Pathological critical value early warning management system based on pathological knowledge graph

Abstract

The invention, which relates to the technical field of medical health artificial intelligence, discloses a pathology critical value early warning management system based on a pathology knowledge graph, comprising a pathology data acquisition module, a pathology feature extraction module, a pathology state matching module, a pathology state early warning module and a pathology state management module. After pathological data are collected and preprocessed, a pathological feature extraction module constructs a knowledge graph and generates a composite feature vector, a pathological state matching module recognizes an acute pathological state and quantifies the matching degree, and a pathological state early warning module constructs a model in combination with an analysis result and outputs a pathological critical value. According to the method, intelligent early warning and management of the pathological critical value are realized by integrating the pathological knowledge graph and the artificial intelligence technology, the accuracy and efficiency of pathological analysis are improved, visual early warning information and decision support are provided for clinicians, the acute pathological state can be found and processed in time, and the treatment effect of a patient is improved.

Description

Technical Field

[0001] The present invention relates to the field of medical health artificial intelligence technology, and in particular to a pathology critical value early warning management system based on a pathology knowledge graph. Background Art

[0002] Traditional pathology critical value early warning management refers to a set of early warning management mechanisms established by the pathology department during the pathology examination process for examination results that may indicate that the patient is in a life-threatening state. In traditional management, once the pathology department finds a critical value, it will immediately review and confirm it to ensure the accuracy of the result. Subsequently, the critical value result will be quickly reported to the clinical department doctor through the hospital information system or telephone. After receiving the report, the clinical department doctor needs to evaluate the patient's condition in a timely manner and take appropriate treatment measures. In addition, traditional pathology critical value early warning management also includes detailed recording, tracking and follow-up of critical value reports, as well as regular review and improvement of critical value processing procedures to ensure patient safety and improve the quality of medical services.

[0003] In order to solve the problem that traditional pathology critical value warning management needs to be improved in dynamic pathology risk assessment and fragmented application of clinical knowledge, the existing technology adopts manual review and simple information system transmission to handle it. After the pathologist finds the critical value, he confirms the result through manual review, and then transmits the information to the clinical department through the hospital information system or telephone. However, this method may lead to untimely information transmission and incomplete application of clinical knowledge. Due to the fragmentation of clinical knowledge, doctors may need to spend extra time to find and integrate relevant knowledge after receiving critical value information, resulting in reduced decision-making efficiency. At the same time, the complexity of dynamic pathology risk assessment also makes manual processing prone to omissions, which in turn leads to delays in patient treatment and may even affect the patient's treatment effect and life safety. Therefore, a pathology critical value warning management system based on pathology knowledge graph is proposed. Summary of the invention

[0004] The purpose of the present invention is to provide a pathology critical value warning management system based on a pathology knowledge graph to solve the problems raised in the above-mentioned background technology.

[0005] In order to solve the above technical problems, the technical solution adopted by the present invention is: a pathological critical value warning management system based on pathological knowledge graph, including a pathological data acquisition module, a pathological feature extraction module, a pathological state matching module, a pathological state warning module and a pathological state management module;

[0006] The pathological data acquisition module acquires and pre-processes the patient's pathological data;

[0007] The pathological feature extraction module constructs a knowledge graph entity relationship table based on the preprocessed pathological data, extracts dynamic features and binds them to the knowledge graph to generate a composite feature vector;

[0008] The pathological state matching module is used to identify acute pathological states in combination with composite feature vectors and quantify the matching degree between pathological data and disease rules in the knowledge graph;

[0009] The pathological state warning module combines the analysis results of the pathological state matching module with the hierarchical attention mechanism to construct a pathological warning model and output a pathological critical value;

[0010] The pathological status management module combines the pathological critical value with the knowledge graph entity relationship table to perform graded warning and construct a three-dimensional heat map.

[0011] A further improvement of the technical solution of the present invention is that the pathological state matching module includes a pathological identification unit, a pathological prediction unit and a pathological matching unit:

[0012] The pathology recognition unit is used to combine the composite feature vector and the generalized linear model architecture to construct a logistic regression recognition model to identify acute pathological conditions;

[0013] The pathology prediction unit is used to combine the dynamic characteristics of the pathology data, use the long short-term memory network model architecture, build a pathology prediction model, and output the predicted value of the pathology data;

[0014] The pathology matching unit is used to combine the collected pathology data with the graph neural network node propagation algorithm, construct a knowledge graph reasoning model, generate rule triggering strength, and quantify the matching degree between the actual collected pathology data and the disease rules in the knowledge graph.

[0015] A further improvement of the technical solution of the present invention is that in the pathological data acquisition module, the acquisition and preprocessing process of serum potassium ion concentration data and platelet count data includes:

[0016] An ion selective electrode detection system is deployed in the pathological critical value warning management system to calculate the potassium ion concentration by measuring the potential change value. The ion selective electrode detection system has a built-in potassium ion dedicated electrode module, and the valinomycin selective sensitive membrane is covered on the electrode surface. The detection channel is connected to the sample delivery track, and the sample is injected into the detection cabin. A potential difference is generated inside and outside the electrode membrane, and then the potential change value is measured to obtain the potassium ion concentration.

[0017] A dual-sheath flow optical counting system is deployed in the pathological critical value warning management system to count platelets by measuring the forward scattered light intensity and the side fluorescence intensity. The dual-sheath flow optical counting system integrates a semiconductor laser with a wavelength of 633nm and a dual-angle scattered light detector, and controls the temperature of the detection chamber at 37±0.5℃, and then processes the samples into single cells using sheath flow technology.

[0018] The obtained potassium ion concentration data and platelet data were preprocessed. For the preprocessing of potassium ion concentration data, when the potassium ion concentration detection value fluctuated by more than 0.3mmol / L for three consecutive times, the automatic re-test program was triggered, and the median of the potassium ion concentration of the three tests was taken as the calculation result. For the preprocessing of platelet data, when the platelet histogram showed a multi-peak distribution, the impedance method was automatically switched for review, and the cell volume was measured using the Coulter principle to exclude the interference of small red blood cells. The particles with a volume of 2-20 femtoliters were counted and recorded as the effective platelet count.

