Structural causal model construction method and device for acute myocardial infarction monitoring

Abstract

The invention relates to a structural causal model construction method and device for acute myocardial infarction monitoring, belongs to the technical field of patient data processing, and solves the problems that patient data cannot be monitored in real time and predicted data is inaccurate in the prior art. The construction method of the structural causal model comprises the following steps: step 1, collecting and processing historical medical big data to obtain a structured data set; 2, discretizing the structured data set according to medical knowledge; 3, constructing an AMI causal structure based on medical constraint conditions and a causal discovery method according to the discretized data set; and 4, based on the discretized data set in the step 2, performing parameter learning on the AMI causal structure by using a Bayesian parameter estimation method to obtain a final AMI structure causal model. The method can more accurately quantify the features and identify the significant influence factors of the AMI through the effective information values, and assists the prevention, diagnosis and treatment of the AMI.

Description

Technical Field

[0001] The present invention belongs to the technical field of patient data processing, and in particular relates to a method and device for constructing a structural causal model for acute myocardial infarction monitoring. Background Art

[0002] With the rapid development of medical means, disease monitoring and evaluation models based on medical informationization have been vigorously developed, and the data of corresponding diseases are processed and predicted and warned in advance based on the models, such as: Chinese patents CN109949936B and CN110957036B.

[0003] Acute myocardial infarction (AMI) is a disease with a high rate of critical illness and mortality, which is often acute and has severe symptoms. It is very important for the life and health of patients to make diagnosis, early warning and treatment of AMI before it occurs and develops seriously. However, among AMI patients, some patients have no obvious early clinical symptoms and patient data monitoring is not timely, which leads to many patients neglecting health management in daily life and may miss the best time to treat the disease. Summary of the invention

[0004] In view of the above problems, the present invention provides a method and device for constructing a structural causal model for acute myocardial infarction monitoring, which solves the problems in the prior art of being unable to monitor patient data in real time and inaccurate predicted data.

[0005] The present invention provides a method for constructing a structural causal model for acute myocardial infarction monitoring, and the specific steps are as follows: Step 1: Collect and process historical medical big data to obtain a structured dataset for training the AMI structural causal model; Among them, the steps for processing historical medical big data are: Determine the rules for extracting information from historical medical big data; Structural data organization of historical medical big data according to information extraction rules; Identify feature information with regular expressions through regular expressions; Use natural language processing methods to process feature information in historical medical big data that does not have regular expressions; Finally, a structured data set is obtained; Step 2: discretize the structured data set according to medical knowledge to obtain a discretized data set; Step 3: According to the discretized data set, the AMI causal structure is constructed based on medical constraints and causal discovery methods; the variables in the AMI causal structure are AMI, associated factors that cause AMI, and manifestation factors caused by AMI, and the directed edges are the generation process from cause to effect between variables; Step 4: Based on the discretized data set in step 2, use the Bayesian parameter estimation method to learn the parameters of the AMI causal structure and estimate the conditional probability parameters; based on the conditional probability parameters, the final AMI structural causal model is obtained.

[0006] Optionally, historical medical big data includes medical record information in real hospital scenarios; medical record information includes the patient's medical history information, the doctor's description of the patient's condition, and the diagnosis results based on the patient's actual situation; medical record information includes the patient's biochemical test index information, daily living habits and past medical history; the doctor's description of the patient's condition includes the patient's oral description and the doctor's records.

[0007] Optionally, the variables include AMI incidence information variables and corresponding characteristic variables of the AMI incidence information variables.

[0008] Optionally, medical prior knowledge is introduced into the AMI causal structure.

[0009] Optionally, an additive noise model and / or medical knowledge is employed to update the AMI causal structure.

[0010] Optionally, the causal discovery method is a fast causal inference algorithm based on conditional independence or a fast greedy equivalence class search algorithm based on scores.

[0011] Another aspect of the present invention discloses an intelligent wearable device for diagnosing acute myocardial infarction for monitoring acute myocardial infarction, comprising a signal acquisition device, a signal processing device and a display processing unit; the signal processing device is provided with an AMI structural causal model, and the AMI structural causal model is constructed using the aforementioned structural causal model construction method.

