of the mixup initial model, and output the mixup initial model as a metadata model if the recognition accuracy exceeds a preset accuracy threshold. In this embodiment, the output value recognition accuracy can be 0.9-0.95.

[0123] The embodiment of the present invention provides a method for constructing a clinical data service model based on preprocessed medical source data. Figure 5 A schematic diagram of the implementation process of a method for constructing a clinical data service model based on preprocessed medical source data is shown. The method for constructing a clinical data service model based on preprocessed medical source data specifically includes:

[0124] Step S201, obtaining pre-processed medical source data;

[0125] Step S202, retrieve the pre-established corpus and the mixup initial model, and update the mixup initial model using the gradient descent method;

[0126] Step S203, loading medical source data, and dividing the preprocessed medical source data into a training set and a validation set, wherein the ratio of the training set to the validation set is 3:1;

[0127] Step S204, taking a certain characteristic attribute value of the data group in the training set as input, and continuously inputting characteristic Gaussian noise into the mixup initial model in the multi-step forward diffusion performed in sequence, so as to form multiple groups of noise characteristic values;

[0128] Step S205, using the neural network regressor in the mixup initial model to perform embedding transformation on the noise eigenvalue to obtain an embedding transformation vector;

[0129] Step S206, using the gradient descent method to adjust the neural network parameters of the mixup initial model until convergence;

[0130] Step S207, obtain the validation set, use the validation set as input, perform n rounds of iterative training on the mixup initial model, and determine if the classification accuracy and association accuracy of the clinical data service model are greater than the preset classification threshold and association threshold. Then the training of the mixup initial model is completed and the clinical data service model is output.

[0131] In this embodiment, when it is judged that the classification accuracy and association accuracy of the clinical data service model are greater than the preset classification threshold and association threshold, after n rounds of iterative training of the mixup initial model, the classification accuracy and association accuracy of the clinical data service model can be evaluated by ROC curve analysis and Hosmer-Lemeshow test, and the preset classification threshold and association threshold can be set to 0.85 and 0.9.

[0132] The embodiment of the present invention provides a method for extracting medical source data from a clinical data service model and constructing a disease database for a single disease. Figure 6 The schematic diagram of the implementation process of the method for extracting medical source data and constructing a disease database for a single disease by the clinical data service model is shown. The method for extracting medical source data and constructing a disease database for a single disease by the clinical data service model specifically includes:

[0133] Step S301, traversing the medical source data, the clinical data service model identifies the medical source data, and obtains the first classification degree of the medical source data;

[0134] Step S302, using a data stream classification algorithm based on frequent closed itemsets to calculate a second classification degree of the first classification degree;

[0135] Step S303, obtaining the first classification degree and the second classification degree, and calculating the association weight of the medical source data based on the high-dimensional data stream clustering algorithm of Top-K frequent patterns;

[0136] Step S304, the clinical data service model randomly generates a unique index and index link that is greater than the associated weight based on the empi technology, integrates the unique index and index link, and adds the unique index and index link to the disease database to complete the update and construction of the disease database.

[0137] In this embodiment, on the basis of constructing a general clinical data service model, it also supports the construction of a data model for specific single disease characteristics. Then, by setting the entry and exit conditions and the data screening timer, the data of the data center is regularly extracted into the single disease database.

[0138] The embodiment of the present invention provides a method for establishing and training a prognosis prediction model based on the disease database. Figure 7 The schematic diagram of the implementation flow of the method for establishing and training a prognosis prediction model based on the disease database is shown. The method for establishing and training a prognosis prediction model based on the disease database specifically includes:

[0139] Step S401, traverse the disease database, generate a prognosis data set, and process the prognosis data set.

[0140] In this embodiment, RStudio4.0.1 software and Python3.7.0 software can be used for data processing. When preprocessing data, deletion, replacement or interpolation can be considered to handle missing data values. When preprocessing features, the following three data standardization methods are included:

[0141] 1) Min-Max standardization, max is the maximum value, min is the minimum value, the formula is as follows:

[0142] x'=(x-min) / (max-min)

[0143] 2) Normalization: linear transformation of data is performed to map the final result to the range of [0,1], thus achieving geometric scaling of the original data. The formula is as follows:

[0144]

[0145] 3) Discrete data such as strings cannot be used directly, so they are preprocessed using OneHotEncoding. Continuous features are processed using Z-Score standardization, where μ is the mean and σ is the standard deviation.