[0019] A further improvement of the technical solution of the present invention is that: in the pathological feature extraction module, the process of extracting features from the pre-processed serum potassium ion concentration data and platelet count data includes:

[0020] Parse and extract the test item-critical threshold-associated pathological state triple from the patient's electronic medical record text, calculate the associated edge strength weight of the triple according to the evidence level of evidence-based medicine, store the triple as a graph database node, and use the average weight of the corresponding associated edge strength as the initial weight of the pathological state node in the knowledge graph. The graph database node attributes include test indicator parameters, pathological state coding, and treatment timeliness requirements;

[0021] The serum potassium coefficient of variation was calculated by combining the preprocessed serum potassium concentration data with a 6-hour sliding window. When the serum potassium coefficient of variation exceeded 15%, the acute kidney injury node in the knowledge graph was automatically associated. The least squares method was used to fit the trend slope of the patient's platelet within 12 hours. When the platelet trend slope was lower than -5×10 9 / L / h, the disseminated intravascular coagulation node is bound, where the coefficient of variation of serum potassium is the percentage of the standard deviation and mean of serum potassium concentration within the sliding window;

[0022] The deviation of serum potassium concentration and platelet count was calculated, and the deviation of serum potassium concentration and platelet count was combined with the weight of knowledge graph nodes to construct a weighted risk index. The critical threshold of serum potassium concentration was set to 3.2 mmol / L, and the critical threshold of platelet count was set to [30,50]×10 9 / L, when the serum potassium concentration is lower than 3.2mmol / L and is associated with the heart failure node, a binary feature F1=1 is generated, and when the platelet count is between [30,50]×109 / L interval and associated with the postoperative bleeding node to generate feature F2=1, serum potassium variation coefficient, platelet trend slope, F1, F2 and weighted risk index are fused into a spatiotemporal feature vector, its timestamp and node code are encapsulated, and the spatiotemporal feature vector is transmitted to the pathological state matching module;

[0023] Among them, the deviation of serum potassium ion concentration is the ratio of the difference between the current measured value of serum potassium ion concentration and the critical threshold of serum potassium ion concentration and the critical threshold of serum potassium ion concentration; the deviation of platelet count is the ratio of the difference between the lower limit of the critical threshold of platelet count and the current measured value of platelet count and the lower limit of the critical threshold of platelet count; the weighted risk index is the cumulative sum of the product of the deviation of the test indicator and the weight of the associated edge strength of the pathological node in the corresponding knowledge graph.

[0024] A further improvement of the technical solution of the present invention is that: in the pathology recognition unit, a logistic regression recognition model is constructed, and the process of recognizing acute pathological states includes:

[0025] The pathology recognition unit receives the spatiotemporal feature vector with serum potassium variation coefficient, platelet trend slope, F1, F2 and weighted risk index as input features, sets the initial value of the training weight coefficient of each input feature according to the knowledge graph node weight, performs weighted summation of the input feature and the training weight coefficient of each input feature to obtain a linear value, and constructs a logistic regression recognition model;

[0026] Based on the constructed logistic regression recognition model, the linear value is mapped to the probability of the pathological state through the S-type function, and the pathological state threshold is set. When the pathological state probability exceeds the pathological state threshold, it is judged as an acute pathological state, and when the pathological state probability is lower than the pathological state threshold, it is judged as a normal state;

[0027] The logarithmic loss function is used to calculate the difference between the predicted pathological state probability and the true label to obtain the loss value. The knowledge graph node weight is introduced as a regularization term. When F1=1, the learning rate of the heart failure-related weight coefficient is automatically increased to 1.5 times the basic value. If the platelet trend slope is lower than -5×10 for three consecutive hours, 9 / L / h, the locked trend slope weight is no longer updated.

[0028] A further improvement of the technical solution of the present invention is that: in the pathology prediction unit, the process of constructing a pathology prediction model and outputting a predicted value of serum potassium ion concentration and a predicted value of platelet decrease rate includes:

[0029] The pathology prediction unit receives 6 consecutive hours of serum potassium concentration data and integrates it into a serum potassium concentration detection sequence, and receives 12 consecutive hours of platelet count data and integrates it into a platelet count detection sequence;

[0030] Based on the serum potassium concentration detection sequence and the platelet count detection sequence, the difference between the serum potassium concentration in the current hour and the serum potassium concentration in the previous hour and the percentage of the serum potassium concentration in the previous hour are taken as the serum potassium concentration change rate, and the difference between the platelet count in the current hour and the platelet count in the previous hour and the percentage of the time interval are taken as the platelet decrease rate;

[0031] The change rate of serum potassium concentration is multiplied by the weight of the acute kidney injury node in the knowledge graph to generate a weighted change feature, and the platelet decrease rate is combined with the weight of the disseminated intravascular coagulation node to construct a risk-weighted rate.

[0032] The weighted change characteristics and risk weighted rate are input into the input layer of the long short-term memory network architecture. The input gate in the long short-term memory network architecture linearly combines the weighted change characteristics, risk weighted rate and the hidden state at the previous moment through the Sigmoid function to generate a gating value between 0 and 1. The forget gate determines the retention ratio of the historical state through the Sigmoid function, and uses the hyperbolic tangent function to generate a new state recommended by the current input. The fully connected layer of the long short-term memory network architecture multiplies the hidden state with the acute kidney injury node weight and the disseminated intravascular coagulation node weight respectively, adds the bias term, and outputs the predicted value of serum potassium ion concentration and the predicted value of platelet decrease rate;

[0033] Based on the collected serum potassium concentration data, platelet count data, predicted serum potassium concentration, predicted platelet decline rate, 6-hour serum potassium concentration sample size, and 12-hour platelet count sample size, the mean square error was used to calculate the prediction deviation to obtain serum potassium loss, platelet loss, and total loss;

[0034] When the predicted serum potassium value exceeds the threshold interval of the associated disease in the knowledge graph, the cell state update amount in the pathology prediction model is adjusted; when the platelet prediction rate is lower than -5×10 9 When the value of the forget gate in the pathology prediction model is locked at 1, the historical state is maintained.

