CN112420201B - Deep cascading framework for ICU mortality prediction - Google Patents
Deep cascading framework for ICU mortality prediction Download PDFInfo
- Publication number
- CN112420201B CN112420201B CN202011341957.4A CN202011341957A CN112420201B CN 112420201 B CN112420201 B CN 112420201B CN 202011341957 A CN202011341957 A CN 202011341957A CN 112420201 B CN112420201 B CN 112420201B
- Authority
- CN
- China
- Prior art keywords
- network
- sub
- cascade
- node
- disease
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Images
Classifications
-
- G—PHYSICS
- G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
- G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
- G16H50/00—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics
- G16H50/50—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics for simulation or modelling of medical disorders
Landscapes
- Health & Medical Sciences (AREA)
- Engineering & Computer Science (AREA)
- Medical Informatics (AREA)
- Public Health (AREA)
- Biomedical Technology (AREA)
- Data Mining & Analysis (AREA)
- Databases & Information Systems (AREA)
- Pathology (AREA)
- Epidemiology (AREA)
- General Health & Medical Sciences (AREA)
- Primary Health Care (AREA)
- Medical Treatment And Welfare Office Work (AREA)
Abstract
The invention provides a deep cascading framework for ICU mortality prediction and an ICU mortality prediction method, wherein the deep cascading framework for the ICU mortality prediction comprises the following steps: the sign sub-network and the disease sub-network are connected in a cascade mode according to node failure conditions, and both the sign sub-network and the disease sub-network are used for outputting failure distribution so as to conduct ICU mortality prediction through the failure distribution. The invention has the beneficial effects that: the method can conveniently predict the death rate of ICU patients and can make the prediction interpretable.
Description
Technical Field
The invention relates to the field of ICU mortality prediction, in particular to a deep cascade framework for ICU mortality prediction.
Background
Medical risk detection is an important topic for improving the clinical practice capability of the ICU, and many biometric learning methods and deep learning methods based on quantitative tables and characteristics can predict the mortality rate of a specific patient and assist physicians in making corresponding clinical decisions. However, current methods not only require experts to manually define risk factors but also rely heavily on the effectiveness of the pre-training features. From the model perspective, the end-to-end prediction model lacks interpretability, and the black box type neural network structure is difficult to obtain an inference path.
Disclosure of Invention
The problem addressed by the present invention is how to facilitate the prediction of ICU patient mortality and render the prediction interpretable.
The invention provides a deep cascading framework for ICU mortality prediction, which comprises the following steps: the sign sub-network and the disease sub-network are connected in a cascade mode according to node failure conditions, and both the sign sub-network and the disease sub-network are used for outputting failure distribution so as to conduct ICU mortality prediction through the failure distribution.
Further, the node failure condition comprises a failure set, and the failure set comprises the front part with the maximum failure risk in the cascade stagenAnd (4) each node.
Further, the node failure condition further includes network connectivity, the network connectivity includes an effective edge probability of the interaction edge, the effective edge probability represents a probability that one interaction edge is randomly selected and the edge connects two effective nodes, and the effective edge probability generated in the current cascade stage is used for updating the effective edge probability in the next cascade stage.
Further, the sign sub-network and the disease sub-network each include a degree distribution unit and an excess degree distribution unit, the sign sub-network and the disease sub-network each determine the failure set through the degree distribution unit, and the sign sub-network and the disease sub-network each determine an effective edge probability of a next cascade stage through the excess degree distribution unit according to the effective edge probability of a current cascade stage.
Further, the degree distribution unit includes a degree distribution generating function, the excess distribution unit includes an excess distribution generating function, and the recursive relation of the depth cascade framework includes:
in the first placelIn the secondary cascade, the effective node proportions of the disease sub-network and the sign sub-network are respectively as follows:(ii) a Or the like, or, alternatively,
in the first placelIn the secondary cascade, the effective node proportions of the sign sub-network and the disease sub-network are respectively as follows:;
wherein the content of the first and second substances,a sub-network of a disease is represented,a sub-network of signs is represented,anda degree distribution generating function representing the disease sub-network and the signs sub-network respectively,anda hyperdynamic distribution generating function representing the disease sub-network and the signs sub-network respectively,lthe number of cascades is indicated,represents the number of cascades asl-1 a hypernym distribution generation function of the vital sign sub-network,represents the number of cascades asl-1 a hyperdistribution generating function of the disease sub-network,represents the number of cascades aslGenerating a function of the hyperdistribution of the disease sub-network,represents the number of cascades aslOf the sign sub-networkThe degree distribution is generated as a function of,represents the number of cascades aslA function is generated of the degree distribution of the disease sub-network,represents the number of cascades aslThe degree distribution of the sign sub-network of (1) generates a function,the probability of the valid edge is represented,represents the number of cascades aslOf the disease sub-network,represents the number of cascades aslThe effective edge probability of the sign subnetwork.
Further, the sign sub-network and the disease sub-network both determine the failure distribution according to the effective probabilities of the nodes, wherein the failure distribution of the current cascade stage is generated by normalizing the effective probabilities of all the nodes of the cascade stage.
Further, in the sign sub-network and the disease sub-network, the effective probability of the node is determined according to the number of failed nodes in each node neighborhood, wherein the failure set generated in the current cascade stage is used for updating the number of failed nodes in each node neighborhood in the next cascade stage.
Further, the calculation formula of the effective probability of the node is as follows:
wherein, the first and the second end of the pipe are connected with each other,the nodes are represented as a list of nodes,mrepresenting the number of failed nodes in the neighborhood of nodes,which represents the number of neighbors of the node,indicating the number of valid neighbors of the node,it is shown that the distribution is tolerated,which represents the tolerance factor of the current signal,Xrepresenting the signs subnetwork or the diseases subnetwork.