[0146] x'=(x-μ) / σ

[0147] like Figure 8 As shown, an operation interface diagram for processing a prognosis data set is shown. At the same time, the preprocessing of the prognosis data set also includes data normalization (selecting a numeric type field for normalization operation), standardization (selecting a numeric type field for standardization operation) and discretization (selecting a numeric type field for discretization operation) processing.

[0148] Step S402, dividing the prognosis data set into a training set and a validation set, and splitting them in a ratio of 7:3, wherein the training set is used to train the model, and the validation set is used to evaluate the model, and 10 times of 10-fold cross validation is used when sampling the training set and the validation set to test the accuracy of the algorithm and avoid prediction errors;

[0149] Step S403, using the MLP neural network module to build a model, the MLP neural network module can adjust the number of network layers and the number of neurons in each layer.

[0150] In this embodiment, after the prognosis data set is acquired and segmented, a disease risk warning model needs to be further constructed. Machine learning plays an important role in this. It is an analysis method derived from artificial intelligence and statistics that can effectively process high-dimensional data. It builds a statistical model through a data set, relies on the self-learning ability of the machine, and obtains the potential laws of the data, thereby improving the efficiency of processing data. Carrying out corresponding research in the medical field is conducive to the progress of translational medicine and precision medicine, and has an important role in promoting clinical decision-making. The algorithms of traditional machine learning include logistic regression, naive Bayes, Bayesian networks, decision trees, random forest algorithms, etc., whose features need to be determined by professionals, and the optimization algorithm relies on the accuracy of features, which is time-consuming and labor-intensive. Considering the defects of traditional machine learning methods, deep learning is increasingly applied to the medical field and exceeds traditional machine learning in many medical tasks. It uses multiple hidden layer neural network models, learns the intrinsic high-order features of the data through a large number of vector calculations, and uses these feature decision-making processes to complete complex tasks. Deep learning is end-to-end learning, which can automatically extract high-level features, greatly reducing the feature engineering that traditional machine learning relies on. In addition, deep learning has the advantages of strong learning ability, wide coverage, and good adaptability. Deep neural networks are the basis of deep learning. Artificial neural networks with multiple hidden layers are used to perform complex pattern recognition and data classification tasks, including a variety of network structures, such as convolutional neural networks, recurrent neural networks, and multi-layer perceptrons (Multi-Layer Perceptron, MLP). MLP, also known as multi-layer neural networks, is the basic algorithm of deep neural networks. Perceptrons are single neuron models. The power of neural networks lies in their ability to learn the representations in training data and how to link them with the output variables they want to predict. They can learn any mapping function and have been proven to be a universal approximation algorithm. MLP has shown its superiority over traditional machine learning algorithms in basic emotional facial expression classification and measuring the degree of obsessive-compulsive disorder. Multi-layer perceptrons are similar to single-layer neural networks and are fully connected between layers. When there is an output, it can be compared with a known label and the weights are adjusted accordingly. This process is repeated until the calculation process reaches the maximum number of allowed iterations or an acceptable error rate. Its main features are: the neuron nodes between adjacent layers are fully connected; the neuron nodes in the same layer are not connected; each connection has a weight. Neural network layers can be divided into three layers: input layer, hidden layer and output layer. Generally speaking, the first layer is the input layer, the last layer is the output layer, and the layers in between are all hidden layers, and the layers are fully connected, such as Fig. 9 As shown, a structural schematic diagram of the MLP neural network module is shown.

[0151] At the same time, it should be noted that Fig.11As shown, a method for constructing a model using an MLP neural network module is shown. The method for constructing a model using an MLP neural network module specifically includes:

[0152] Step S4031, determine the input layer and output layer, add multiple hidden layers, each hidden layer has a certain number of neurons, in this embodiment, the neurons can be 100-1000 groups. Select the appropriate number of layers and neurons according to the complexity of the problem and the amount of data, where the number of nodes in the input layer is the number of features; the number of nodes in the output layer is the number of categories.