[0012] Compared with the prior art, the present invention has at least the following beneficial effects: (1) The structural causal model for acute myocardial infarction monitoring of the present invention can realize the intelligent monitoring of AMI, provide an AMI intelligent data processing and monitoring method with interpretability and generalization ability, improve the accuracy of prediction results, and establish a systematic understanding of AMI as a complex disease, which can provide guidance for the prevention and treatment of AMI. Based on the structural causal model, the effective information value can be used to more accurately quantify the characteristics and identify the significant influencing factors of AMI, assisting the prevention, diagnosis and treatment of AMI. More importantly, the structural causal model can also be used to predict the risk of individual morbidity in the short term.

[0013] (2) The wearable device of the present invention, in view of the rapid onset and high risk of AMI, can effectively improve the device user's ability to predict AMI data through real-time detection in daily life, thereby allowing the user to seek medical treatment earlier and reduce the risk of complications and severe illness after recovery.

[0014] (3) The wearable device of the present invention can collect individual user data in real time, and realize real-time monitoring and early warning of the risk of acute myocardial infarction.

[0015] (4) The wearable device of the present invention can provide decision support for doctors, relieve the pressure of emergency visits, and improve the accuracy of clinical diagnosis of AMI. In addition, the device is also suitable for daily monitoring of the general population or patients with hypertension, so as to identify the risk of myocardial infarction in users and issue early warnings a few days in advance, realize real-time monitoring and evaluation of personal health, early screening and pre-diagnosis of diseases, so as to actively intervene in advance.

[0016] (5) The wearable device of the present invention is equipped with a physiological index regular monitoring program based on medical knowledge, so that the present invention can more accurately and early detect potential health problems, prompting the user to seek medical treatment as soon as possible, or suggesting that the user make adjustments to their lifestyle habits.

[0017] (6) The wearable device of the present invention has strong applicability and can be used for daily risk monitoring of other diseases, and only the corresponding medical data needs to be replaced. It can create a new field of health management, realize cloud diagnosis and treatment, save traditional medical resources, and open up new medical methods.

[0018] (7) The wearable device of the present invention can use the collected data of all users to conduct big data analysis and integration through a cloud server, incorporate it into the training of structural causal models, and apply it to the research and analysis of the epidemiological characteristics of myocardial infarction in the future, providing new treatment experience and scientific basis for disease prediction and prevention and control. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] The drawings are only for the purpose of illustrating particular embodiments and are not to be construed as limiting the invention.

[0020] Figure 1 The present invention is a flowchart of the method for constructing a structural causal model for acute myocardial infarction monitoring.

[0021] Figure 2 This is a schematic diagram of the wearable device for intelligent diagnosis of acute myocardial infarction for monitoring acute myocardial infarction according to the present invention.

[0022] Figure 3 This is a causal structure diagram of the AMI structure in the actual case of the present invention.

[0023] Figure 4It is a schematic diagram of the significance of the causal effect of relevant variables on AMI in the actual case of the present invention.

[0024] Figure 5 It is a schematic diagram of the significance of the causal impact of AMI on related variables in the actual case of the present invention. DETAILED DESCRIPTION

[0025] In order to more clearly understand the above-mentioned purpose, features and advantages of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments can be combined with each other. In addition, the present invention can also be implemented in other ways different from those described herein, and therefore, the protection scope of the present invention is not limited by the specific embodiments disclosed below.

[0026] A specific embodiment of the present invention, as Figure 1-Figure 5 , discloses an acute myocardial infarction intelligent diagnosis wearable device based on a structural causal model, including a signal acquisition device, a signal processing device and a display processing unit; The signal acquisition instruments include blood pressure measuring instruments, electrocardiogram measuring instruments, blood oxygen measuring instruments and pulse measuring instruments.

[0027] A blood pressure measuring instrument, used to collect the user's blood pressure signal in real time; An electrocardiogram measuring instrument is used to collect the user's electrocardiogram signals in real time; A blood oxygen meter, used to collect the user's blood oxygen saturation signal in real time; A pulse measuring instrument, used to collect the user's pulse signal in real time; Signal processor, which includes a big data analyzer and an AMI structural causal model, and is used to process blood pressure signals, electrocardiogram signals, blood oxygen saturation signals, pulse signals and manually input data to obtain causal data of acute myocardial infarction and AMI prediction risk prediction value; The display processing unit is used to manually input data, display historical monitoring results, display causal data of acute myocardial infarction, and display early warning risk prediction values. It realizes the visualization of physiological indicators, can dynamically observe their changing patterns, and can also trace the time period when abnormal values ​​occur.