Further, each node of the signs sub-network and the diseases sub-network has a personalized tolerance distributionWherein:
the tolerance coefficient is calculated by the formula:
wherein, the first and the second end of the pipe are connected with each other,representing the importance weight of the node's neighbors to the node.
Further, in the sign sub-network and the disease sub-network, when the number of nodes connected with valid nodes is greater than or equal to a first preset number, the nodes are valid, otherwise, the nodes are invalid.
The invention has the beneficial effects that: the sign sub-network and the disease sub-network are used for constructing a deep cascade framework, physiological domino effect is established through a cascade failure theory, the sign sub-network and the disease sub-network are cascaded according to node failure conditions, in cascade failure, the failure conditions of the nodes, such as failure between the nodes or interaction edges of two node connections, can cause other nodes to also fail through coupling relations between the nodes, so that cascade effect is generated, therefore, the sign sub-network and the disease sub-network generate failure distribution in respective cascade stages to integrally react on failure conditions of disease types and sign types in the sub-networks, finally, the ICU death rate of a patient can be determined according to the output, and due to the interrelations between the sign types and the disease types, the death rate prediction has interpretability.
The present invention also provides a computing device, including a memory, a processor, and a computer program stored on the memory and executable on the processor, where the processor executes the computer program to implement the following steps:
acquiring an examination type of a patient, wherein the examination type comprises a disease type of the patient or an abnormal sign type in signs of examination of the patient;
generating an initial failure set according to the inspection type;
inputting the initial failure set into a deep cascading framework for ICU mortality prediction as described above for cascading;
sequentially inputting failure distribution output by each cascading stage during cascading into a serialization model;
and obtaining the mortality prediction result of the patient according to the last hidden state of the serialization model at the end of the cascade.
Further, the processor, when executing the computer program, is further configured to implement the following steps:
generating importance weights through the graph attention network;
determining the effective probability of each node of each cascade stage during cascade connection according to the importance weight;
normalizing the effective probabilities of all nodes in the cascade stage to generate the failure distribution.
The computing device of the present invention has similar benefits to the above-described deep cascading framework for ICU mortality prediction compared to the prior art, and will not be described herein again.
The invention also proposes a computer-readable storage medium, on which a computer program is stored which, when being executed by a processor, carries out the steps of:
acquiring an examination type of a patient, wherein the examination type comprises a disease type of the patient or an abnormal sign type in signs of examination of the patient;
generating an initial failure set according to the inspection type;
inputting the initial failure set into a deep cascading framework for ICU mortality prediction as described above for cascading;
sequentially inputting failure distribution output by each cascading stage during cascading into a serialization model;
and obtaining the mortality prediction result of the patient according to the last hidden state of the serialization model at the end of the cascade.
Further, the computer program, when executed by the processor, is adapted to perform the steps of:
generating importance weights through the graph attention network;
determining the effective probability of each node of each cascade stage during cascade connection according to the importance weight;
normalizing the effective probabilities of all nodes in the cascade stage to generate the failure distribution.
The advantages of the computer-readable storage medium of the present invention over the prior art are similar to the above-described deep cascading framework for ICU mortality prediction, and are not described herein again.
Drawings
FIG. 1 is a schematic representation of a prediction model for ICU mortality prediction using a deep cascade framework in accordance with the present invention;
FIG. 2 is a flow chart of a method for ICU mortality prediction in an embodiment of the present invention.
Detailed Description
In order to make the aforementioned objects, features and advantages of the present invention comprehensible, embodiments accompanied with figures are described in detail below.
It is noted that the terms first, second and the like in the description and in the claims of the present invention and in the drawings described above are used for distinguishing between similar elements and not necessarily for describing a particular sequential or chronological order. It is to be understood that the data so used is interchangeable under appropriate circumstances such that the embodiments of the invention described herein are capable of operation in other sequences than those illustrated or described herein.
Referring to fig. 1, a deep cascading framework for ICU mortality prediction includes a sign sub-network and a disease sub-network, where the sign sub-network includes a plurality of nodes corresponding to sign types, the disease sub-network includes a plurality of nodes corresponding to disease types, the sign sub-network and the disease sub-network have interaction edges, and the interaction edges are connection edges between the nodes corresponding to the disease types and the nodes corresponding to the sign types, where the sign sub-network and the disease sub-network cascade according to node failure conditions, and both the sign sub-network and the disease sub-network are used for outputting failure distributions to perform ICU mortality prediction through the failure distributions.
Medical risk detection is an important topic for improving the clinical practice capability of the ICU, and many biometric learning methods and deep learning methods based on quantitative tables and characteristics can predict the mortality of a specific patient and assist a physician to make corresponding clinical decisions. However, current methods not only require experts to manually define risk factors but also overly rely on the effectiveness of pre-training features. From the model perspective, the end-to-end prediction model lacks interpretability, and the black box type neural network structure is difficult to obtain an inference path.
In embodiments of the invention, the body system is considered to be a complex dynamic network, the interaction between physiological functions maintaining the balance of the network. Once a part or a function of the physiological network fails, the steady state is broken, resulting in a more serious and dangerous functional failure, even a breakdown of the entire physiological network until death, for example: complications, infections, immune disorders, and the like.