[0153] Step S4032, using the He or Xavier initialization method to set initial values ​​for the weights of the prognosis prediction model, input neurons 1, 2, 3, 4, ..., x n , a total of x n Neuron input, weighted sum of neurons, get weighted sum value w 1 ,w 2 ,w 3 ,…,w n , calculate the weights of each neuron, perform function operation on the weighted sum, and get the weighted sum vector. The formula is:

[0154]

[0155] Step S4033, load the weighted sum vector, and output the variable after processing by the activation function σ, where the activation function is:

[0156] o=σ(z)=σ(WX+b);

[0157] Step S4034, establishing the neural network input, output, network connection mode, number of network layers and number of nodes in each layer through the forward propagation algorithm;

[0158] Among them, input: input vector x, network layer number L, weight matrix W in the neural network, offset vector b; output: σ;

[0159] Step S4035, call the loss function to evaluate the difference between the predicted value and the true value. When evaluating the difference between the predicted value and the true value, the binary cross-entropy loss is selected for the binary classification problem, and the formula is as follows:

[0160]

[0161] For classification problems, the classification variable cross entropy loss (Cross-Entropy Loss) is used, and the formula is as follows:

[0162] CE=-∑y_true*log(y_pred)

[0163] Also select the optimizer: Stochastic Gradient Descent (SGD) or Adam.

[0164] Step S404, loading the completed prognosis prediction model and performing model evaluation on the prognosis prediction model;

[0165] Step S405 , using the validation set to validate the model performance of the prognosis prediction model, monitoring loss and accuracy, and adjusting the hyperparameters of the prognosis prediction model based on the validation set validation results.

[0166] It should be noted that when adjusting the hyperparameters of the prognosis prediction model based on the validation results of the validation set, hyperparameters refer to fixed parameters related to the neural network model architecture, which can be manually adjusted before or during training. Grid optimization is used to find the optimal parameters and obtain the best model. Hyperparameters include optimization parameters and network parameters, such as learning rate, learning rate decay coefficient, parameters of different optimizers, etc.; network parameters, such as network depth, number of neurons, number of batches, etc.

[0167] The basic hyperparameters of the prognostic prediction model are set as hidden layer: 3 layers; nodes: 34, 58, 60; maximum number of iterations: 100 times; learning function: back propagation algorithm; hidden layer activation function: "Act_Logistic" function; output layer activation function: "Act_Logistic" function; decay: 1e-7, and when adjusting the hyperparameters, the model performance can be adjusted by trying different hyperparameter combinations, such as learning rate, batch size, number and size of hidden layers, activation function, etc. In order to improve the model performance, the best hyperparameter combination is found through grid search to improve the model performance.

[0168] In this embodiment, the method of loading the prognosis prediction model after modeling and evaluating the prognosis prediction model specifically includes:

[0169] The evaluation index is calculated by the confusion matrix, where Table 1 shows the composition structure of the confusion matrix;

[0170] Table 1 Confusion matrix composition structure

[0171]

[0172] Evaluate whether the prognostic prediction model correctly classifies the data Accuracy;

[0173] Accuracy=(TP+TN) / (TP+TN+FP+FN)

[0174] Evaluate the accuracy of the prediction model, where the accuracy is calculated by the ratio of data evaluated as positive to real positive data.

[0175] Precision=TP / (TP+FP)

[0176] Calculate the recall rate of the model. The recall rate of the model represents the proportion of all positive data that are evaluated as positive.

[0177] Recall=TP / (TP+FN)

[0178] Calculate the comprehensive evaluation index F1 of the classification effect of the prognostic prediction model;

[0179] F1=2TP / (2TP+FP+FN)

[0180] Among them, the receiver operating characteristic curve is an indicator for evaluating the predictive value, with the X-axis representing the false positive rate FPR and the Y-axis representing the true positive rate TPR;

[0181] FPR = FP / (FP + TN)

[0182] TPR=TP / (TP+FN)

[0183] When training the model, the generalization ability of the model must also be verified. The model is trained using training data. After each epoch (a complete traversal of the training data), the model performance is evaluated using validation data. Indicators such as loss and accuracy are monitored to determine whether the model is overfitting or underfitting. During the training process, the learning rate and other hyperparameters are adjusted to optimize model performance. The model is evaluated on the validation set: Model performance is regularly evaluated on the validation set to check for overfitting and select the best model.

