CN115274132A

CN115274132A - Respiratory infectious disease monitoring and early warning system and method

Info

Publication number: CN115274132A
Application number: CN202210832406.0A
Authority: CN
Inventors: 陈航; 蒋荣猛; 韩冰
Original assignee: Beijing Ditan Hospital
Current assignee: Beijing Ditan Hospital
Priority date: 2022-07-15
Filing date: 2022-07-15
Publication date: 2022-11-01

Abstract

The invention provides a monitoring and early warning system and a method for respiratory infectious diseases, which comprises a community monitoring and moving terminal subsystem, a hospital terminal individual monitoring subsystem and a group monitoring and early warning subsystem; the community monitoring mobile terminal subsystem, the hospital terminal individual monitoring subsystem and the group monitoring early warning subsystem all comprise the following layer structures: a source data layer for acquiring source data; the basic resource layer is used for providing data encryption service and calling MySQL, docker and HyperLegger Fabric according to data encryption requirements; the data service layer is used for providing query service and alarm service and carrying out statistical analysis on the source data through a knowledge graph and a rule engine; storing a rule base in the data service layer; and the data visualization layer is used for visualizing the data, calling an alarm rule to alarm according to the visualized data, and managing the alarm rule according to the requirement. The invention realizes the aims of information source multi-channel, monitoring and early warning intellectualization and multipoint triggering so as to improve the emergency handling capacity for dealing with the emergent public health events.

Description

Respiratory infectious disease monitoring and early warning system and method

Technical Field

The invention belongs to the technical field of intelligent monitoring, and particularly relates to a monitoring and early warning system and method for respiratory infectious diseases.

Background

A knowledge graph was first proposed in 2012 by Google to describe the information results that its search engine obtained from different sources. The resulting nature of these messages is a Multi-relational Graph (Multi-relational Graph) consisting of different messages (nodes) + relations (edges).

The knowledge graph is logically divided into a data layer and a mode layer, and the data layer is used for storing real data. The mode layer is above the data layer and is the core of the knowledge graph, and the refined knowledge is stored. The construction of the knowledge graph is an iterative updating process, and each iteration comprises three stages according to the logic of knowledge acquisition: extracting information, namely extracting entities, attributes and interrelations among the entities from various data sources, and forming ontology knowledge expression on the basis; knowledge fusion, after obtaining new knowledge, needs to integrate it to eliminate contradictions and ambiguities, for example, some entities may have multiple expressions, a certain name may correspond to multiple different entities, etc.; and in the knowledge processing, for the new fused knowledge, the qualified part can be added into the knowledge base after quality evaluation (part of the new fused knowledge needs to be manually screened) so as to ensure the quality of the knowledge base.

In the prior art, a new coronary pneumonia multipoint triggering, monitoring and early warning response system automatically reports basic information of people with early symptoms such as fever through mobile APP, wherein the basic information comprises basic information and symptoms of people. Early warning administrators are established in disease prevention control centers in counties, cities and provinces, and reported information is analyzed and disposed in time. But the method can not solve the problems that the information filling content of the crowd needing key monitoring in the community is not comprehensive, the working cost is high, the efficiency is low, and the early monitoring and early warning capability of the community can not be improved

In view of the above, a monitoring and early warning system and method for respiratory infectious diseases are needed to be provided.

Disclosure of Invention

Therefore, the technical problem to be solved by the present invention is to provide a monitoring and early warning system and method for respiratory infectious diseases, which achieve the objectives of multi-channel information sources, intelligent monitoring and early warning, and multi-point triggering, so as to improve the emergency handling capability for dealing with emergency public health events.