[0028] Among them, the manually input data is the user's examination data in the hospital, living habits, medical history and physical condition and / or clinical manifestations.

[0029] Furthermore, an OCR (optical character recognition module) photo recognition module is provided for users to upload medical test report data.

[0030] Furthermore, a cloud server and a communication module are provided, and the signal processor is provided on the cloud server; the communication module is used for transmission of various measurement signals and causal data.

[0031] Preferably, the communication module is a 4G communication module.

[0032] Preferably, an alarm is also provided to issue an alarm according to the early warning risk prediction value; when the early warning risk prediction value is higher than the risk threshold, an alarm is issued; Furthermore, the alarm is a loudspeaker; further, when the alarm is issued, the alarm information is simultaneously uploaded to the cloud server.

[0033] The wearable device of the present invention can provide targeted suggestions through real-time early warning of the risk of AMI and with the help of the interpretability of the AMI structural causal model. After the device side (blood pressure meter, electrocardiogram meter, blood oxygen meter and pulse meter) and the display processing unit successfully obtain the user data, the data is forwarded to the signal processor (cloud server) by the API interface, and after storage and analysis, the current AMI risk prediction value is given. The wearable device is also provided with a suggestion library based on medical principles, which will return targeted suggestions to the wearable device based on the various measurement signals collected by the user, manually input data and AMI predicted risk prediction values, guide the user to seek medical treatment in time, screen and control the risk hazards before the onset of AMI, strangle the disease process in the early stage, and avoid the serious consequences of AMI attacks.

[0034] Furthermore, the big data analyzer is used to collect and process historical medical big data to obtain a structured data set for training the AMI structural causal model. The historical medical big data includes medical record information in real hospital scenarios.

[0035] Among them, medical record information includes the patient's medical history information, the doctor's description of the patient's condition and the diagnosis results based on the patient's actual situation; medical record information includes the patient's biochemical test index information, daily living habits and past medical history, etc.; the doctor's description of the patient's condition includes the patient's oral statement and the doctor's records.

[0036] Furthermore, the big data analysis module processes the historical medical big data. First, the information extraction rules of the historical medical big data are summarized, and structured data is sorted according to the information rules. Regular expressions are used to identify key information, and finally, natural language processing methods are used to process complex text information to obtain a structured data set. The specific processing steps are as follows: First, summarize the historical medical big data information extraction rules: For the natural language information of patients' oral statements and doctors' records in historical medical big data, we analyze their regular characteristics and formulate information extraction rules. The extraction rules include two aspects: 1) confirming the normal value and / or abnormal value division range of biochemical test indicators in the natural language information of patients' oral statements and doctors' records, as well as the standard value of each biochemical test indicator; 2) according to the terminology habits used by doctors to describe diseases in diagnosis, a type of disease may have multiple expressions in medicine. For example, abnormal heart function can be expressed as "heart function grade x (NYHA classification)", "heart function grade x (Killip) classification", "heart failure", etc. According to the terminology habits, the description rules of various diseases in the natural language information of patients' oral statements and doctors' records are confirmed.

[0037] Then, structured data is sorted according to information rules: for biochemical test index information with fixed rules and clear structure, structured data is sorted using the data processing method in python pandas; Regular expressions are used to identify key information: For information with fixed expressions and strong keyword positioning, regular expressions are used to identify keywords and obtain disease results, such as "chronic renal insufficiency" and "type 2 diabetes".

[0038] Finally, a structured data set was obtained by processing complex text information using natural language processing methods. For historical medical big data whose semantics cannot be identified by keywords alone and which do not have a fixed form of expression, natural language processing methods were used for processing and analysis. For example, if there is semantic information in the medical record description that cannot be directly identified by keywords, such as "the patient has a history of smoking" or "the patient does not smoke", the intelligent identification of fuzzy information is achieved through the natural language processing model.