[0184] The embodiment of the present invention provides a method for analyzing and identifying medical data sources by a prognosis prediction model, identifying factors affecting disease prognosis, and calculating the impact values ​​of the factors. Fig.13 The following is a schematic diagram showing the implementation process of the method for analyzing and identifying the medical data source, identifying the factors affecting the prognosis of the disease, and calculating the influence value of the influencing factors by the prognosis prediction model. The method for analyzing and identifying the medical data source, identifying the factors affecting the prognosis of the disease, and calculating the influence value of the influencing factors by the prognosis prediction model specifically includes:

[0185] Step S401, obtaining a medical data source, and extracting information from the obtained medical data source, wherein information is extracted from the data source based on an information extraction function of a prognosis prediction model, and the information extraction function is:

[0186]

[0187] Among them, Jx is the information extraction value of the information extraction function, and λ is the extraction function curvature of the information extraction function. In this embodiment, the extraction function curvature can be 0.1-0.3, x is the summation upper limit, and is the variable of the information extraction function. f(x n ) to enter a value for the data source, and Enter the variance value for the data source;

[0188]

[0189] Among them, x n is a data set containing n sets of data sources, and λ is the extraction function curvature of the information extraction function;

[0190]

[0191] Step S402, performing dimensionality reduction feature decomposition on the information extraction value of the information extraction function to obtain a feature decomposition value;

[0192]

[0193] Among them, J τ x is the eigendecomposition value, and x n is a data set containing n groups of data sources;

[0194] Step S403, obtaining the largest k groups of disease prognosis influencing factors based on the eigendecomposition values, and calculating the influence values ​​of the influencing factors.

[0195]

[0196] Among them, J τ x is the characteristic decomposition value, α is the pre-coefficient of the influencing factor extraction function, β is the post-coefficient of the influencing factor extraction function, the pre-coefficient and the post-coefficient are both constants. In this embodiment, the pre-coefficient and the post-coefficient can be in the range of 1-5, Y x Output values ​​for influencing factors;

[0197]

[0198] Among them, Y x is the output value of the influencing factor, is the influence value of the influencing factor, α is the pre-coefficient of the influencing factor extraction function, and β is the post-coefficient of the influencing factor extraction function.

[0199] The embodiment of the present invention can analyze and identify the medical data source by constructing a prognosis prediction model, identify the factors affecting the prognosis of the disease, and calculate the influence value of the influencing factors, thereby helping medical personnel understand and predict the trend and results of the patient's prognosis based on a large number of medical data sources, and providing a scientific basis for decision-making. In addition, the evaluation prediction results of the prognosis prediction model constructed in this embodiment have higher accuracy and smaller errors than the existing technology, which can solve the problem that the prognosis evaluation of stroke patients at this stage mainly relies on the subjective experience of doctors, lacks an accurate risk warning model for prognosis evaluation, and leads to the problem that the risk of patient prognosis cannot be timely warned and evaluated. At the same time, the prognosis prediction model in the embodiment of the present invention establishes the neural network input, output, network connection mode, number of network layers and number of nodes in each layer through the forward propagation algorithm during construction and training, and calls the loss function to evaluate the difference between the predicted value and the true value, and uses the validation set to verify the model performance of the prognosis prediction model, monitor the loss and accuracy, so that the constructed prognosis prediction model has higher accuracy and smaller errors than the existing technology.

[0200] On the other hand, the embodiment of the present invention provides a stroke prognosis warning system based on deep learning. Fig.12 The schematic diagram of the structure of the stroke prognosis warning system based on deep learning is shown. The stroke prognosis warning system based on deep learning specifically includes:

[0201] A source data acquisition module 100, the source data acquisition module 100 is used to acquire medical source data and pre-process the medical source data based on a pre-built metadata model, wherein the medical source data includes basic patient information, patient prognosis information, patient follow-up information, and multi-source heterogeneous medical system data;

[0202] A database construction module 200, wherein the database construction module 200 is used to load medical source data, and to construct a clinical data service model based on the preprocessed medical source data, and the clinical data service model extracts the medical source data and constructs a disease database for a single disease;

[0203] An influence value calculation module 300 is used to establish and train a prognosis prediction model based on the disease database, extract medical data sources from the disease database, analyze and identify the medical data sources, identify factors affecting disease prognosis, and calculate the influence values ​​of the factors;

[0204] The impact warning module 400 is used to obtain the impact value of the impact factor, determine whether the impact value of the impact factor exceeds the preset impact threshold, and if it exceeds the impact threshold, trigger the warning protocol, generate a warning instruction, and upload the warning instruction to the disease database.