Abnormalities in vital signs are the internal cause of disease, while diseases are the external manifestations of abnormal signs. One disease is indicative of some related signs being marginal or abnormal, and if the severity of these abnormal signs exceeds a tolerable threshold for other diseases, more physiological functional failure can occur, even resulting in a cascade. Therefore, there is a typical interaction between signs and diseases, which together maintain the balance of the physiological system. Based on this, the deep cascade framework for ICU mortality prediction of the present application, hereinafter referred to as deep cascade framework, includes: sign subnetworkAnd disease sub-networkSign subnetworkComprising a plurality of nodes representing types of physical signs, a disease sub-networkThe node comprises a plurality of disease types, wherein the sign types are basic physiological characteristics of human beings, such as urine osmotic pressure, plasma albumin parameters and the like, and correspondingly, the disease types are external manifestations of abnormal signs, such as renal failure disease characteristics related to the urine osmotic pressure and the plasma albumin parameters and the like. Thus a two-part network is constructed, in whichAndindividual disease sub-networksSign and sign sub-networkThe set of nodes of (a) is,is a collection of interacting edges between two sub-networks,representing a set of triplesWeight, sign sub-network ofAnd disease sub-networksWith interacting edges, i.e. connecting edges of nodes of disease type and nodes of sign type, i.e. representing the interrelation between two sub-networks, for sign sub-networksAnd disease sub-networkAnd the nodes in the network have no edge connection, and only the sign sub-network and the disease sub-network interact with each other. On-sign subnetworkIn (1), failure of a certain node representing the type of physical sign may cause disease, and the corresponding disease sub-networkOne or more nodes representing the disease type may fail, and correspondingly, the disease subnetworkIn (1), failure of a certain node representing the disease type may result in a sign sub-networkWherein one or more nodes representing a type of a sign may fail, thereby producing a chain reaction, and thus referring to fig. 1, the sign subnetworkAnd disease sub-networkInteract with each other to generate cascade failure, and in time period, sign sub-networkAnd disease sub-networksThe method has a plurality of cascading stages, and in each cascading process, the failure condition of a node is used as a main factor influencing the subsequent cascading stage.
Thus, the present embodiment passes through the sign subnetworkAnd disease sub-networkConstructing deep cascadesFramework, physiological domino effect and physical sign sub-network established by cascade failure theoryAnd disease sub-networkCascading is carried out according to node failure conditions, in cascading failure, the failure conditions of nodes, such as the failure between the nodes or the interaction edges of the connection of the two nodes, can cause other nodes to also fail through the coupling relation between the nodes, so that a cascading effect is generated, and therefore, in each stage of cascading by utilizing a deep cascading framework, a physical sign sub-network can be outputAnd disease sub-networkIn this embodiment, in particular, the sign subnetworkAnd disease sub-networkGenerating failure distributions in respective cascade stages to integrally react to failure conditions of disease types and sign types in respective sub-networks, and finally determining the ICU death rate of the patient according to the output, in an optional embodiment of the present embodiment, as shown in FIG. 1, when ICU death rate prediction is performed, the failure distributions are processed through a serialization model, and based on the sign sub-networks in time periodOr disease subnetworkAlternately cascading, and outputting fail-overCloth, sign subnetworkFailure distribution and disease subnetwork ofThe failure distribution is integrally input into the serialization model to be processed so as to control the input dimensionality to be unchanged, and the final hidden state of the serialization model can be classified by a classifier to directly obtain an ICU mortality prediction result. Thus, based on the deep cascade framework constructed in the embodiment, when applied to ICU mortality prediction, the potential failure risk of all physiological functions in each clinical stage can be predicted. In contrast to other feature-based or scale-based models, the deep cascade framework in this embodiment, when applied to ICU mortality prediction, the sign subnetworkAnd disease sub-networkThe nodes which are mutually influenced mutually influence each other to carry out cascade failure and output failure distribution for ICU mortality prediction, the output of a deep cascade framework and the cascade process have strong corresponding relations and are different from a black box type neural network structure, and an inference path can be obtained by applying the deep cascade framework in the embodiment to carry out ICU mortality prediction, so that the deep cascade framework has the characteristic of interpretability, is suitable for various prediction tasks and can be continuously learned from medical common sense and clinical experience.
The interaction between the physical signs and the physiological function of the disease can be learned or summarized from medical practice.
In an alternative embodiment of the invention, in the sign subnetworkAnd disease sub-networkWhen the number of the nodes connected with the effective nodes is larger than or equal to a first preset number, the nodes are effective, otherwise, the nodes are invalid.
For a sub-network, if a node in the sub-networkNumber of active nodes connected to another sub-networkGreater than or equal to a first preset numberThen degree isNode (a) ofIs set toI.e. indicating that the node remains active, otherwise the state is set toI.e. indicating node failure, to judge the validity and failure of the node in the sub-network.
Thus, for the nodeFormalizing nodesTolerance coefficient ofIs a conditional probabilityAnd determining the failure risk of the nodes in the sub-network, wherein the piecewise function expression is as follows:
In an optional embodiment of the invention, the node failure condition comprises a failure set, and the failure set comprises the nodes before the maximum failure risk in the cascade stagenAnd (4) each node.
In ICU mortality prediction, a sub-network based on signs is requiredAnd disease sub-networkThe overall failure condition of (a) is judged for mortality. In an alternative embodiment, in the cascade stage, the first stage will benAnd adding the node with the maximum failure risk into the failure set input in the current cascade stage to update the failure set to be output. Referring to fig. 1, when the mortality of an ICU is predicted through a deep cascade framework, an initial failure set is usually input, the initial failure set includes a disease detection type or a sign abnormality type when the ICU is input, and in the first cascade process, the failure set output at this stage is added to the initial failure set in the current cascadenThe node with the largest failure risk fails in the subsequent cascade processThe set is updated in this way and is used as input in the later cascading stages, so that the failure condition of the whole in each cascading stage can be finally determined through the failure set, such as determining the failure distribution of the sub-networks.