[0039] Furthermore, the natural language inference task method in the natural language processing method is used for analysis and processing. The natural language inference task method can judge the relationship between the premise of the natural language statement and the given hypothesis by identifying the two. Taking the specific case as the premise, the hypothesis is set that the case has a certain feature, and whether the hypothesis is established is identified, thereby realizing the intelligent extraction of case features and realizing the intelligent processing of large amounts of text information.

[0040] Preferably, an advanced deep language processing discriminant model based on BERT is used to process natural language inference tasks, and is applied to the recognition and processing of medical texts with complex expressions (i.e., medical big data whose semantics cannot be recognized by keywords alone and there is no fixed form of expression). It has good sentence processing capabilities. In order to enhance its recognition ability for long texts, after adjusting the task application mode, the model can complete the recognition task based on the prompt words, and extract feature information of complex semantic expressions such as "smoking", "chest pain" and "drinking" from the medical record text with good accuracy.

[0041] Furthermore, the big data analysis module discretizes the structured data set based on medical knowledge.

[0042] For example, in clinical diagnosis, doctors usually make a diagnosis based on clinical symptoms, biochemical test indicators and electrocardiogram characteristics, comprehensively consider the individual characteristics of the patient, and finally judge whether the patient has AMI based on medical knowledge. Therefore, it is necessary to combine medical knowledge to judge the diagnostic items suitable for AMI diagnosis in the used data set. In addition, since the range of continuous data such as some biochemical test indicator information is large, which affects the calculation results, the present invention starts from the real pathophysiological mechanism, and the standard of discretization processing is formulated according to medical principles. With reference to the division basis of various indicator values ​​in the medical sense, the data information is discretized to obtain a structured data set containing AMI incidence information variables and their related factor characteristic variables (including demographics, living habits, past medical history and physical condition, clinical manifestations).

[0043] Among them, the AMI structural causal model is used to describe the complex causal relationship between AMI and AMI-related factors.

[0044] Furthermore, the establishment of AMI structural causal model introduces AMI causal structure.

[0045] Compared with other data-driven model methods, the AMI structural causal model method of the present invention emphasizes the causal logical relationship between variables rather than just considering correlation. For example, the causal relationship emphasizes the chronological logic of variables. For example, if variable A is the cause of variable B, then variable A occurs before variable B.

[0046] See also Figure 2 The AMI structural causal model is constructed based on a directed acyclic causal graph (i.e., AMI causal structure) to describe the causal relationship between variables in a variable set. The variables in the AMI causal structure are AMI, related factors that cause AMI, and manifestation factors caused by AMI. Directed edges represent the generation process from cause to effect between variables.

[0047] It is understandable that the variable entries may be AMI incidence information variables and their related characteristic variables (including demographics, lifestyle, past medical history, physical condition and clinical manifestations).

[0048] Furthermore, the AMI causal structure (i.e., directed acyclic causal graph) is constructed based on the causal discovery algorithm.

[0049] Furthermore, when using the causal discovery algorithm to construct the AMI causal structure, medical prior knowledge is introduced into the AMI causal structure, which avoids incorrect causality in advance and generally improves the accuracy of the structural causal model construction; Specifically, when using the causal discovery algorithm to construct the causal structure of AMI, the constraints of the causal discovery process are designed based on medical prior knowledge: The variable items are classified according to demographics, living habits, medical history, physical condition, and clinical manifestations. On this basis, constraints are designed according to medical prior knowledge. The specific constraints are: (1) It is prohibited to have variables other than demographics in the variable entry point to demographics or the causal relationship between demographics; (2) Based on medical knowledge, summarize the temporal order of variables such as demographics, lifestyle, medical history, physical condition, and clinical manifestations associated with the onset of acute myocardial infarction (AMI) in historical medical big data, and prohibit causal relationships that violate the temporal order; (3) The internal causal relationship between clinical manifestations such as chest pain and lung rales is prohibited.

[0050] Furthermore, combined with medical knowledge constraints, the AMI causal structure is constructed based on constraint conditions and causal discovery methods.