The invention discloses a monitoring and early warning system for respiratory infectious diseases, which comprises:

the system comprises a community monitoring mobile terminal subsystem, a hospital terminal individual monitoring subsystem and a group monitoring early warning subsystem;

the community monitoring mobile terminal subsystem, the hospital terminal individual monitoring subsystem and the group monitoring early warning subsystem all comprise the following layer of architectures:

a source data layer for acquiring source data;

the basic resource layer is used for providing data encryption service and calling MySQL, docker and HyperLegendric Fabric according to data encryption requirements;

the data service layer is used for providing query service and alarm service and carrying out statistical analysis on the source data through a knowledge graph and a rule engine; a rule base is stored in the data service layer;

and the data visualization layer is used for visualizing the data, calling the alarm rule to alarm according to the visualized data, and managing the alarm rule according to the requirement.

Furthermore, the community monitoring terminal subsystem realizes the self-filling of basic information of community residents, community health monitoring groups and major monitoring industry workers, wherein the basic information comprises basic information of the workers, basic diseases, epidemiological history, symptoms and the like;

after the information is submitted, the system automatically feeds back diagnosis and treatment suggestions and protection suggestions according to the knowledge graph and the rule engine diagnosis;

and after data desensitization, encrypting and sending the data to the group monitoring and early warning subsystem.

Furthermore, the individual monitoring system at the hospital end is integrated with an HIS system and a PACS system in the hospital, diagnosis and treatment information of the patient is collected, which kind of diseases the patient may suffer from is fed back according to a knowledge map and a rule engine, and diagnosis and protection suggestions are given out;

after desensitizing the relevant information of the patient, the system feeds back the desensitized information to the group monitoring and early warning subsystem in an encryption mode.

Further, the group monitoring and early warning subsystem comprises the functions of visualization, statistical analysis, warning management and the like;

the data is from an individual monitoring and early warning subsystem and a hospital-end individual monitoring system, and based on a space-time analysis technology, the space-time aggregation of people with specific syndrome characteristics is monitored from a group perspective, early warning is realized on potential spread of diseases, and key data display, multi-dimensional data trend analysis and multi-level warning management are realized.

Further, the representative model for knowledge graph knowledge representation learning comprises a distance model, a single-layer neural network model, a bilinear model, a nerve tensor model, a matrix decomposition model and a translation model;

the distance model firstly represents the entities by vectors, then projects the entities into a vector space of a relationship pair with the entities through a relationship matrix, and finally judges the confidence coefficient of the existing relationship between the entities by calculating the distance between the projected vectors;

the bilinear model is used for describing semantic relevance of entities under the relationship through bilinear transformation based on the relationship between the entities;

the neural tensor model is used for linking the entities under different dimensions and representing complex semantic links among the entities;

the TransE model is to consider the relationship between entities in the knowledge base as some translation from the entities and is represented by a vector.

Further, the knowledge map comprises the steps of carrying out knowledge fusion

The preliminary screening is used for preliminarily screening entity data with the same fusion identifier;

judging attribute similarity, configuring similar attributes and similarity functions, and judging attribute similarity between data;

and (3) knowledge fusion: fusing data with attribute similarity reaching a threshold condition;

and judging the similarity of the entities, and obtaining the similarity of the entities according to the attribute similarity vector.

Further, the storing of the knowledge graph comprises storing based on a table structure and storing based on a graph structure;

based on the storage of the table structure, storing data in the knowledge graph by using a two-dimensional data table, wherein the data comprises a three-tuple table, a type table and a relational database;

and storing data in the knowledge graph, including a graph database, by using a graph mode based on the storage of the graph structure.

Further, the data service layer is provided with a rule engine which comprises

Forward link, based on inserted Fact object and update of Fact object, extracting more Fact objects by using available Fact inference rule until final target is calculated, finally matching one or more rules, and planning to execute;

reverse links, starting from the conclusion of the rules engine assumptions, search for sub-goals that can satisfy the assumptions if they cannot be directly satisfied.