In an optional embodiment of the present invention, the node failure condition further includes network connectivity, the network connectivity includes effective edge probability of the interaction edge, the effective edge probability represents a probability that one interaction edge is randomly selected and the edge connects two effective nodes, and the effective edge probability generated in the current cascade stage is used to update the effective edge probability in the next cascade stage, so as to update the effective edge probability in the sign sub-networkAnd disease sub-networkIn the cascade process, the sign sub-network is measured by the effective edge probabilityAnd disease sub-networkConnectivity of the constructed physiological network.
In an optional embodiment of the present invention, the sign sub-network and the disease sub-network each include a degree distribution unit and an excess degree distribution unit, and both the sign sub-network and the disease sub-network determine the failure set through the degree distribution unit, and both the sign sub-network and the disease sub-network determine the effective edge probability of the next cascade stage according to the effective edge probability of the current cascade stage through the excess degree distribution unit.
When one of two connected nodes fails, the interaction edge between them will also be removed, for example: the kidney loses the physiological function of detoxification due to the occurrence of renal failure, and then the supply interaction of urine osmotic pressure and physical stability of plasma albumin is no longer realized, and the cascade frequency is changedWill continuously impair the robustness of the physiological network, and therefore, in this embodiment, referring to fig. 1, the sign subnetworkAnd disease sub-networkThe degree distribution generating unit based on the degree distribution generating function and the excess degree distribution generating unit based on the excess degree distribution generating function are both arranged in the computer. Sign subnetworkAnd disease sub-networkEach cascade stage sign sub-network is measured and determined by a degree distribution generating functionAnd disease sub-networksThe effective probability distribution of the middle nodes, and further the node with the maximum failure risk can be determined through the effective probability of the effective nodes, so as to determine the failure setnAdding the node with the maximum failure risk into the failure set input in the current cascade stage to update the failure set, and correspondingly, adding the sign sub-network into the failure setAnd disease sub-networkThe effective edge probability is determined through an excess distribution generating function, so that the connectivity of the network is determined, and the robustness is improved.
In an alternative embodiment of the invention, the sign subnetworkAnd disease sub-networkDegree distribution generating function ofThe expression of (a) is:
whereinXRepresenting the sign sub-networkOr the disease subnetwork,Is thatThe degree distribution of the network is such that,the probability that an interaction edge is randomly selected and the edge connects two effective nodes, namely the effective edge probability, is used for measuring the effective probability distribution of each cascade stage node by using a degree distribution generating function.
Excess distribution generating functionIs a function of generating a distribution of degree of contrastAccording to the effective edge probability of the current cascade stageCalculating effective edge probability of network of next stageOver-distribution generating functionThe expression of (a) is:
wherein, the first and the second end of the pipe are connected with each other,representing sub-networksXSo as to measure the connectivity of the network by using the over-distribution generating function.
Thus, in an alternative embodiment of the present invention, the recursive relationship of the depth cascade framework comprises:
,representing the probability of valid nodes, inlIn the secondary cascade, the effective node proportions of the disease sub-network and the sign sub-network are respectively as follows:(ii) a Or the like, or, alternatively,
in the first placelIn the secondary cascade, the effective node proportions of the sign sub-network and the disease sub-network are respectively as follows:;
wherein the content of the first and second substances,a sub-network of a disease is represented,a sub-network of signs is represented,anda degree distribution generating function representing the disease sub-network and the signs sub-network respectively,anda hyperdynamic distribution generating function representing the disease sub-network and the signs sub-network respectively,lthe number of cascades is indicated,represents the number of cascades asl-1 a hypernym distribution generation function of the vital sign sub-network,represents the number of cascades asl-1 a hyperdistribution generating function of the disease sub-network,represents the number of cascades aslOver-distribution generation function of the disease sub-network,represents the number of cascades aslThe superscale distribution of sign sub-networks of (a) to (b) generate a function,represents the number of cascades aslA function is generated of the degree distribution of the disease sub-network,represents the number of cascades aslThe degree distribution of the physical sign sub-networks of (1) generates a function,the probability of the valid edge is represented,represents the number of cascades aslOf the disease sub-network,represents the number of cascades aslThe effective edge probability of the sign subnetwork.
It will be appreciated that in the deep cascade framework described in this embodiment, cascades may be derived from disease subnetworksStarting or starting from a sign subnetworkStarting with sub-networks, starting with cascades of different sub-networks, with some distinction in cascade recursion relationshipsInitially, the recursive relationship for the depth cascade framework includes:in the first placelIn the secondary cascade, the disease subnetworkAnd the sign subnetworkThe effective node ratios of (1) are respectively as follows:。
but in the sign sub-networkAt the beginning, the recursive relationship of the depth cascade framework includes:in the first placelIn the secondary cascade, the sign subnetworkAnd the disease subnetworkThe effective node ratios are respectively as follows:。
generally, patients usually suffer from a small number of diseases at the beginning of their hospitalization, so in this embodiment, the sub-networks of diseases are cascadedStarting with an example, for a disease sub-networkNode set ofRemoving a proportion of disease nodes, e.g., removing a proportion ofCorrespondingly, another sign subnetworkIs not influenced by cascade connection in the first cascade connection stage, and is integratedThe remaining effective node proportionWill decide on the current disease sub-networkAnd sign subnetworkConnectivity of two physiological networks formed, in which case the effective edge probabilityGenerating a degree distribution into a functionApplication to sign subnetworkThen sign subnetworkThe ratio of the middle effective node will be And (6) updating. Since the disease subnetwork at this timeThe network connectivity cannot be reflected according to the proportion of the effective nodes because the network connectivity is destroyed by the initial failure nodes, that is, the inequality exists:. At this time, according to the physical meaning of the super distribution, there isThus, it can be determined that in the present embodiment, slave disease subnetworks are cascadedThe recursive relation of the recursion starts. Correspondingly, this determination can be made to cascade slave sign subnetworksThe recursive relation of the recursion is started to achieve the corresponding technical effect, and the description is omitted here.