[0051] Specifically, the causal discovery method adopted by the present invention is a fast causal inference algorithm based on conditional independence or a fast greedy equivalence class search algorithm based on scores, which respectively mines the causal relationship between relevant factors from historical medical big data based on different rules and constructs the AMI causal structure; Among them, the fast causal inference algorithm examines causal relationships from historical medical big data based on the conditional independence rule and identifies the causal direction by identifying the V structure between variables; Specifically, for the identification of V-structure causal relationships, the expression is:

[0052] Among them, V represents V structure; A and B represent variables A and variable B respectively, C represents the common adjacent variable C of variables A and variable B; its causal meaning is that variables A and variable B are the causes of the common adjacent variable C respectively; its statistical characteristics are that, without giving the common adjacent variable C, variables A and variable B are independent, and given the common adjacent variable C, variables A and variable B are conditionally not independent.

[0053] Specifically, the fast causal inference algorithm determines whether the V structure exists by identifying the separation set. The outlier set is a set of variable combinations that can make two variables satisfy conditional independence, expressed as:

[0054] Among them, A and B represent variable A and variable B respectively; X represents a set of variables. Given the variable set X, variable A and variable B are conditionally independent. Their conditional independence can be judged by statistical test methods such as chi-square test, thereby determining the range of the separation set.

[0055] The fast causal inference algorithm starts from an undirected graph with fully connected variables, searches for a separation set between any two variables, deletes the causal direction between the two variables if a separation set exists, traverses all variables, and constructs a directed acyclic causal graph.

[0056] It can be understood that if the common adjacent variable C of two variables A and B does not belong to their separation set X, a V structure is determined. .

[0057] Based on this, the fast causal inference algorithm obtains the complete causal structure by deleting causal edges and determining the V structure.

[0058] Among them, the fast greedy equivalence class search algorithm obtains the optimal AMI causal structure based on the Bayesian information criterion and other scoring searches; Specifically, the fast greedy equivalence class search algorithm determines the causal direction by identifying the V-structure relationship between variables. The fast greedy equivalence class search algorithm starts from a causal graph that does not contain edges (that is, all variables are considered independent of each other) and identifies the optimal AMI causal structure by assigning a causal model score.

[0059] Specifically, a fast greedy equivalence class search algorithm can evaluate causal structure using scores such as the Bayesian Information Criterion; The expression of Bayesian information criterion scoring is:

[0060] in, Indicates the rating value; k The degrees of freedom that represent the causal structure (i.e., the number of parameters used to describe the statistical characteristics of the causal structure, which increases with the number of causal directions in the causal structure); N represents the sample size (i.e. the amount of data used to train the model); L represents the likelihood function; Indicates variables, Indicates The causal variable of each variable, represents the probability, which is calculated based on the statistical characteristics of the variable given the cause variable in the causal structure. n Indicates the total number of variables.

[0061] Specifically, the fast greedy equivalence class search algorithm is divided into two stages: the first stage selects the direction with the best score according to the direction score, and adds causal directions to the direction with the best score; the second stage identifies the causal directions that can improve the score of the causal graph model by removing them, thereby obtaining a complete AMI causal structure.

[0062] Furthermore, the additive noise model and medical knowledge are used to further refine the AMI causal structure and identify the causal directions that the above two types of causal discovery algorithms cannot identify, and the final AMI causal structure is obtained. This solves the problem that some causal directions in the structural causal model with the same edge relationship and collision structure cannot be identified due to the Markov equivalence class problem in the causal discovery algorithm.

[0063] The expression of the AMI causal structure of the additive noise model is:

[0064] in, is the cause variable entry, is the result variable entry, is the residual term.

[0065] It can be understood that the additive noise model confirms the causal direction by testing the conditional independence between the noise term and the independent variable. For example, if the causal variable term For AMI, set the result variable entry For "chest pain", we can identify the residual term and the cause variable entry The independence between variables can be used to determine the true causal direction, and the causal relationship between other variables can be determined by analogy.

[0066] Since the data used in the present invention are feature data discretized by medical knowledge, an improved additive noise model calculation method suitable for discrete variables is introduced. In addition, for the causal direction that the additive noise model cannot determine, medical knowledge is introduced to determine the causal relationship between variables, and finally a complete structural causal model structure is obtained.