[0205] In this embodiment, the source data acquisition module 100, the database construction module 200, the impact value calculation module 300 and the impact warning module 400 can use local area network or 5G communication to achieve data interaction. The embodiment of the present invention can pre-process the collected medical source data through a pre-built metadata model, thereby realizing automatic cleaning and data labeling of the medical source data, and normalizing and extracting a large amount of semi-structured and unstructured data in the medical data source, which facilitates the construction of a disease database. The prognosis prediction model constructed in this application integrates the scientific research needs and clinical experience of frontline clinical experts, and standardizes the construction of a data platform and tools for data collection, collation and sharing by clinical researchers, which can efficiently solve the problem of integrated processing of massive data in clinical research.

[0206] On the other hand, an embodiment of the present invention provides a computer device, which includes a display screen, a memory, a processor, and a computer program, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the deep learning-based stroke prognosis warning method.

[0207] It is understandable that, in the preferred embodiment provided by the present invention, the computer device may also be a notebook computer, a personal digital assistant (PDA), a mobile phone or other device capable of communication.

[0208] On the other hand, an embodiment of the present invention further provides a readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the processor executes the steps of the stroke prognosis warning method based on deep learning.

[0209] Exemplarily, the computer program may be divided into one or more modules, one or more modules are stored in a memory and executed by a processor to complete the present invention. One or more modules may be a series of computer program instruction segments capable of completing specific functions, and the instruction segments are used to describe the execution process of the computer program in a terminal device. For example, the above-mentioned computer program may be divided into units or modules of the berth status display system provided by each of the above-mentioned system embodiments.

[0210] The processor may be a central processing unit (CPU), other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor, etc. The processor is the control center of the terminal device, and uses various interfaces and lines to connect various parts of the entire user terminal.

[0211] The above-mentioned memory can be used to store computer programs and / or modules, and the above-mentioned processor realizes various functions of the above-mentioned terminal device by running or executing computer programs and / or modules stored in the memory, and calling data stored in the memory. The memory may mainly include a program storage area and a data storage area, wherein the program storage area may store an operating system, an application required for at least one function; the data storage area may store data created according to the use of a stroke prognosis warning system based on deep learning, etc. In addition, the memory may include a high-speed random access memory, and may also include a non-volatile memory, such as a hard disk, a memory, a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (Flash Card), at least one disk storage device, a flash memory device, or other volatile solid-state storage devices.

[0212] Finally, it should be noted that the computer-readable storage medium (e.g., memory) herein may be a volatile memory or a nonvolatile memory, or may include both a volatile memory and a nonvolatile memory. As an example and not by way of limitation, a nonvolatile memory may include a read-only memory (ROM), a programmable ROM (PROM), an electrically programmable ROM (EPROM), an electrically erasable programmable ROM (EEPROM), or a flash memory. A volatile memory may include a random access memory (RAM), which may act as an external cache memory. As an example and not by way of limitation, RAM may be obtained in a variety of forms, such as synchronous RAM (DRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), and direct Rambus RAM (DRRAM). The storage devices of the disclosed aspects are intended to include, but are not limited to, these and other suitable types of memory.

[0213] In summary, the present invention provides a stroke prognosis warning system, method and storage medium based on deep learning. The embodiments of the present invention can analyze and identify medical data sources by constructing a prognosis prediction model, identify factors affecting disease prognosis, and calculate the influence values ​​of the factors, thereby helping medical personnel understand and predict the trends and results of patient prognosis based on a large number of medical data sources, and providing a scientific basis for decision-making. In addition, the evaluation and prediction results of the prognosis prediction model constructed in this embodiment have higher accuracy and smaller errors than the existing technology, which can solve the problem that the current prognosis evaluation of stroke patients mainly relies on the subjective experience of doctors and lacks an accurate risk warning model for prognosis evaluation, resulting in the inability to timely warn and evaluate the risk of patient prognosis.

[0214] Through the prognosis prediction model in the embodiment of the present invention, prevention and treatment strategies can be formulated in advance for high-risk patients, such as more frequent examinations, earlier interventions or more intensive treatments; patients can be classified to provide personalized treatment and care; resources can be allocated more effectively, and beds and care for high-risk patients can be given priority; high-risk groups can be identified at an early stage, providing a basis for precision treatment, which has important economic and social value.