The invention also provides a monitoring and early warning method of respiratory infectious diseases, which comprises the following steps:

step 1, collecting basic information of community residents, community health monitoring groups and workers in a key monitoring industry by a community end; the medical end individual monitoring subsystem collects diagnosis and treatment information of a patient;

step 2, diagnosing the basic information and the diagnosis and treatment information according to the knowledge graph and the rule engine to obtain diagnosis and treatment suggestions and protection suggestions corresponding to the basic information and the diagnosis and treatment information respectively;

step 3, desensitizing and encrypting the basic information and the diagnosis and treatment information and uploading the information to a group monitoring subsystem;

and 4, the group monitoring subsystem intensively displays the processed basic information and the diagnosis and treatment information.

Further, in step 4, the display process is specifically

And performing trend analysis and displaying by using the time dimension, the line graph stack and the bar graph and the stack region graph.

Further, in step 2, the diagnosis specifically includes the following steps:

abstracting basic information and diagnosis and treatment information into a symptom description text, and simultaneously inputting the symptom description text and a data field of blood examination into a decision tree algorithm;

extracting the features of the symptom description text to generate a corresponding 512-dimensional feature vector;

carrying out data normalization processing on the data field of the blood examination to obtain a blood characteristic vector;

the decision tree carries out iterative training by taking the 512-dimensional feature vector and the blood feature vector as training data;

the corresponding disease classification probability is diagnosed.

Compared with the prior art, the technical scheme of the invention has the following advantages:

the invention constructs a multidimensional and multi-field comprehensive, linkage and cooperative monitoring and early warning platform by comprehensively adopting technologies such as big data, artificial intelligence, privacy data protection and the like, establishes an infectious disease monitoring sentinel, increases intelligent monitoring and early warning of syndrome symptoms and infectious diseases with unknown reasons on the basis of legal infectious disease monitoring, and realizes the aims of information source multi-channel, monitoring and early warning intellectualization and multi-point triggering so as to improve the emergency handling capacity for dealing with emergent public health events.

The invention solves the problems that the information filling content of the important monitoring crowd in the community is not comprehensive, the working cost is high, the efficiency is low, and the early monitoring and early warning capability of the community cannot be improved. Meanwhile, the problem of monitoring and early warning of the known and unknown fever with respiratory infectious diseases of the community is solved.

The invention realizes the self-filling of basic information of community related personnel, community health monitoring crowd and major monitoring industry staff through the mobile terminal, wherein the basic information comprises information of personnel, basic diseases, epidemiological history, symptoms and the like. After the information is filled and submitted, the system automatically feeds back diagnosis and treatment suggestions and protection suggestions according to knowledge graph and rule engine diagnosis.

Drawings

Fig. 1 is a schematic structural diagram of a system according to an embodiment of the present invention.

Detailed Description

The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.

Example one

The monitoring and early warning system for respiratory infectious diseases of the present embodiment, as shown in fig. 1, includes:

the community monitoring mobile terminal subsystem, the hospital terminal individual monitoring subsystem and the group monitoring early warning subsystem all comprise the following layer structures:

a source data layer for acquiring source data;

and the data visualization layer is used for visualizing the data, calling an alarm rule to alarm according to the visualized data, and managing the alarm rule according to the requirement.

Further, an individual monitoring system at a hospital end is integrated with an HIS system and a PACS system in the hospital, diagnosis and treatment information of a patient is collected, the patient is fed back which kind of diseases the patient possibly suffers from according to a knowledge map and a rule engine, and diagnosis and protection suggestions are given out;

the data is from an individual monitoring and early warning subsystem and a hospital-side individual monitoring system, and based on a space-time analysis technology, the space-time aggregation of the population with the specific syndrome characteristics is monitored from the population perspective, the early warning is realized on the potential spread of the disease, and the key data display, the multi-dimensional data trend analysis and the multi-level alarm management are realized.