[0067] Further, the parameters of the AMI causal structure are learned based on the above. The conditional probability parameters of the AMI structural causal model are estimated using the Bayesian parameter estimation method. , that is, to estimate the variables The parent node variable collection In the case of variables The conditional distribution parameters of are used to obtain the final AMI structural causal model.

[0068] Furthermore, based on the AMI structural causal model, according to various real-time measurement signals and manually input data, the conditional probability of AMI (i.e. ,in, Represents a given set of evidence, thereby achieving the inference of the probability of AMI incidence and obtaining causal data of acute myocardial infarction and AMI prediction risk prediction value.

[0069] The present invention evaluates the predictive performance of the structural causal model established based on the historical medical big data of AMI. Specifically, given the patient condition indicators of the historical medical big data, the probability of AMI is inferred based on the AMI structural causal model, and compared with the true value, and the F1 score, accuracy, recall rate, precision and other indicators of the AMI structural causal model prediction results are calculated. According to these indicators, the present invention is compared with the traditional data-driven method, and the predicted data obtained by the present invention is more accurate.

[0070] Furthermore, the wearable system for intelligent diagnosis of acute myocardial infarction also includes a real-time measurement data input and output module, which is used to collect data signals from blood pressure measuring devices, blood oxygen measuring devices and pulse detection devices in real time, and allows users to pre-set personal demographics, medical history, clinical symptoms and other information in the software, and output patient risk information.

[0071] Optionally, the real-time measurement data input and output module is a mobile phone APP.

[0072] Furthermore, the effective information value (EI) of the real-time measurement signals and manually input data (i.e., indicators) was calculated based on the AMI structural causal model, and the causal relationship significance of the real-time measurement signals and manually input data for the occurrence of AMI was ranked according to the effective information value. Combined with the feasibility of data monitoring and transmission, and the medical connection between the real-time measurement signals and manually input data and AMI symptoms, a comprehensive analysis and evaluation was conducted from three aspects, thereby preliminarily confirming the range of indicators that play a key role in AMI prediction.

[0073] When in use, the signal data measured in real time on the device side is transmitted to the device, and the pre-input information is combined with the actual measurement value. Based on the built-in AMI intelligent diagnosis model based on the structural causal model, the probability of AMI disease is calculated in real time, and the risk of disease is promptly prompted.

[0074] Specifically, since the user's blood pressure curve shows different performance within 24 hours, with dynamic variability and circadian rhythm, the blood pressure measuring instrument of the wearable device automatically measures blood pressure every 15 to 30 minutes.

[0075] Furthermore, during the day, the user keeps the wearable device at the same height as the heart; at night, the user lies flat to ensure normal data collection. It is reasonable and feasible to use the wearable device of the present invention for daily blood pressure monitoring. Too dense sampling intervals may affect the user's normal life and cause frequent compression and contraction of blood vessels, thereby causing measurement errors. Therefore, the present invention combines the actual situation of the blood pressure measuring instrument and the developed prototype to determine that the blood pressure timing monitoring interval of the blood pressure measuring instrument is every half hour within 24 hours, which is helpful for long-term and lasting monitoring of the user's blood pressure, thereby scientifically achieving disease prevention.

[0076] Furthermore, five minutes was selected as the time interval for the pulse meter (i.e., heart rate) to monitor the pulse rate regularly. The arithmetic mean of the heart rate within 24 hours was selected as the input of the model, which improved the accuracy of the results.

[0077] A blood pressure measuring instrument, used to collect the user's blood pressure signal in real time; An electrocardiogram measuring instrument is used to collect the user's electrocardiogram signals in real time; A blood oxygen meter, used to collect the user's blood oxygen saturation signal in real time; A pulse measuring instrument, used to collect the user's pulse signal in real time; Furthermore, when the wearable device is used at night, blood oxygen sampling every five minutes improves the accuracy of the results; further, the arithmetic mean of all blood oxygen saturation data at night eliminates random errors.