[0035] A further improvement of the technical solution of the present invention is that in the pathology matching unit, the process of constructing a knowledge graph reasoning model and generating rule trigger strength includes:

[0036] The pathology matching unit receives the collected serum potassium ion concentration data and platelet count data, loads a graph database containing test item nodes, pathological status nodes and associated edge strength weights, and simultaneously obtains the serum potassium prediction value and platelet decline rate prediction value output by the pathology prediction unit;

[0037] Based on the graph neural network node propagation algorithm, the initial states of serum potassium nodes and platelet nodes are initialized as the ratio of the corresponding measured value to the critical threshold. The state weight of the pathological node is the mean of the associated edge strength weight. The state weight of each pathological node is calculated by aggregating neighborhood information, and the weighted attenuation mechanism is used to update the state weight of the pathological node in the knowledge graph. , the calculation process is as follows:

[0038] ;

[0039] in, is the state value of the pathological node at the previous moment, is the state value of the adjacent check node, is the strength weight of the associated edge, and uses the S-type function to compress the state of each pathological node to the [0,1] interval to build a knowledge graph reasoning model;

[0040] Based on serum potassium concentration and platelet count deviation, combined with the associated edge strength weight of the pathological node and the real-time pathological node state weight, the weighted cumulative formula is used to calculate the cumulative sum of the product of serum potassium concentration deviation, platelet count deviation and the corresponding associated edge strength weight and the real-time state weight of the associated pathological node, and the corresponding rule trigger strength is obtained to quantify the matching degree between the collected data and the knowledge graph rules.

[0041] Set the trigger strength threshold of a single pathology node and the total trigger strength threshold of multiple pathology nodes. When the trigger strength of a single pathology node exceeds the trigger strength threshold of a single pathology node, the corresponding handling rule is activated. When the total strength of two or more pathology nodes exceeds the total trigger strength threshold of multiple pathology nodes, it is set to the highest warning level. If the prediction error continues to exceed 15%, the weight of the associated edge is attenuated at a rate of 0.95, and a temporary rule edge with an initial weight of 0.5 is generated for abnormal data that is unrelated but has a deviation higher than 0.5.

[0042] A further improvement of the technical solution of the present invention is that: in the pathological state warning module, the process of constructing a pathological warning model and outputting a pathological critical value includes:

[0043] The pathological state warning module receives and uniformly quantifies the acute pathological state probability value, the serum potassium prediction value for the next two hours, the platelet decline rate prediction value, and the rule trigger strength vector, expands the serum potassium prediction value and the platelet decline rate prediction value to generate a two-hour prediction sequence, and performs six-hour sliding window mean smoothing on the rule trigger strength;

[0044] The weights of the logistic regression recognition model, pathology prediction model and knowledge graph reasoning model are allocated through a trainable parameter matrix and a normalized exponential function. The prediction contribution weights of the current moment and the future moment are calculated using an S-type function for the time series output of the pathology prediction model, forming a double-layer attention allocation mechanism and constructing a pathology early warning model.

[0045] Based on the weights of each model and the predicted contribution weight, spatial fusion calculation is performed to output the pathological critical value , the calculation process is as follows:

[0046] ;

[0047] in, , and is the weight of the logistic regression recognition model, pathology prediction model and knowledge graph reasoning model, P is the pathological state probability output by the logistic regression recognition model, is the 6-hour sliding mean of the rule triggering intensity, and The weights assigned to the predicted values ​​at different future time points output by the pathology prediction unit, and The predicted values ​​of serum potassium concentration and platelet decrease rate in the next 1 hour and 2 hours;

[0048] The pathological criticality judgment threshold is adjusted according to the remaining proportion of each pathological treatment time window in the knowledge graph. When the pathological criticality value exceeds the pathological criticality judgment threshold, the warning level is determined in combination with the maximum node weight in the rule trigger intensity vector, and the standard treatment plan in the graph is associated to generate clinical decision recommendations.

[0049] A further improvement of the technical solution of the present invention is that: in the pathological state early warning module, the process of importing the latest output results of each model into the pathological early warning model and outputting the pathological critical value includes:

[0050] The latest acute pathological state probability values ​​output by the logistic regression recognition model, pathology prediction model and knowledge graph reasoning model, the predicted values ​​of serum potassium and platelet decrease rate in the next two hours, and the rule triggering intensity vector are imported into the pathology warning model to obtain the pathological critical value.

[0051] A further improvement of the technical solution of the present invention is that in the pathological state management module, the process of combining the pathological critical value with the knowledge graph entity relationship table for graded warning and constructing a three-dimensional heat map includes:

[0052] Set the upper and lower limits of the thresholds for associated disease nodes, and divide the pathological critical values ​​into special warning, level I warning, and level II warning;

[0053] If the pathological critical value exceeds 1.2 times the upper limit of the threshold of the associated disease node and two or more high-weight pathological nodes are activated at the same time, it is judged as a special-level warning. If the pathological critical value is within the threshold interval of the associated disease node and triggers a single high-weight node, it is judged as a level I warning. If the pathological critical value is close to the lower limit of the threshold of the associated disease node and the rule triggering intensity meets the standard, it is judged as a level II warning. The warning level is dynamically increased in combination with the remaining proportion of the treatment time window. When the remaining time is less than 30 minutes, the warning level is automatically upgraded, and the knowledge graph is linked to extract the standard treatment plan corresponding to the highest weight node;

[0054] The coordinate system of the three-dimensional heat map is defined as the horizontal 24-h time dimension axis, the vertical serum potassium and platelet dual-scale test index axis, and the vertical pathological state axis. The thermal value of each space-time point is calculated based on the rule trigger intensity and time attenuation factor, and the upper and lower limits of the space-time point thermal value thresholds are set. The red, orange, and yellow warning areas are mapped according to the threshold interval of the space-time point thermal value. The continuous high-risk area implementation blocks are displayed in an aggregated manner. When the space-time point thermal value suddenly increases by more than 2 times the historical average, a pulse flashing mark is added.