In the embodiment, because information encryption needs to be performed among hospitals, communities and sentry points, the encryption in the embodiment adopts an end-to-end encryption method for the characteristics of protecting user privacy and data security, and the end-to-end encryption allows data to exist in a ciphertext form all the time in the transmission process from a source point to a destination point. With end-to-end encryption (also called offline encryption or packet encryption), messages are not decrypted when they are transmitted until they reach the end point, and because they are protected during the whole transmission process, they will not be leaked even if nodes are damaged.

End-to-end encryption systems are more reliable, easier to design, implement, and maintain than link encryption and node encryption. End-to-end encryption also avoids synchronization problems inherent in other encryption systems, because each packet is encrypted independently, transmission errors occurring in one packet do not affect subsequent packets. Furthermore, end-to-end encryption is more natural from the user's intuition of security requirements. This encryption method may be chosen by a single user so as not to affect other users on the network, and only requires that the source and destination nodes be kept secret.

End-to-end encryption systems typically do not allow the destination address of a message to be encrypted because the node through which each message passes uses the address to determine how to transmit the message. This encryption method is vulnerable to preventing an attacker from analyzing the traffic, since it cannot mask the origin and destination of the transmitted message.

Statistical analysis in this embodiment refers to the analysis of collected data by effective (appropriate) statistical analysis methods, which are aggregated and understood and digested to extract useful information and form conclusions, which provide basis or advice for decision making.

Purpose of statistical analysis: the information hidden behind a large amount of disordered data is concentrated and refined, and the internal rules of the researched object are summarized.

Three steps of statistical analysis:

collecting data is a prerequisite and basis for statistical analysis. The data can be collected through a plurality of ways, and direct data can be obtained through experiments, observation, measurement, investigation and the like, and indirect data can also be obtained through document retrieval, reading and the like. In addition to the authenticity and reliability of the data, special attention is paid to distinguishing two types of data with different properties in the data collection process: the first is continuous data, also called metering data, which refers to data obtained through actual measurement; the second is discontinuous data, also called counting data, which refers to data obtained by counting the attributes of object types, grades, and the like.

The data sorting is a process of classifying and summarizing collected data according to certain standards. As most of the collected data are disordered, scattered and non-systematic, before statistical operation, the data need to be verified according to the research purpose and requirements, the non-real parts in the data are removed, and the data are grouped and gathered or tabulated, so that the original data are simplified, visualized and systematized, and the distribution characteristics of the data can be preliminarily reflected.

The analysis of data refers to the process of obtaining a conclusion through statistical operation on the basis of sorting the data, and is the core and key of statistical analysis. Data analysis can generally be divided into two levels: the first level is to calculate the index with external representativeness reflecting the trend, the discrete degree and the relevant intensity in the data set by using a descriptive statistical method; the second level is to process the data by using an inference statistical method on the basis of description statistics, infer the general situation by sample information, and analyze and infer the characteristics and the rules of the general

The statistical analysis method comprises the following steps:

in view of statistical theory, most of the statistical works seen by us are statistical analysis methods except for the theories and methods of statistical design, statistical investigation, statistical arrangement, etc., and the statistical analysis methods have a common point and are quantitative analysis methods.

In statistical practice, the role of statistical analysis is realized by quantitative analysis. Statistical analysis has three roles in human cognition: firstly, quantifying objective objects, including reflecting quantitative representation of the rules of the objective objects; secondly, confirming the quality of the object according to the quantitative change degree, namely determining the quantity limit for distinguishing the quality of the object; third, a new rule is revealed, that is, a rule of finding something that has not been recognized by analyzing a quantitative relationship. None of the above effects were by quantitative analysis.

From the characteristics of statistics, statistics is one of powerful weapons but not the only weapons in the understanding society, each science is the understanding of objective laws, and statistics is different from other subjects, namely the quantitative nature of statistics. Because of this, statistics cannot be kept in the substantive sciences of all departments, and the substantive sciences can achieve precision through statistical quantitative analysis.