[0215] It should be noted that, for the above-mentioned embodiments, for the sake of simplicity, they are all described as a series of action combinations, but those skilled in the art should know that the present invention is not limited by the described order of actions, because according to the present invention, some steps may be performed in other orders or simultaneously. Secondly, those skilled in the art should also know that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily required by the present invention.

[0216] In the several embodiments provided in the present application, it should be understood that the disclosed device can be implemented in other ways. For example, the device embodiments described above are only schematic, such as the division of the above-mentioned units, which is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or communication connection shown or discussed can be through some interfaces, and the indirect coupling or communication connection between devices or units can be in the form of telecommunication or other forms.

[0217] The units described above as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0218] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the scope of protection of the invention. Obviously, the described embodiments are only some embodiments of the present invention, rather than all embodiments. Based on these embodiments, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in this field can still combine, add, delete or make other adjustments to the features in the various embodiments of the present invention according to the circumstances without conflict, without making creative work, so as to obtain different other technical solutions that do not deviate from the concept of the present invention in essence, and these technical solutions also belong to the scope of protection of the present invention.

Claims

1. A stroke prognosis warning method based on deep learning, characterized in that: The stroke prognosis warning method based on deep learning includes: Acquire medical source data and pre-process the medical source data based on a pre-built metadata model, wherein the medical source data includes basic patient information, patient prognosis information, patient follow-up information, and multi-source heterogeneous medical system data; Load medical source data and build a clinical data service model based on the preprocessed medical source data. The clinical data service model extracts the medical source data and builds a disease database for a single disease. A prognosis prediction model is established and trained based on the disease database, and medical data sources are extracted from the disease database. The prognosis prediction model analyzes and identifies the medical data sources, identifies factors affecting disease prognosis, and calculates the impact values ​​of the factors. Obtain the impact value of the influencing factor, and determine whether the impact value of the influencing factor exceeds the preset impact threshold. If it exceeds the impact threshold, trigger the early warning protocol, generate an early warning instruction, and upload the early warning instruction to the disease database; The method for constructing the metadata model specifically includes: Traverse the medical source data, take the medical source data as input, build a corpus, divide the medical source data into blocks based on the data block algorithm to form multiple groups of source data blocks, and perform hash calculation on the multiple groups of source data blocks; A sampling algorithm is used to extract multiple hash element values ​​from the hash calculation value as the characteristic values ​​of the medical source data; Integrate the feature values ​​and input them into the corpus to update the corpus; Building a mixup initial model based on deep learning, wherein the mixup initial model is a neural network model including at least one set of neural network regressors, and the mixup initial model is an initial model of the metadata model; The mixup initial model sets hyperparameters and extracts training and validation sets from the corpus based on the hyperparameters. Take the training set as input, execute the mixup initial model, and obtain the mel feature spectrum of the training set respectively; The training set and mel feature spectrum are respectively input into the neural network regressor of the mixup initial model. The neural network regressor calculates the predicted value of the training set based on the LOSS function and determines whether the predicted value exceeds the preset prediction threshold. If it exceeds, the LOSS function is used to reversely update the hyperparameters of the mixup initial model. Load the validation set, use the validation set as input, execute the mixup initial model after adjusting the parameters, and judge the recognition accuracy of the output value of the mixup initial model. If the recognition accuracy exceeds the preset accuracy threshold, output the mixup initial model as the metadata model. The method for constructing a clinical data service model based on preprocessed medical source data specifically includes: Obtain pre-processed medical source data; Retrieve the pre-established corpus and mixup initial model, and use the gradient descent method to update the mixup initial model; Load the medical source data and divide the preprocessed medical source data into a training set and a validation set, where the ratio of the training set to the validation set is 3:1; Take a certain characteristic attribute value of the data group in the training set as input, and continuously input characteristic Gaussian noise into the mixup initial model in the multi-step forward diffusion to form multiple groups of noise characteristic values; The neural network regressor in the mixup initial model is used to embed the noise eigenvalues ​​to obtain the embedding transformation vector; The gradient descent method is used to adjust the neural network parameters of the mixup initial model until convergence; Obtain the validation set, use the validation set as input, perform n rounds of iterative training on the mixup initial model, and determine if the classification accuracy and association accuracy of the clinical data service model are greater than the preset classification threshold and association threshold. The training of the mixup initial model is completed, and the clinical data service model is output; The method of extracting medical source data and constructing a disease database for a single disease in the clinical data service model specifically includes: The medical source data is traversed, and the clinical data service model identifies the medical source data to obtain the first classification degree of the medical source data; A data stream classification algorithm based on frequent closed itemsets is used to calculate a second classification degree of the first classification degree; Obtain the first classification degree and the second classification degree, and calculate the association weight of the medical source data based on the high-dimensional data stream clustering algorithm of Top-K frequent patterns; The clinical data service model randomly generates unique indexes and index links that are greater than the associated weight based on the empi technology, integrates the unique indexes and index links, and adds the unique indexes and index links to the disease database to complete the update and construction of the disease database; The method for establishing and training a prognosis prediction model based on the disease database specifically includes: Traverse the disease database, generate a prognosis data set, and process the prognosis data set; The prognosis dataset was divided into a training set and a validation set with a ratio of 7:3, where the training set was used to train the model and the validation set was used to evaluate the model; The model is constructed using the MLP neural network module, which can adjust the number of network layers and the number of neurons in each layer; Loading the completed prognosis prediction model, and performing model evaluation on the prognosis prediction model; The validation set was used to verify the model performance, monitor the loss and accuracy of the prognostic prediction model, and the hyperparameters of the prognostic prediction model were adjusted based on the validation set verification results; The method of analyzing and identifying the medical data source by the prognosis prediction model, identifying the factors affecting the disease prognosis, and calculating the impact value of the factors specifically includes: A medical data source is obtained, and information is extracted from the obtained medical data source, wherein information is extracted from the data source based on an information extraction function of a prognosis prediction model, and the information extraction function is: Among them, J x is the information extraction value of the information extraction function, λ is the extraction function curvature of the information extraction function, x is the summation upper limit, and is the variable of the information extraction function, f(x n ) to enter a value for the data source, and Enter the variance value for the data source; Among them, x n is a data set containing n sets of data sources, and λ is the extraction function curvature of the information extraction function; Performing dimension reduction feature decomposition on the information extraction value of the information extraction function to obtain a feature decomposition value; Among them, J τ x is the eigendecomposition value, and x n is a data set containing n groups of data sources; Find the largest k groups of disease prognosis influencing factors based on the eigendecomposition value, and calculate the influence value of the influencing factors; The method of obtaining the largest k groups of disease prognosis influencing factors based on the eigendecomposition values ​​and calculating the influence values ​​of the influencing factors specifically includes: Among them, J τ x is the eigendecomposition value, α is the pre-coefficient of the influencing factor extraction function, β is the post-coefficient of the influencing factor extraction function, and the pre-coefficient and post-coefficient are both constants. x Output values ​​for influencing factors; Among them, Y x is the output value of the influencing factor, is the influence value of the influencing factor, α is the pre-coefficient of the influencing factor extraction function, and β is the post-coefficient of the influencing factor extraction function.