[0055] Due to the adoption of the above technical solution, the present invention has the following technical advances compared with the prior art:

[0056] 1. The present invention provides a pathology critical value warning management system based on a pathology knowledge graph. Through the collaborative work of a pathology data acquisition module and a pathology feature extraction module, key features can be efficiently and accurately extracted from massive pathology data, and bound to the knowledge graph entity relationship table to generate a composite feature vector, which greatly improves the accuracy and efficiency of pathology data analysis and provides a solid foundation for subsequent pathology status matching and warning.

[0057] 2. The present invention provides a pathology critical value warning management system based on a pathology knowledge graph. The pathology state matching module can combine the composite feature vector to quickly identify the acute pathological state and quantify the matching degree between the pathology data and the disease rules in the knowledge graph. This function enables the system to timely detect the patient's pathology critical value and provide clinicians with timely and accurate warning information, which helps to shorten the diagnosis time and improve the treatment effect.

[0058] 3. The present invention provides a pathology critical value warning management system based on a pathology knowledge graph. The pathology status management module can combine the pathology critical value with the knowledge graph entity relationship table to perform graded warnings and construct a three-dimensional heat map. This function makes the warning information more intuitive and easy to understand, making it convenient for clinicians to quickly grasp the patient's condition and degree of criticality, thereby formulating a more reasonable and effective treatment plan. At the same time, the construction of the three-dimensional heat map also provides a powerful tool for the visualization and in-depth analysis of pathology data. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.

[0060] Figure 1 A block diagram of the present invention. DETAILED DESCRIPTION

[0061] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0062] Examples, such as Figure 1 As shown, the present invention provides a pathological critical value warning management system based on a pathological knowledge graph, including a pathological data acquisition module, a pathological feature extraction module, a pathological state matching module, a pathological state warning module and a pathological state management module;

[0063] The pathological data acquisition module collects and pre-processes the patient's pathological data. The ion selective electrode detection system is deployed in the pathological critical value early warning management system. The potassium ion concentration is calculated by measuring the potential change value. The ion selective electrode detection system has a built-in potassium ion dedicated electrode module. The electrode surface is covered with a valinomycin selective sensitive membrane. The detection channel is connected to the sample delivery track. The sample is injected into the detection cabin. A potential difference is generated inside and outside the electrode membrane, and then the potential change value is measured to obtain the potassium ion concentration. The dual sheath flow optical counting system is deployed in the pathological critical value early warning management system. Platelets are counted by measuring the forward scattered light intensity and the side fluorescence intensity. The dual sheath flow optical counting system integrates a semi-circular light with a wavelength of 633nm. Conductor laser and dual-angle scattered light detector, and control the temperature of the detection chamber at 37±0.5℃, and then use sheath flow technology to process the sample into single cells, and pre-process the acquired potassium ion concentration data and platelet data. For the pre-processing of potassium ion concentration data, when the potassium ion concentration detection value fluctuates by more than 0.3mmol / L for three consecutive times, the automatic re-test program is triggered, and the median of the potassium ion concentration of the three tests is taken as the calculation result. For the pre-processing of platelet data, when the platelet histogram shows multi-peak distribution, the impedance method is automatically switched for review, and the cell volume is measured by the Coulter principle to exclude the interference of small red blood cells, and the particles with a volume of 2-20 femtoliters are counted and recorded as the effective platelet count;

[0064] The pathological feature extraction module builds a knowledge graph entity relationship table based on the preprocessed pathological data, extracts dynamic features and binds them to the knowledge graph, generates a composite feature vector, parses and extracts the test item-critical threshold-associated pathological state triple from the patient's electronic medical record text, calculates the associated edge strength weight of the triple according to the evidence level of evidence-based medicine, stores the triple as a graph database node, and uses the average weight of the corresponding associated edge strength as the initial weight of the pathological state node in the knowledge graph. The graph database node attributes include test index parameters, pathological state coding and treatment timeliness requirements. The serum potassium variation coefficient is calculated by combining the preprocessed serum potassium ion concentration data and a 6-hour sliding window. When the serum potassium variation coefficient exceeds 15%, it is automatically associated with the acute kidney injury node in the knowledge graph. The least squares method is used to fit the trend slope of the patient's platelet within 12 hours. When the platelet trend slope is lower than -5×10 9 / L / h, the disseminated intravascular coagulation node is bound, where the serum potassium coefficient of variation is the percentage of the standard deviation and mean of the serum potassium ion concentration in the sliding window. The deviation of serum potassium ion concentration and platelet count are calculated, and the deviation of serum potassium ion concentration and platelet count are combined with the weight of the knowledge graph node to construct a weighted risk index. The critical threshold of serum potassium ion concentration is set to 3.2mmol / L, and the critical threshold of platelet count is set to [30,50]×10 9 / L, when the serum potassium concentration is lower than 3.2mmol / L and is associated with the heart failure node, a binary feature F1=1 is generated, and when the platelet count is between [30,50]×10 9 / L interval and associated with the postoperative bleeding node, generate feature F2=1, fuse the serum potassium coefficient of variation, platelet trend slope, F1, F2 and weighted risk index into a spatiotemporal feature vector, encapsulate its timestamp and node code, and transmit the spatiotemporal feature vector to the pathological state matching module, where the serum potassium ion concentration deviation is the difference between the current serum potassium ion concentration measured value and the serum potassium ion concentration critical threshold and the ratio of the serum potassium ion concentration critical threshold, the platelet count deviation is the difference between the lower limit of the platelet count critical threshold and the current platelet count measured value and the ratio of the lower limit of the platelet count critical threshold, and the weighted risk index is the cumulative sum of the product of the test indicator deviation and the associated edge strength weight of the pathological node in the corresponding knowledge graph;