Spatio-temporal analysis

Spatiotemporal data mining finds application in many fields, such as disease monitoring, environmental monitoring, public health and medical health, etc. As an emerging research field, time-space data mining is dedicated to developing and applying emerging computing technologies to analyze massive and high-dimensional time-space data and to mine valuable information in the time-space data.

Data has two basic inherent properties: spatial and temporal properties. The spatial attribute may refer to an absolute position attribute or a relative spatial relationship, and in general, the spatial relationship may be divided according to the following two ways:

local or global. Locally: relationships between the attribute dimensions within the same node (or object); global: relationships between different nodes (or objects).

Implicit or explicit. Implicit: the dimension or the relationship between the nodes can be considered as a characteristic dimension and is not presented in the form of a relationship graph structure; explicit: obvious graph structures of relationships can be constructed between nodes (or objects).

For the time relationship, it is generally presented in the form of a sequence. Therefore, the spatio-temporal sequence is analyzed as a unified and basic data structure, which can be divided into four basic data structures:

single node, single dimension sequence; single-node, multi-dimensional sequences; a multi-node, single-dimensional sequence; multi-node, multi-dimensional sequences.

Seven main methods of spatiotemporal data analysis include: the spatiotemporal data is visualized, and the purpose is to make assumptions and select an analysis model through visual inspiration; time sequence analysis of the spatial statistical indexes reflects the change of the spatial pattern along with time; the time-space change index reflects the comprehensive statistics of the time-space change; detecting the space-time pattern and the abnormity, and revealing the invariant and the variant part of the space-time process; performing space-time interpolation to obtain the numerical value of the non-sampling point; performing space-time regression, and establishing a statistical relationship between a dependent variable and an explanatory variable; modeling a time-space process, and establishing a mechanism mathematical model of the time-space process; and the spatial-temporal evolution tree reconstructs a spatial-temporal evolution path by using the spatial data.

Encrypted transmission of private data in the present embodiment

In the era of digital economy, data has become an important production factor, and the scale and speed at which data is collected, used, and analyzed is rapidly increasing. However, the ground for data sharing is still slow in terms of data security and personal information protection. Therefore, in order to break the data isolated island and fully release the value of the data, the privacy calculation technology is developed.

According to the definition of the global working group of the federated national big data, privacy-Preserving Computation (Privacy-Preserving Computation) refers to a series of technologies that can compute and analyze data in a state that the original data is guaranteed to be encrypted or invisible, and includes secure multi-party Computation, homomorphic encryption, trusted execution environment, differential Privacy, federal learning and the like. In privacy calculation, original data cannot leave the owner, all participants only share the final calculation result, and the separation of data ownership and use right is realized, so that the value of data elements can be mined to the maximum extent, and a solution is provided for data fusion requirements and privacy protection requirements.

The core idea of the privacy computing technology is to construct a trust basis of data sharing by methods such as code encryption and algorithm. Currently, industry technologies can be mainly divided into three categories:

the cryptography technology is represented by safe multi-party calculation and homomorphic encryption, and original data is encrypted or divided through an algorithm, so that a data receiver cannot be identified.

And the trusted execution environment is used for constructing a safety area based on the protection capability of hardware, and calculating after gathering the data needing to be protected to the area.

And Federal learning, namely, the latest application of privacy calculation in the field of artificial intelligence, training data is not required to be shared among a plurality of participants, and only model parameters are transmitted, so that joint modeling is realized on the premise that original data does not leave an enterprise private domain.

Secure multiparty computing

Secure multiparty computing mainly addresses the problem of how to securely compute an agreed function without a trusted third party. The technical specification of multi-party secure computing financial application defines the technical specification as a cryptographic technology which completes the computing target based on multi-party data and realizes that private data of all parties are not leaked except the computing result and derivable information thereof. "

Common cryptographic techniques used for secure multiparty computing are garbled circuits, inadvertent transmissions, secret sharing, homomorphic encryption, and the like. By operating the safe multiparty computing protocol, a plurality of participants can cooperate without exposing data plaintext to jointly execute computing on the data, and when the protocol is finished, each participant only knows own input and computing result but cannot know the input of other participants. Secure multiparty computing is applicable to situations where multiple participants wish to compute data jointly, but cannot share data directly for data security and privacy concerns. For example, secure multiparty computing may allow bidders to identify who won an auction without revealing any information about the actual bid.