In an optional embodiment of the present invention, the sign subnetwork and the disease subnetwork both determine the failure distribution by the effective probabilities of the nodes, wherein the failure distribution of the current cascade stage is generated by normalizing the effective probabilities of all nodes of the cascade stage.
Efficient probability means that there are neighbors of a nodemThe probability that an individual node has failed, itself remains valid.
In this embodiment, when the mortality is predicted by using the deep cascade framework, in each cascade stage, the deep cascade framework outputs failure distribution for mortality prediction, the failure distribution is generated by normalizing the effective probability of all nodes, and the effective probability is calculated by the following formula:;
wherein the content of the first and second substances,the nodes are represented as a list of nodes,mindicating sectionThe number of failed nodes in the neighborhood of a point,which represents the number of neighbors of the node,indicating the number of valid neighbors of the node,it is shown that the distribution is tolerated,the value of the tolerance factor is represented,Xrepresenting the sign sub-networkOr the disease subnetwork。
In this embodiment, in the cascade stage, the effective probabilities of all the nodes are obtainedThen, all the effective probabilities can be matchedNormalization is performed so that a failure distribution can be obtainedAnd further as output of the deep cascade framework for ICU mortality prediction, see fig. 1, in the deep cascade framework, the sign subnetworkAnd disease sub-networkCascading and outputting failure distributionAnd failure distributionTo the serialization model, as sign subnetworkAnd disease sub-networkAre all carried outnWhen the secondary cascade then ends, the serialization model receives 2nAnd distributing the failures, and classifying the last hidden state of the serialized model through a classifier to obtain an ICU mortality prediction result.
In an alternative embodiment of the invention, each node of the signs sub-network and the diseases sub-network has a personalized tolerance distributionWherein:
the tolerance coefficient is calculated by the formula:
wherein the content of the first and second substances,representing the importance weight of the node's neighbors to the node.
Degree distribution generating functionThe method can be used for measuring the proportion of effective nodes in the interaction network in each cascade stage, and because each node has difference, the uniform tolerance coefficient is adopted for each node with the same degreeThis can lead to inaccuracies and heterogeneity that violates physiological functions. For example, assume that at some cascade stage a diabetic node and a hypertensive node have the same number of neighborsAnd the same number of valid neighborsThen both disease nodes have the same risk of failure when cascading without regard to node variability. However, since the two physiological functions differ in risk tolerance and pathological characteristics, the risk of failure is significantly different between the two.
Therefore, in the embodiment, as for the degree distribution generating function and the over-degree distribution generating function, the degree distribution generating function is generated by fusing the node characteristicsAnd an over-distribution generating functionConversion to node personalization form, generating function for degree distributionAnd an over-distribution generating functionDecomposing degree distribution in functionWhereinIs at a degree ofkThe set of nodes of (a) is,thus, the two generating functions can be redefined as:
in this embodiment, a personalized tolerance profile is generated for each nodeTolerant of distributionIs a tolerance factorWith the effective number of neighbor nodes of the nodeIncrease of (2), tolerance factor thereofWill also rise, provided thatIs a nodeIs/are as followsThe neighbor nodes which have failed are provided with mutual independence assumption among the neighbors in the bipartite network, so that the nodesEqual to each neighbor node and nodeSum of causal probabilities between, i.e.And satisfyTherefore, for the tolerance coefficient, the calculation formula is:
distinguished from tolerance factorThe calculation method of the piecewise, linear or polynomial function, the tolerance coefficient in this embodimentConversion from simple quantitative statistics to AND-nodesThereby more accurately confirming the failure risk of the sub-network.
Wherein, referring to FIG. 1, the importance weight may be generated by a graph attention network (GAT)Thereby facilitating reflection of the interaction weights of a pair of nodes.
The invention also provides an ICU mortality prediction method, which comprises the following steps:
s1, obtaining the examination type of the patient, wherein the examination type comprises the disease type of the patient or the abnormal sign type in the signs of the examination of the patient;
s2, generating an initial failure set according to the check type;
s3, inputting the initial failure set into the deep cascading framework for cascading;
s4, sequentially inputting failure distribution output by each cascade stage during cascade connection into a serialization model;
and S5, obtaining a mortality prediction result of the patient according to the last hidden state of the serialization model at the end of the cascade.
In this embodiment, when the patient's ICU mortality is predicted according to the deep cascade framework, the examination type of the patient is obtained, the examination type may include a disease diagnosis type when entering the ICU or a sign examination type within a certain time after entering the ICU, and for a certain patient, there is usually a certain disease when entering the ICU, and a part with abnormality in the sign type can be detected, and correspondingly, the part with abnormality or a disease item can be used as a sign sub-networkOr disease subnetworkThe initial failure node is used to construct the initial failure set and serve as the initial input of the deep cascading framework, referring to fig. 1, in this embodiment, the sign sub-network is used to cascade failureIn the first cascade, the patient is usually suffering from the initial hospitalizationA small number of diseases, hence the disease sub-networkA certain proportion of disease nodes need to be removed, e.g., the removal proportion isCorrespondingly, another sign subnetworkThe disease subnetwork will not be affected by cascade in the first cascade stageRemaining effective node proportionWill decide the current disease sub-networkSign and sign sub-networkConnectivity of two physiological networks formed, in which case the effective edge probabilityWhereby said physical sign subnetworkAnd cascading according to the initial failure set.