[0065] The pathological state matching module is used to identify acute pathological states by combining composite feature vectors and quantify the matching degree between pathological data and disease rules in the knowledge graph. The pathological state matching module includes a pathological recognition unit, a pathological prediction unit and a pathological matching unit. The pathological recognition unit is used to combine composite feature vectors and a generalized linear model architecture to construct a logistic regression recognition model to identify acute pathological states. The pathological recognition unit receives the spatiotemporal feature vector with serum potassium variation coefficient, platelet trend slope, F1, F2 and weighted risk index as input features, sets the initial value of the training weight coefficient of each input feature according to the knowledge graph node weight, and compares the input features with the training weight coefficient of each input feature. The weighted sum is performed to obtain the linear value, and a logistic regression recognition model is constructed. Based on the constructed logistic regression recognition model, the linear value is mapped to the probability of the pathological state through the S-type function, and the pathological state threshold is set. When the pathological state probability exceeds the pathological state threshold, it is judged as an acute pathological state. When the pathological state probability is lower than the pathological state threshold, it is judged as a normal state. The logarithmic loss function is used to calculate the difference between the predicted pathological state probability and the true label to obtain the loss value. The knowledge graph node weight is introduced as a regularization term. When F1=1, the learning rate of the heart failure-related weight coefficient is automatically increased to 1.5 times the basic value. If the platelet trend slope is lower than -5×10 9 / L / h, the locked trend slope weight is no longer updated. The pathology prediction unit is used to combine the dynamic characteristics of the pathology data, use the long short-term memory network model architecture, build a pathology prediction model, and output the predicted value of the pathology data. The pathology prediction unit receives 6 consecutive hours of serum potassium ion concentration data and integrates it into a serum potassium ion concentration detection sequence. It receives 12 consecutive hours of platelet count data and integrates it into a platelet count detection sequence. Based on the serum potassium ion concentration detection sequence and the platelet count detection sequence, the difference between the current hour serum potassium ion concentration and the previous hour serum potassium ion concentration and the percentage of the previous hour serum potassium ion concentration are used as the serum potassium ion concentration change rate. The difference between the current hour platelet count and the previous hour platelet count and the percentage of the time interval are used as the platelet decrease rate. The serum potassium ion concentration change rate is multiplied by the weight of the acute kidney injury node in the knowledge graph to generate a weighted change feature. The platelet decrease rate is combined with the weight of the disseminated intravascular coagulation node to construct a risk-weighted rate. The weighted change feature and the risk-weighted rate are input into the long short-term memory network architecture. The input gate in the long short-term memory network architecture uses the Sigmoid function to linearly combine the weighted change characteristics, risk weighted rate and hidden state of the previous moment to generate a gating value between 0 and 1. The forget gate determines the retention ratio of the historical state through the Sigmoid function, and uses the hyperbolic tangent function to generate the new state recommended by the current input. The fully connected layer of the long short-term memory network architecture multiplies the hidden state with the acute kidney injury node weight and the disseminated intravascular coagulation node weight respectively, adds the bias term, and outputs the predicted value of serum potassium concentration and the predicted value of platelet decrease rate. Based on the collected serum potassium concentration data, platelet count data, serum potassium concentration predicted value, platelet decrease rate predicted value, 6 consecutive hours of serum potassium concentration sample volume and 12 hours of platelet count sample volume, the mean square error is used to calculate the prediction deviation to obtain serum potassium loss, platelet loss and total loss. When the serum potassium prediction value exceeds the threshold interval of the associated disease in the knowledge graph, the cell state update amount in the pathology prediction model is adjusted. When the platelet prediction rate is lower than -5×10 9 / L / h, the forget gate value in the pathology prediction model is locked to 1, and the historical state is maintained. The pathology matching unit is used to combine the collected pathology data with the graph neural network node propagation algorithm, build a knowledge graph reasoning model, generate rule trigger strength, and quantify the degree of matching between the actual collected pathology data and the disease rules in the knowledge graph. The pathology matching unit receives the collected serum potassium ion concentration data and platelet count data, loads the graph database containing the test item nodes, pathology status nodes and the associated edge strength weights, and synchronously obtains the serum potassium prediction value and the platelet decrease rate prediction value output by the pathology prediction unit. Based on the graph neural network node propagation algorithm, the initial states of the serum potassium node and the platelet node are initialized to the ratio of the corresponding measured value to the critical threshold value, and the pathology node state weight is the average of the associated edge strength weights. The state weight of each pathology node is calculated by aggregating neighborhood information, and the weighted attenuation mechanism is used to update the state weight of the pathology node in the knowledge graph. , the calculation process is as follows:

[0066] ;

[0067] in, is the state value of the pathological node at the previous moment, is the state value of the adjacent check node, is the strength weight of the associated edge, and the S-type function is used to compress the state of each pathological node to the interval [0,1], and a knowledge graph reasoning model is constructed. Based on serum potassium ion concentration and platelet count deviation, combined with the strength weight of the associated edge of the pathological node and the weight of the real-time pathological node state, the weighted accumulation formula is used to calculate the cumulative sum of the product of the deviation of serum potassium ion concentration and platelet count and the corresponding strength weight of the associated edge and the real-time state weight of the associated pathological node, and the corresponding rule trigger strength is obtained. The matching degree between the collected data and the knowledge graph rules is quantified, and the single pathological node trigger strength threshold and the multi-pathological node trigger total strength threshold are set. When the trigger strength of a single pathological node exceeds the single pathological node trigger strength threshold, the corresponding treatment rule is activated. When the total strength of two or more pathological nodes exceeds the multi-pathological node trigger total strength threshold, it is set to the highest warning level. If the prediction error continues to exceed 15%, the weight of the associated edge is attenuated by 0.95 times, and a temporary rule edge with an initial weight of 0.5 is generated for abnormal data that is not associated but has a deviation higher than 0.5.