(II) homomorphic encryption

Homomorphic encryption refers to an encryption algorithm that satisfies the homomorphic arithmetic property of a ciphertext and allows calculation of encrypted data (ciphertext) to obtain an encrypted result, which, after decryption, is identical to a result obtained by directly calculating original data (plaintext).

The homomorphic encryption can be calculated without decrypting the encrypted data, and is suitable for the situation that the data processing process is outsourced to a third party organization (such as a cloud service provider). The method can ensure that the user fully utilizes the computing resources of the cloud service provider on the premise of ensuring the data privacy.

(III) trusted execution Environment

The trusted execution environment is a hardware and operating system based security architecture that isolates data running therein from the general purpose execution environment to ensure confidentiality of the data. When private calculation is needed, the general execution environment sends the encrypted data to the trusted execution environment, and calculation is carried out after the encrypted data is decrypted into a plaintext.

The trusted execution environment provides a solution for realizing data privacy protection based on hardware protection capability, and the security of the trusted execution environment depends on the security of the hardware device, so that a trusted hardware manufacturer or a platform service provider needs to be selected when the trusted execution environment is actually used.

(IV) Federal learning

Federated learning is a distributed machine learning framework that enables joint modeling without multiple participants revealing their raw data.

In a conventional machine learning framework, a plurality of participants need to gather data held by each participant into one data center for training a model. Under a federal learning framework, a plurality of participants train a machine learning submodel by utilizing own original data, an intermediate result (such as gradient) is transmitted through an encryption mechanism, and a common model is finally established and maintained, so that the data privacy of each participant is protected, and the cost of concentrated transmission of a large amount of data is reduced.

Multimodal learning is one of the leading and most potential developing techniques in the field of artificial intelligence. Each source or form of information may be referred to as a modality. For example, humans have touch, hearing, vision, smell; information media such as voice, image, video, text and the like; a wide variety of sensors such as radar, infrared, accelerometer, etc. Each of the above may be referred to as a modality. MultiModal Machine Learning, known as MultiModal Machine Learning (MMML), aims to achieve the capability of processing and understanding multi-source modal information through a Machine Learning method. The current popular research direction is multi-modal learning among images, videos, audios and semantics.

The multi-modal representation learning means that redundancy among the modalities is eliminated by utilizing complementarity among the multiple modalities, so that better feature representation is learned. The method mainly comprises two research directions: joint Representations (Joint Representations) and collaborative Representations (Coordinated Representations).

Joint representation maps information of a plurality of modes to a unified multi-mode vector space together;

the collaborative representation is responsible for mapping each modality in the multi-modality to a respective representation space, but certain correlation constraints (such as linear correlation) are satisfied between the mapped vectors.

Features learned using multi-modal representations can be used for information retrieval and also for classification/regression tasks.

With the continuous development and progress of medical imaging technology and computer technology, medical image analysis has become an indispensable tool and technological means in medical research, clinical disease diagnosis and treatment. In recent years, deep Learning (DL), in particular, deep Convolutional Neural Networks (CNNs), has rapidly developed into a research hotspot of medical image analysis, which can automatically diagnose diseases and characteristics implicit in a special area from medical image big data.

The main medical image analysis tasks:

medical image classification and recognition

Clinicians need to assist in diagnosing whether a human body has a focus by means of medical images and quantitatively grade the degree of focus, so that automatically identifying focus regions and normal tissues and organs in the images is a basic task of medical image analysis.