2. A stroke prognosis warning system based on deep learning, implemented by the stroke prognosis warning method based on deep learning as claimed in claim 1, characterized in that: The stroke prognosis warning system based on deep learning specifically includes: A source data acquisition module, the source data acquisition module is used to acquire medical source data and pre-process the medical source data based on a pre-built metadata model, wherein the medical source data includes basic patient information, patient prognosis information, patient follow-up information, and multi-source heterogeneous medical system data; A database construction module, wherein the database construction module is used to load medical source data, and to construct a clinical data service model based on the preprocessed medical source data, and the clinical data service model extracts the medical source data and constructs a disease database for a single disease; An influence value calculation module, wherein the influence value calculation module establishes and trains a prognosis prediction model based on the disease database, extracts medical data sources from the disease database, and the prognosis prediction model analyzes and identifies the medical data sources, identifies factors affecting disease prognosis, and calculates the influence values ​​of the factors; The impact warning module is used to obtain the impact value of the influencing factor and determine whether the impact value of the influencing factor exceeds the preset impact threshold. If it exceeds the impact threshold, the warning protocol is triggered, a warning instruction is generated, and the warning instruction is uploaded to the disease database.

3. A computer-readable storage medium, characterized in that: Computer program instructions are stored, and when the computer program instructions are executed, the stroke prognosis warning method based on deep learning as claimed in claim 1 is implemented.