[0068] The pathological state warning module combines the analysis results of the pathological state matching module with the hierarchical attention mechanism to construct a pathological warning model and output the pathological critical value. The pathological state warning module receives and uniformly quantifies the acute pathological state probability value, the serum potassium prediction value for the next two hours, the platelet decline rate prediction value, and the rule trigger strength vector. The serum potassium prediction value and the platelet decline rate prediction value are expanded to generate a two-hour prediction sequence, and the rule trigger strength is smoothed with a six-hour sliding window mean. The weights of the logistic regression recognition model, the pathological prediction model, and the knowledge graph reasoning model are allocated through a trainable parameter matrix and a normalized exponential function. For the time series output of the pathological prediction model, the S-type function is used to calculate the prediction contribution weights of the current moment and the future moment, forming a double-layer attention allocation mechanism, constructing a pathological warning model, and performing spatial fusion calculation based on the weights of each model and the prediction contribution weights to output the pathological critical value. , the calculation process is as follows:

[0069] ;

[0070] in, , and is the weight of the logistic regression recognition model, pathology prediction model and knowledge graph reasoning model, P is the pathological state probability output by the logistic regression recognition model, is the 6-hour sliding mean of the rule triggering intensity, and The weights assigned to the predicted values ​​at different future time points output by the pathology prediction unit, and The predicted values ​​of serum potassium concentration and platelet decline rate in the next 1 hour and 2 hours are used. The pathological criticality judgment threshold is adjusted according to the remaining proportion of each pathological treatment time window in the knowledge graph. When the pathological criticality value exceeds the pathological criticality judgment threshold, the warning level is determined in combination with the maximum node weight in the rule trigger strength vector, and the standard treatment plan in the graph is associated to generate clinical decision recommendations. The latest acute pathological state probability values ​​output by the logistic regression recognition model, pathological prediction model and knowledge graph reasoning model, the predicted serum potassium value and platelet decline rate in the next two hours, and the rule trigger strength vector are imported into the pathological warning model to obtain the pathological criticality value.

[0071] The pathological status management module combines the pathological critical value with the knowledge graph entity relationship table for graded warning, and constructs a three-dimensional heat map, sets the upper and lower limits of the associated disease node threshold, and divides the pathological critical value into special warning, level I warning and level II warning. If the pathological critical value exceeds 1.2 times the upper limit of the associated disease node threshold and activates two or more high-weight pathological nodes at the same time, it is judged as a special warning. If the pathological critical value is in the threshold interval of the associated disease node and triggers a single high-weight node, it is judged as a level I warning. If the pathological critical value is close to the lower limit of the associated disease node threshold and the rule triggering intensity meets the standard, it is judged as a level II warning. Combined with the treatment The remaining ratio of the window dynamically increases the warning level. When the remaining time is less than 30 minutes, the warning level is automatically upgraded. The knowledge graph is linked to extract the standard disposal plan corresponding to the highest weight node. The horizontal 24-hour time dimension axis, the vertical serum potassium and platelet dual-scale test index axis, and the vertical pathological state axis are defined to form the coordinate system of the three-dimensional heat map. The thermal value of each space-time point is calculated based on the rule trigger intensity and time attenuation factor, and the upper and lower limits of the space-time point thermal value threshold are set. The red, orange, and yellow warning areas are mapped according to the threshold interval of the space-time point thermal value, and the continuous high-risk area implementation blocks are displayed in an aggregated manner. When the space-time point thermal value suddenly increases by more than 2 times the historical average, a pulse flashing mark is added.

[0072] First, the pathological data of the patient is collected and preprocessed through the pathological data acquisition module. Then, the pathological feature extraction module uses the preprocessed data to construct the knowledge graph entity relationship table, and extracts dynamic features and binds them to the knowledge graph to generate a composite feature vector. Then, the pathological state matching module combines the composite feature vector to identify the acute pathological state and quantify the degree of matching between the pathological data and the disease rules in the knowledge graph. Subsequently, the pathological state warning module combines the analysis results of the matching module with the hierarchical attention mechanism to construct a pathological warning model and output the pathological criticality value. Finally, the pathological state management module performs graded warnings based on the pathological criticality value and the knowledge graph entity relationship table, and constructs a three-dimensional heat map so that clinicians can intuitively and quickly understand the patient's condition and criticality.

[0073] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A pathology critical value warning management system based on pathology knowledge graph, characterized by: It includes a pathological data acquisition module, a pathological feature extraction module, a pathological state matching module, a pathological state early warning module and a pathological state management module; The pathological data acquisition module acquires and pre-processes the patient's pathological data; The pathological feature extraction module constructs a knowledge graph entity relationship table based on the preprocessed pathological data, extracts dynamic features and binds them to the knowledge graph to generate a composite feature vector; The pathological state matching module is used to identify acute pathological states in combination with composite feature vectors and quantify the matching degree between pathological data and disease rules in the knowledge graph; The pathological state warning module combines the analysis results of the pathological state matching module with the hierarchical attention mechanism to construct a pathological warning model and output a pathological critical value; The pathological status management module combines the pathological critical value with the knowledge graph entity relationship table to perform graded warning and construct a three-dimensional heat map.

2. According to claim 1, a pathology critical value early warning management system based on pathology knowledge graph is characterized by: The pathological state matching module includes a pathological identification unit, a pathological prediction unit and a pathological matching unit: The pathology recognition unit is used to combine the composite feature vector and the generalized linear model architecture to construct a logistic regression recognition model to identify acute pathological conditions; The pathology prediction unit is used to combine the dynamic characteristics of the pathology data, use the long short-term memory network model architecture, build a pathology prediction model, and output the predicted value of the pathology data; The pathology matching unit is used to combine the collected pathology data with the graph neural network node propagation algorithm, construct a knowledge graph reasoning model, generate rule triggering strength, and quantify the matching degree between the actual collected pathology data and the disease rules in the knowledge graph.