[0078] Based on this, the present invention can provide a relatively scientific and complete physiological index monitoring solution to better adapt the user's prediction function of the risk of acute myocardial infarction in daily situations. The specific contents are as follows: Blood pressure is measured every half an hour, heart rate is measured every five minutes, and blood oxygen saturation is measured every five minutes at night. All the above data are transmitted to the cloud platform by the device via 4G. After preprocessing the data on the cloud server, the daytime peak and nighttime peak values ​​of blood pressure are taken, and the arithmetic mean of all data collected in one day is taken for blood oxygen and heart rate. Finally, the processed data is input into the structural causal model.

[0079] In another embodiment of the present invention, the wearable device for intelligent diagnosis of acute myocardial infarction based on the structural causal model also includes a module for identifying significant factors of causal effects related to AMI, which is used to provide guidance for the prevention and treatment of AMI and establish a systematic understanding of AMI.

[0080] Furthermore, the significance of the causal relationship between the relevant factors in the AMI structural causal model and AMI is evaluated, so as to determine the factors with significant causal effects on AMI, and provide guidance for the front-end AMI risk warning and subsequent AMI clinical treatment. The effective information index is used to quantify the significance of the causal relationship between the factors in the structural causal model.

[0081] Specifically, the effective information index is used to quantify the set probability intervention and equal probability intervention of the AMI structural causal model, and the KL divergence between the two probability intervention results is obtained. The KL divergence is used to describe the significance of the causal effect, and the expression is:

[0082]

[0083]

[0084]

[0085] in, Representation variables A For variables the extent of the impact; Representation variables The total number of states in the state space of ; Representation variables The state space of status; Representation variables The total number of states in the state space of Representation variables The state space of status; Indicates intervention in the state of the variable; represents probability; represents an intervention state of equal probability.

[0086] For example, to calculate the effect of AMI on chest pain For example, AMI is represented by " A ”, chest pain is expressed as “ ”.

[0087] The proposed wearable device for intelligent diagnosis of acute myocardial infarction is applied through a real case. The historical medical record data of the emergency department of Beijing Jishuitan Hospital is selected, which contains the medical record information of 3521 patients. The information includes biochemical tests, past medical history, living habits, clinical diagnosis results and other dimensions.

[0088] Step 1: Processing and analysis of medical big data; The medical record dataset contains information on 3521 patients in total, of which the raw data of biochemical tests (part) are shown in Table 1.

[0089] Table 1 Raw data of biochemical tests (partial)

[0090] Information extraction was performed on the medical record dataset. The key-value pair matching, regular expression recognition, and natural language processing methods were used to extract biochemical test data, text information in fixed expression forms, and text information without fixed expression forms. Invalid entries with missing items, duplicate items, and duplicate items in the dataset were deleted, and finally structured data was obtained.

[0091] Based on the inductive items of medical knowledge, by investigating the medical background knowledge and analyzing the available information in the data, four categories of AMI-related characteristics that should be taken into consideration were identified, including demographics, lifestyle, medical history and physical condition, and clinical manifestations, with a total of 56 key items. Table 2 shows the important AMI-related items (part).

[0092] Table 2 Important items related to AMI (partial)

[0093] Furthermore, the data were discretized and some data discretization standards (partial) were developed based on medical knowledge. Some of the standards are shown in Table 3.

[0094] Table 3 Data discretization processing standards (partial)

[0095] A total of 3521 medical records described in digital form were obtained and can be used for calculation. The data processing results (partial) are shown in Table 4.

[0096] Table 4 Data processing results (partial)

[0097] Step 2: Establishment of structural causal model of AMI; Based on the data obtained in step 1, a structural causal model is constructed through a causal discovery algorithm. First, the medical causal relationship in each item is analyzed to determine the medical prior knowledge constraints, including: 1) prohibiting pointing to demographics and the causal relationship between them; 2) prohibiting the causal relationship between clinical manifestations; 3) constructing a model according to the causal order of demographics → lifestyle → past history and physical symptoms → AMI → clinical manifestations.

[0098] Next, a structural causal model is constructed based on a fast greedy equivalence class search algorithm.

[0099] Furthermore, based on the discrete additive noise model and the medical knowledge refined structural causal model structure, Table 5 shows the causal relationship (partial) identified based on the ANM model.

[0100] Table 5 Causal relationships identified based on the ANM model (partial)

[0101] Finally, the structure of the structural causal model is obtained as follows Figure 3 shown.