Referring to FIG. 1, the present embodiment utilizes the above described deep cascading framework for ICU mortality prediction, in which cascading failures are in sign sub-networksInitially, at this point, an initial failure set is entered, which includes abnormalities in the signs corresponding to the patient examinationIs determined, whereby the initial failed node passes through the sign subnetworkDegree distribution unit and excess distribution unit to obtain maximum failure risknEach node updates the failure set to obtain the effective edge probability representing the connectivity of the network at the next stage, and the failure set and the effective edge probability are used as the second cascade of the whole deep cascade framework, namely the disease subnetworkThe first cascade of input, disease subnetworksCascading is also performed according to the input to obtain a failure set and effective edge probability for the next cascading stage, so that the failure set and the effective edge probability are input into the sign sub-network againTo perform the physical sign sub-networkThereby carrying out sign subnetworkAnd disease sub-networksAnd, during the cascade, sign subnetworksAnd disease sub-networkThe failure distributions are output in their respective cascade stages as outputs of a deep cascade framework, in particular, in fig. 1, the sign subnetworkFirst cascade output failure distributionAnd sign subnetworkAnd disease sub-networkAre all carried outlFailure distribution of output after secondary cascadeAndand with disease subnetworksCascade connectionnFailure distribution of output at sub-endDue to physical signs subnetworkAnd disease sub-networkAlternately cascaded, the cascaded outputs of each stage are also input into the serialized neural network as a whole, corresponding to the serialized neural network in FIG. 1h1…h2l-1、h2l…h2nThe input of each stage is processed by a serialized neural network, and the final hidden state can obtain the result of ICU mortality prediction by a classifier.
When the disease diagnosis type is used as the examination type, the disease diagnosis type is used as the initial failure set, i.e., the diagnosis type is used as the initial failure setCorrespondingly, when the sign inspection type is taken as the inspection type, the abnormal item of the sign inspection type can be taken as an initial failure set.
In the cascade process, disease subnetworksAnd sign subnetworkOutput failure distributionAndin which the failure distribution is determined by the effective probability of all nodes in the cascade stageIs normalized based on the disease sub-networkSign and sign sub-networkCascaded in sequence, failure distributionAndreferring to fig. 1, that is, in the input serialized neural network, the serialized model may use transform, LSTM, or GRU, and when the number of cascading times reaches a set threshold or no node fails, it indicates that the cascading is finished, and when the cascading is finished, the last hidden state of the serialized model may pass through a classifier, such as a softmax classifier, so as to obtain the patient suffering from the diseaseICU mortality prediction in the subject.
In an alternative embodiment of the present invention, the ICU mortality prediction method further comprises:
generating importance weights through the graph attention network;
determining the effective probability of each node of each cascade stage during cascade connection according to the importance weight;
normalizing the effective probabilities of all nodes in the cascade stage to generate the failure distribution.
Referring to fig. 1, in the present embodiment, importance weights are generated by the graph attention network to generate personalized tolerance distributions for each node, as in fig. 1, with sign sub-networksThe cascade is started, and the sign sub-networkPerforming various stages of the cascade, the graph attention network each generating weights for generating individualized tolerance distributions with tolerance coefficientsIn the disease subnetworkIn each stage of the cascade, the graph attention network generates weights for generating personalized tolerance distributions with tolerance coefficientsThus, to physical sign sub-networkAnd disease sub-networkThe determination of the effective probabilities of the nodes of (A) each have a corresponding tolerance coefficient, e.g.Figure 1 shows a sign subnetworkNo. 1 …lSub-cascaded stages, each having a tolerance factor…Etc., and a disease sub-network as shown in figure 1To (1) al…nSub-cascaded stages, each having a tolerance factor…Etc. it should be noted that, in fig. 1, the sub-network is due to physical signsAnd disease sub-networkAlternatively cascaded, thus in a sign subnetwork 1 st 1 …lBetween the secondary cascade stages, there is also a 1 st 1 … not shown in fig. 1lSub-disease sub-networkCorresponding to the 1 st 1 … not shown in fig. 1lSub-disease sub-networkHas tolerance coefficients corresponding to the cascade stages of (2), and in addition, in a disease subnetworkTo (1) al…nBetween the secondary cascade stages, there is also a second stage which is not shown in FIG. 1l…nSub-sign sub-networkWhile also corresponding to the second stage not shown in fig. 1l…nSub-sign sub-networkThe tolerance coefficient of the tolerance distribution is converted from simple quantitative statistics to the nodeThereby more accurately confirming the failure risk of the sub-network.
The present invention also provides a computing device, including a memory, a processor, and a computer program stored on the memory and executable on the processor, where the processor executes the computer program to implement the following steps:
acquiring an examination type of a patient, wherein the examination type comprises a disease type of the patient or an abnormal sign type in signs of examination of the patient;
generating an initial failure set according to the inspection type;
inputting the initial failure set into a deep cascading framework for ICU mortality prediction as described above for cascading;
sequentially inputting failure distribution output by each cascading stage during cascading into a serialization model;
and obtaining the mortality prediction result of the patient according to the last hidden state of the serialization model at the end of the cascade.
In an alternative embodiment of the present invention, when the processor executes the computer program, the processor is further configured to implement the following steps:
generating importance weights through the graph attention network;
determining the effective probability of each node of each cascade stage during cascade connection according to the importance weight;
normalizing the effective probabilities of all nodes in the cascade stage to generate the failure distribution.
The computing device of this embodiment is similar to the above-described deep cascading framework for ICU mortality prediction in comparison with the prior art, and will not be described herein again.
A computer-readable storage medium of another embodiment of the invention, having a computer program stored thereon, which, when executed by a processor, performs the steps of:
acquiring an examination type of a patient, wherein the examination type comprises a disease type of the patient or an abnormal sign type in signs of examination of the patient;
generating an initial failure set according to the inspection type;
inputting the initial failure set into a deep cascading framework for ICU mortality prediction as described above for cascading;
sequentially inputting failure distribution output by each cascading stage during cascading into a serialization model;
and obtaining the mortality prediction result of the patient according to the last hidden state of the serialization model at the end of the cascade.