3. According to claim 2, a pathology critical value early warning management system based on pathology knowledge graph is characterized in that: In the pathological data acquisition module, the acquisition and preprocessing process of serum potassium ion concentration data and platelet count data includes: Deploy an ion selective electrode detection system in the pathological critical value warning management system to calculate the potassium ion concentration by measuring the potential change value; A dual-sheath flow optical counting system is deployed in the pathological critical value warning management system to count platelets by measuring the forward scattered light intensity and side fluorescence intensity; The obtained potassium ion concentration data and platelet data were preprocessed. For the preprocessing of potassium ion concentration data, when the potassium ion concentration detection value fluctuated by more than 0.3mmol / L for three consecutive times, the automatic re-test program was triggered, and the median of the potassium ion concentration of the three tests was taken as the calculation result. For the preprocessing of platelet data, when the platelet histogram showed a multi-peak distribution, the impedance method was automatically switched for review, and the cell volume was measured using the Coulter principle to exclude the interference of small red blood cells. The particles with a volume of 2-20 femtoliters were counted and recorded as the effective platelet count.

4. According to claim 3, a pathology critical value early warning management system based on pathology knowledge graph is characterized in that: In the pathological feature extraction module, the process of extracting features from the pre-processed serum potassium ion concentration data and platelet count data includes: Parse and extract the test item-critical threshold-associated pathological state triple from the patient's electronic medical record text, calculate the associated edge strength weight of the triple according to the evidence level of evidence-based medicine, store the triple as a graph database node, and use the average weight of the corresponding associated edge strength as the initial weight of the pathological state node in the knowledge graph. The graph database node attributes include test indicator parameters, pathological state coding, and treatment timeliness requirements; The serum potassium coefficient of variation was calculated by combining the preprocessed serum potassium concentration data with a 6-hour sliding window. When the serum potassium coefficient of variation exceeded 15%, the acute kidney injury node in the knowledge graph was automatically associated. The least squares method was used to fit the trend slope of the patient's platelet within 12 hours. When the platelet trend slope was lower than -5×10 9 / L / h, the disseminated intravascular coagulation node is bound, where the coefficient of variation of serum potassium is the percentage of the standard deviation and mean of serum potassium concentration within the sliding window; The deviation of serum potassium concentration and platelet count was calculated, and the deviation of serum potassium concentration and platelet count was combined with the weight of knowledge graph nodes to construct a weighted risk index. The critical threshold of serum potassium concentration was set to 3.2 mmol / L, and the critical threshold of platelet count was set to [30,50]×10 9 / L, when the serum potassium concentration is lower than 3.2mmol / L and is associated with the heart failure node, a binary feature F1=1 is generated, and when the platelet count is between [30,50]×10 9 / L interval and associated with the postoperative bleeding node to generate feature F2=1, serum potassium variation coefficient, platelet trend slope, F1, F2 and weighted risk index are fused into a spatiotemporal feature vector, its timestamp and node code are encapsulated, and the spatiotemporal feature vector is transmitted to the pathological state matching module; Among them, the deviation of serum potassium ion concentration is the ratio of the difference between the current measured value of serum potassium ion concentration and the critical threshold of serum potassium ion concentration and the critical threshold of serum potassium ion concentration; the deviation of platelet count is the ratio of the difference between the lower limit of the critical threshold of platelet count and the current measured value of platelet count and the lower limit of the critical threshold of platelet count; the weighted risk index is the cumulative sum of the product of the deviation of the test indicator and the weight of the associated edge strength of the pathological node in the corresponding knowledge graph.

5. According to claim 4, a pathology critical value early warning management system based on pathology knowledge graph is characterized in that: In the pathology recognition unit, a logistic regression recognition model is constructed, and the process of recognizing acute pathological states includes: The pathology recognition unit receives the spatiotemporal feature vector with serum potassium variation coefficient, platelet trend slope, F1, F2 and weighted risk index as input features, sets the initial value of the training weight coefficient of each input feature according to the knowledge graph node weight, performs weighted summation of the input feature and the training weight coefficient of each input feature to obtain a linear value, and constructs a logistic regression recognition model; Based on the constructed logistic regression recognition model, the linear value is mapped to the probability of the pathological state through the S-type function, and the pathological state threshold is set. When the pathological state probability exceeds the pathological state threshold, it is judged as an acute pathological state, and when the pathological state probability is lower than the pathological state threshold, it is judged as a normal state; The logarithmic loss function is used to calculate the difference between the predicted pathological state probability and the true label to obtain the loss value. The knowledge graph node weight is introduced as a regularization term. When F1=1, the learning rate of the heart failure-related weight coefficient is automatically increased to 1.5 times the basic value. If the platelet trend slope is lower than -5×10 for three consecutive hours, 9 / L / h, the locked trend slope weight is no longer updated.

6. A pathology critical value early warning management system based on pathology knowledge graph according to claim 5, characterized in that: In the pathology prediction unit, the process of constructing a pathology prediction model and outputting a predicted value of serum potassium ion concentration and a predicted value of platelet decrease rate includes: The pathology prediction unit receives 6 consecutive hours of serum potassium concentration data and integrates it into a serum potassium concentration detection sequence, and receives 12 consecutive hours of platelet count data and integrates it into a platelet count detection sequence; Based on the serum potassium concentration detection sequence and the platelet count detection sequence, the difference between the serum potassium concentration in the current hour and the serum potassium concentration in the previous hour and the percentage of the serum potassium concentration in the previous hour are taken as the serum potassium concentration change rate, and the difference between the platelet count in the current hour and the platelet count in the previous hour and the percentage of the time interval are taken as the platelet decrease rate; The change rate of serum potassium concentration is multiplied by the weight of the acute kidney injury node in the knowledge graph to generate a weighted change feature, and the platelet decrease rate is combined with the weight of the disseminated intravascular coagulation node to construct a risk-weighted rate. The weighted change characteristics and risk-weighted rates are input into the input layer of the long short-term memory network architecture to construct a pathology prediction model, and output the predicted values ​​of serum potassium ion concentration and platelet decrease rate; Based on the collected serum potassium concentration data, platelet count data, predicted serum potassium concentration, predicted platelet decline rate, 6-hour serum potassium concentration sample size, and 12-hour platelet count sample size, the mean square error was used to calculate the prediction deviation to obtain serum potassium loss, platelet loss, and total loss; When the predicted serum potassium value exceeds the threshold interval of the associated disease in the knowledge graph, the cell state update amount in the pathology prediction model is adjusted; when the platelet prediction rate is lower than -5×10 9 / L / h, the forget gate value in the pathology prediction model is locked to 1, and the historical state is maintained.