[0102] Finally, the structural causal model is parameterized based on the existing structural causal model structure, and the conditional probability is estimated based on the Bayesian method to obtain the final AMI structural causal model. The model performance is verified on this basis. By comparing the inferred AMI disease probability of the model with the true value, the performance evaluation results of the AMI structural causal model are calculated and shown in Table 6.

[0103] Table 6 Performance evaluation results of AMI structural causal model

[0104] Step 3: Identify the significant factors related to the causal effect of AMI; According to the above parameterized structural causal model, calculate , quantify the significance of causal effects between AMI-related factors. After calculation, 11 factors were found to be significantly affected by the causal effect of AMI, and 9 factors were found to have significant causal effects on the incidence of AMI, such as Figure 4 , Figure 5 shown.

[0105] Step 4: Implement intelligent diagnosis of AMI based on wearable devices; The above model is connected to smart wearable devices and corresponding mobile phone APPs, and realizes real-time intelligent diagnosis of AMI based on the blood pressure, heart rate, pulse and other information collected by the smart wearable devices on the signal acquisition platform, as well as other indicator information such as personal information, living habits, medical history, etc. manually entered by the user into the APP.

[0106] At the same time, the above-mentioned AMI system analysis results are integrated, and the significance of the causal effects of AMI-related factors is converted into "health tips" to remind users to develop good daily living habits and guide daily health management.

[0107] The detailed variable entries and corresponding abbreviations in the case are shown in Table 7.

[0108] Table 7 Variable entries and corresponding abbreviations

[0109] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by any technician familiar with the technical field within the technical scope disclosed by the present invention should be covered within the protection scope of the present invention.

Claims

1. A method for constructing a structural causal model for acute myocardial infarction monitoring, characterized in that: The specific steps are as follows: Step 1: Collect and process historical medical big data to obtain a structured dataset for training the AMI structural causal model; Among them, the steps for processing historical medical big data are: Determine the rules for extracting information from historical medical big data; Structural data organization of historical medical big data according to information extraction rules; Identify feature information with regular expressions through regular expressions; Use natural language processing methods to process feature information in historical medical big data that does not have regular expressions; Finally, a structured data set is obtained; Step 2: discretize the structured data set according to medical knowledge to obtain a discretized data set; Step 3: According to the discretized data set, the AMI causal structure is constructed based on medical constraints and causal discovery methods; the variables in the AMI causal structure are AMI, associated factors that cause AMI, and manifestation factors caused by AMI, and the directed edges are the generation process from cause to effect between variables; Step 4: Based on the discretized data set in step 2, use the Bayesian parameter estimation method to learn the parameters of the AMI causal structure and estimate the conditional probability parameters; based on the conditional probability parameters, the final AMI structural causal model is obtained.

2. The method for constructing a structural causal model according to claim 1, characterized in that: Historical medical big data includes medical record information in real hospital scenarios; medical record information includes the patient's medical history information, the doctor's description of the patient's condition, and the diagnosis results based on the patient's actual situation; medical record information includes the patient's biochemical test index information, daily living habits and past medical history; the doctor's description of the patient's condition includes the patient's oral description and the doctor's records.

3. The method for constructing a structural causal model according to claim 2, characterized in that: The variables include AMI incidence information variables and corresponding characteristic variables of the AMI incidence information variables.

4. The method for constructing a structural causal model according to any one of claims 1 to 3, characterized in that: Introducing medical prior knowledge into the causal structure of AMI.

5. The method for constructing a structural causal model according to any one of claims 1 to 3, characterized in that: Adopting additive noise models and / or medical knowledge to update the AMI causal structure.

6. The method for constructing a structural causal model according to claim 5, characterized in that: The causal discovery method is a fast causal inference algorithm based on conditional independence or a fast greedy equivalence class search algorithm based on scores.

7. An acute myocardial infarction intelligent diagnosis wearable device for acute myocardial infarction monitoring, characterized in that: It includes a signal acquisition instrument, a signal processing instrument and a display processing unit; the signal processing instrument is provided with an AMI structural causal model, and the AMI structural causal model is constructed using the structural causal model construction method described in any one of claims 1-6.

Images