In an alternative embodiment of the invention, the computer program, when executed by the processor, is further adapted to perform the steps of:
generating importance weights by the graph attention network;
determining the effective probability of each node of each cascade stage during cascade connection according to the importance weight;
normalizing the effective probabilities of all nodes in the cascade stage to generate the failure distribution.
The advantages of the computer-readable storage medium of this embodiment over the prior art are similar to the above-described deep cascading framework for ICU mortality prediction, and are not described herein again.
Although the present disclosure has been described above, the scope of the present disclosure is not limited thereto. Various changes and modifications may be made by one skilled in the art without departing from the spirit and scope of the disclosure, and it is intended that all such changes and modifications fall within the scope of the invention.
Claims (13)
1. A deep cascade framework for ICU mortality prediction comprising: the sign sub-network and the disease sub-network are connected in a cascade mode according to node failure conditions, and the sign sub-network and the disease sub-network are both used for outputting failure distribution so as to conduct ICU mortality prediction through the failure distribution;
the sign sub-network and the disease sub-network both comprise a degree distribution unit and a super degree distribution unit, the degree distribution unit comprises a degree distribution generating function, the super degree distribution unit comprises a super degree distribution generating function, and the recursive relation of the deep cascade framework comprises:
,represents the effective node ratio, inlIn the secondary cascade, the effective node proportions of the disease sub-network and the sign sub-network are respectively as follows:(ii) a Or the like, or, alternatively,
in the first placelIn the secondary cascade, the effective node proportions of the sign sub-network and the disease sub-network are respectively as follows:;
wherein the content of the first and second substances,a sub-network of diseases is represented,a sub-network of physical signs is represented,anda degree distribution generating function representing the disease sub-network and the signs sub-network respectively,anda hyperdynamic distribution generating function representing the disease sub-network and the signs sub-network respectively,lthe number of cascades is indicated,represents the number of cascades asl-1 an overarching distribution generation function of the sub-network of signs,represents the number of cascades asl-1 a hyperdistribution generating function of the disease sub-network,represents the number of cascades aslGenerating a function of the hyperdistribution of the disease sub-network,represents the number of cascades aslThe superscale distribution of sign sub-networks of (a) to (b) generate a function,represents the number of cascades aslA function is generated of the degree distribution of the disease sub-network,represents the number of cascades aslThe degree distribution of the sign sub-network of (1) generates a function,the probability of a valid edge is represented,represents the number of cascades aslOf the disease sub-network,represents the number of cascades aslThe effective edge probability of the sign subnetwork.
2. The deep cascading framework for ICU mortality prediction of claim 1, wherein the node failure condition comprises a failure set comprising pre-stage failures with the greatest risk in cascading stagesnAnd (4) each node.
3. The deep cascading framework for ICU mortality prediction of claim 2, wherein the node failure condition further comprises network connectivity comprising effective edge probabilities of the interacting edges, the effective edge probabilities representing a probability of randomly selecting an interacting edge and connecting two effective nodes, the effective edge probabilities generated in a current cascading stage being used to update the effective edge probabilities in a next cascading stage.
4. The deep cascading framework for ICU mortality prediction of claim 3, wherein the sign sub-network and the disease sub-network each determine the failure set through a degree distribution unit, the sign sub-network and the disease sub-network each determine a valid edge probability of a next cascade stage according to a valid edge probability of a current cascade stage through a super degree distribution unit.
5. The deep cascading framework for ICU mortality prediction according to any of claims 2-4, wherein the sign subnetwork and the disease subnetwork each determine the failure distribution by an effective probability of a node, wherein the failure distribution of a current cascade stage is generated by normalizing the effective probabilities of all nodes of the cascade stage.
6. The deep cascading framework for ICU mortality prediction of claim 5, wherein in the signs subnetwork and the diseases subnetwork, the significance probability of the node is determined according to the number of failed nodes in each neighborhood of nodes, wherein the failure set generated in the current cascading stage is used to update the number of failed nodes in each neighborhood of nodes in the next cascading stage.
7. The deep cascading framework for ICU mortality prediction of claim 6, wherein the effective probability of the node is calculated by:
wherein the content of the first and second substances,a node is represented by a plurality of nodes,mrepresenting the number of failed nodes in the neighborhood of nodes,which represents the number of neighbors of the node,indicating the number of valid neighbors of the node,it is shown that the distribution is tolerated,which represents the tolerance factor of the current signal,Xrepresenting the signs subnetwork or the diseases subnetwork.
8. The deep cascade framework for ICU mortality prediction of claim 7, wherein each node of the signs sub-network and the diseases sub-network has a personalized tolerance distributionWherein:
the tolerance coefficient is calculated by the formula:
9. A deep cascading framework for ICU mortality prediction according to any of claims 1-4, wherein in the signs sub-network and the diseases sub-network, a node is active when the number of nodes connecting active nodes is greater than or equal to a first preset number, and otherwise the node is inactive.
10. A computing device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, the processor implementing the following steps when executing the computer program:
acquiring an examination type of a patient, wherein the examination type comprises a disease type of the patient or an abnormal sign type in signs of examination of the patient;
generating an initial failure set according to the inspection type;
cascading said initial failure set into the deep cascading framework for ICU mortality prediction of any of claims 1-9;
sequentially inputting failure distribution output by each cascading stage during cascading into a serialization model;
and obtaining the mortality prediction result of the patient according to the last hidden state of the serialization model at the end of the cascade.
11. The computing device of claim 10, wherein the processor, when executing the computer program, is further configured to:
generating importance weights through the graph attention network;
determining the effective probability of each node of each cascade stage during cascade connection according to the importance weight;
normalizing the effective probabilities of all nodes in the cascade stage to generate the failure distribution.