7. A pathology critical value early warning management system based on pathology knowledge graph according to claim 6, characterized in that: In the pathology matching unit, the process of constructing a knowledge graph reasoning model and generating rule trigger strength includes: The pathology matching unit receives the collected serum potassium ion concentration data and platelet count data, loads a graph database containing test item nodes, pathological status nodes and associated edge strength weights, and simultaneously obtains the serum potassium prediction value and platelet decline rate prediction value output by the pathology prediction unit; Based on the graph neural network node propagation algorithm, the initial states of serum potassium nodes and platelet nodes are initialized as the ratio of the corresponding measured value to the critical threshold. The state weight of the pathological node is the mean of the associated edge strength weight. The state weight of each pathological node is calculated by aggregating neighborhood information, and the weighted attenuation mechanism is used to update the state weight of the pathological node in the knowledge graph. , and use the S-type function to compress the state of each pathological node to the [0,1] interval to build a knowledge graph reasoning model; Based on serum potassium concentration and platelet count deviation, combined with the associated edge strength weight of the pathological node and the real-time pathological node state weight, the weighted cumulative formula is used to calculate the cumulative sum of the product of serum potassium concentration deviation, platelet count deviation and the corresponding associated edge strength weight and the real-time state weight of the associated pathological node, and the corresponding rule trigger strength is obtained to quantify the matching degree between the collected data and the knowledge graph rules. Set the trigger strength threshold of a single pathology node and the total trigger strength threshold of multiple pathology nodes. When the trigger strength of a single pathology node exceeds the trigger strength threshold of a single pathology node, the corresponding handling rule is activated. When the total strength of two or more pathology nodes exceeds the total trigger strength threshold of multiple pathology nodes, it is set to the highest warning level. If the prediction error continues to exceed 15%, the weight of the associated edge is attenuated at a rate of 0.95, and a temporary rule edge with an initial weight of 0.5 is generated for abnormal data that is unrelated but has a deviation higher than 0.

5.

8. A pathology critical value early warning management system based on pathology knowledge graph according to claim 7, characterized in that: In the pathological state warning module, the process of constructing a pathological warning model and outputting a pathological critical value includes: The pathological state warning module receives and uniformly quantifies the acute pathological state probability value, the serum potassium prediction value for the next two hours, the platelet decline rate prediction value, and the rule trigger strength vector, expands the serum potassium prediction value and the platelet decline rate prediction value to generate a two-hour prediction sequence, and performs six-hour sliding window mean smoothing on the rule trigger strength; The weights of the logistic regression recognition model, pathology prediction model and knowledge graph reasoning model are allocated through a trainable parameter matrix and a normalized exponential function. The prediction contribution weights of the current moment and the future moment are calculated using an S-type function for the time series output of the pathology prediction model, forming a double-layer attention allocation mechanism and constructing a pathology early warning model. Based on the weights of each model and the predicted contribution weight, spatial fusion calculation is performed to output the pathological critical value , the pathological criticality judgment threshold is adjusted according to the remaining proportion of each pathological treatment time window in the knowledge graph. When the pathological criticality value exceeds the pathological criticality judgment threshold, the warning level is determined in combination with the maximum node weight in the rule trigger intensity vector, and the standard treatment plan in the graph is associated to generate clinical decision recommendations.

9. A pathology critical value early warning management system based on pathology knowledge graph according to claim 8, characterized in that: In the pathological state early warning module, the process of importing the latest output results of each model into the pathological early warning model and outputting the pathological critical value includes: The latest acute pathological state probability values ​​output by the logistic regression recognition model, pathology prediction model and knowledge graph reasoning model, the predicted values ​​of serum potassium and platelet decrease rate in the next two hours, and the rule triggering intensity vector are imported into the pathology warning model to obtain the pathological critical value.

10. A pathology critical value early warning management system based on pathology knowledge graph according to claim 9, characterized in that: In the pathological state management module, the process of combining the pathological critical value with the knowledge graph entity relationship table to perform graded warning and construct a three-dimensional heat map includes: Set the upper and lower limits of the thresholds for the associated disease nodes, and divide the pathological critical values ​​into special warning, level I warning, and level II warning; If the pathological critical value exceeds 1.2 times the upper limit of the threshold of the associated disease node and two or more high-weight pathological nodes are activated at the same time, it is judged as a special-level warning. If the pathological critical value is within the threshold interval of the associated disease node and triggers a single high-weight node, it is judged as a level I warning. If the pathological critical value is close to the lower limit of the threshold of the associated disease node and the rule triggering intensity meets the standard, it is judged as a level II warning. The warning level is dynamically increased in combination with the remaining proportion of the treatment time window. When the remaining time is less than 30 minutes, the warning level is automatically upgraded, and the knowledge graph is linked to extract the standard treatment plan corresponding to the highest weight node; The coordinate system of the three-dimensional heat map is defined as the horizontal 24-h time dimension axis, the vertical serum potassium and platelet dual-scale test index axis, and the vertical pathological state axis. The thermal value of each space-time point is calculated based on the rule trigger intensity and time attenuation factor, and the upper and lower limits of the space-time point thermal value thresholds are set. The red, orange, and yellow warning areas are mapped according to the threshold interval of the space-time point thermal value. The continuous high-risk area implementation blocks are displayed in an aggregated manner. When the space-time point thermal value suddenly increases by more than 2 times the historical average, a pulse flashing mark is added.

Images