12. A computer-readable storage medium, on which a computer program is stored, which computer program, when being executed by a processor, carries out the steps of:
acquiring an examination type of a patient, wherein the examination type comprises a disease type of the patient or an abnormal sign type in signs of examination of the patient;
generating an initial failure set according to the inspection type;
cascading said initial failure set into the deep cascading framework for ICU mortality prediction of any of claims 1-9;
sequentially inputting failure distribution output by each cascading stage during cascading into a serialization model;
and obtaining the mortality prediction result of the patient according to the last hidden state of the serialization model at the end of the cascade.
13. The computer-readable storage medium according to claim 12, wherein the computer program, when executed by the processor, is further configured to perform the steps of:
generating importance weights by the graph attention network;
determining the effective probability of each node of each cascade stage during cascade connection according to the importance weight;
normalizing the effective probabilities of all nodes in the cascade stage to generate the failure distribution.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202011341957.4A CN112420201B (en) | 2020-11-25 | 2020-11-25 | Deep cascading framework for ICU mortality prediction |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202011341957.4A CN112420201B (en) | 2020-11-25 | 2020-11-25 | Deep cascading framework for ICU mortality prediction |
Publications (2)
Publication Number | Publication Date |
---|---|
CN112420201A CN112420201A (en) | 2021-02-26 |
CN112420201B true CN112420201B (en) | 2022-09-30 |
Family
ID=74843546
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202011341957.4A Active CN112420201B (en) | 2020-11-25 | 2020-11-25 | Deep cascading framework for ICU mortality prediction |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN112420201B (en) |
Families Citing this family (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN112967816B (en) * | 2021-04-26 | 2023-08-15 | 四川大学华西医院 | Acute pancreatitis organ failure prediction method, computer equipment and system |
Family Cites Families (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106777935A (en) * | 2016-12-05 | 2017-05-31 | 广东石油化工学院 | A kind of disease dynamic prediction method based on network structure |
CN107463796B (en) * | 2017-07-12 | 2019-10-18 | 北京航空航天大学 | Early stage virulence factor detection method based on gene co-expressing Internet communication analysis |
CN107491638A (en) * | 2017-07-28 | 2017-12-19 | 深圳和而泰智能控制股份有限公司 | A kind of ICU user's prognosis method and terminal device based on deep learning model |
CN109119155B (en) * | 2018-07-03 | 2022-01-28 | 厦门大学 | ICU death risk assessment system based on deep learning |
-
2020
- 2020-11-25 CN CN202011341957.4A patent/CN112420201B/en active Active
Also Published As
Publication number | Publication date |
---|---|
CN112420201A (en) | 2021-02-26 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Sailasya et al. | Analyzing the performance of stroke prediction using ML classification algorithms | |
Haq et al. | A hybrid intelligent system framework for the prediction of heart disease using machine learning algorithms | |
Ismaeel et al. | Using the Extreme Learning Machine (ELM) technique for heart disease diagnosis | |
Luengo et al. | A study on the use of statistical tests for experimentation with neural networks: Analysis of parametric test conditions and non-parametric tests | |
Nour et al. | Automatic classification of hypertension types based on personal features by machine learning algorithms | |
Ziasabounchi et al. | ANFIS based classification model for heart disease prediction | |
US11600387B2 (en) | Control method and reinforcement learning for medical system | |
Ranganath et al. | Multiple causal inference with latent confounding | |
Ribeiro et al. | Multi-step ahead meningitis case forecasting based on decomposition and multi-objective optimization methods | |
WO2011060314A1 (en) | Method and system for optimal estimation in medical diagnosis | |
Anouncia et al. | Design of a diabetic diagnosis system using rough sets | |
CN112420201B (en) | Deep cascading framework for ICU mortality prediction | |
Azar et al. | Linguistic hedges fuzzy feature selection for differential diagnosis of Erythemato-Squamous diseases | |
Santhanam et al. | Machine learning and blood pressure | |
Jahangir et al. | Auto-MeDiSine: an auto-tunable medical decision support engine using an automated class outlier detection method and AutoMLP | |
Fatemidokht et al. | Development of a hybrid neuro-fuzzy system as a diagnostic tool for Type 2 Diabetes Mellitus | |
Roversi et al. | A Dynamic Bayesian Network model for simulating the progression to diabetes onset in the ageing population | |
Prasad et al. | Chronic Kidney Disease Risk Prediction Using Machine Learning Techniques | |
Qiu et al. | HFS‐LightGBM: A machine learning model based on hybrid feature selection for classifying ICU patient readmissions | |
Kubus et al. | The use of fuzzy cognitive maps in evaluation of prognosis of chronic heart failure patients | |
Karahoca et al. | Diagnosis of diabetes by using adaptive neuro fuzzy inference systems | |
Muibideen et al. | A Fast Algorithm for Heart Disease Prediction using Bayesian Network Model | |
Abdellatif et al. | Computational detection and interpretation of heart disease based on conditional variational auto-encoder and stacked ensemble-learning framework | |
WO2021196239A1 (en) | Network representation learning algorithm across medical data sources | |
Nisha et al. | Interpretable Machine Learning Models for Assisting Clinicians in the Analysis of Physiological Data. |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant | ||
TR01 | Transfer of patent right |
Effective date of registration: 20230918 Address after: Building 1, Kechuang headquarters, Shenzhen (Harbin) Industrial Park, 288 Zhigu street, Songbei District, Harbin City, Heilongjiang Province Patentee after: Harbin Institute of Technology Institute of artificial intelligence Co.,Ltd. Address before: 150001 No. 92 West straight street, Nangang District, Heilongjiang, Harbin Patentee before: HARBIN INSTITUTE OF TECHNOLOGY |
|
TR01 | Transfer of patent right |