Medical image localization and detection

The positioning of the anatomical structure of human tissues and organs and the lesion area is a very important pretreatment step in clinical treatment planning and intervention procedures, and the positioning precision directly influences the treatment effect. The image targeting task requires not only identifying a particular object in an image, but also determining its specific physical location. The task of image object detection requires that all objects in the image be identified and their physical location and classification determined.

Medical image segmentation task

Image segmentation, which is the key task of clinical surgical image navigation and image-guided tumor radiotherapy, is the identification of voxels inside and outside contours of the target region of interest in the image.

Common models are VGG networks, resNet networks, and EfficientNet networks.

In this embodiment, the knowledge graph is logically divided into a data layer and a mode layer, and the data layer is used for storing real data. The mode layer is above the data layer and is the core of the knowledge graph, and the refined knowledge is stored. The construction of the knowledge graph is an iterative updating process, and each iteration comprises three stages according to the logic of knowledge acquisition: extracting information, namely extracting entities, attributes and interrelations among the entities from various data sources, and forming ontology knowledge expression on the basis; knowledge fusion, after obtaining new knowledge, needs to integrate it to eliminate contradictions and ambiguities, for example, some entities may have multiple expressions, a certain name may correspond to multiple different entities, etc.; and (4) knowledge processing, namely, for the fused new knowledge, after quality evaluation (part of the new knowledge needs to be manually screened), adding the qualified part into a knowledge base to ensure the quality of the knowledge base.

The construction and application of large-scale knowledge bases require support of a plurality of technologies. Knowledge elements such as entities, relationships, attributes and the like can be extracted from data of some disclosed semi-structured, unstructured and third-party structured databases through knowledge extraction technology. The knowledge representation represents the knowledge elements by a certain effective means, so that the knowledge elements are convenient to further process and use. Then, ambiguity between the referents such as entities, relations and attributes and the like and the fact objects can be eliminated through knowledge fusion, and a high-quality knowledge base is formed. Knowledge reasoning is to further mine implicit knowledge on the basis of the existing knowledge base, so that the knowledge base is enriched and expanded. The comprehensive vector formed by the distributed knowledge representation has important significance for the construction, reasoning, fusion and application of the knowledge base.

In this embodiment, the rule engine is developed from the inference engine, and is a component embedded in the application program, which implements separation of the business decision from the application program code, and writes the business decision using a predefined semantic module. And receiving data input, explaining the business rules, and making business decisions according to the business rules. The rules engine integrates the Fact set and the rules set of the incoming system to trigger one or more business operations. The rule is usually implemented in the service code in an declarative manner, which we may assume is rarely changed. In reality, however, the decision conditions of these business logics are often changed. The business logic or rules in this article may be generally expressed as "performing certain tasks under certain writing conditions". In a business system with a large number of rules and Fact objects, it may happen that multiple Fact inputs will all result in the same output, a situation that we often refer to as a rule conflict. The rule engine may employ different conflict resolution schemes to determine the order of execution of the conflicting rules.

The rule engine in this embodiment has two execution modes:

forward linking: this is a form based on "data driven", based on the insertion of Fact objects and the updating of Fact objects, the rules engine extracts more Fact objects with the available Fact inference rules until the final goal is computed, eventually one or more rules are matched and planned for execution. Thus, the rules engine starts with facts and starts with conclusions.

Reverse linking: this is a form of "target-driven" or inference based, as opposed to forward linking. The reverse chain starts with the conclusion of the rules engine assumptions, and if these assumptions cannot be directly met, searches for sub-goals that can meet the assumptions. The rules engine will loop through this process until a conclusion is proved or there are no more demonstrable sub-targets.

Example two

The embodiment provides a monitoring and early warning method of respiratory infectious diseases, which comprises the following steps:

and 4, the group monitoring subsystem intensively displays the processed basic information and the diagnosis and treatment information, and performs trend analysis and display by using the time dimension, the line graph stack, the bar graph and the stack area graph.

In the embodiment, when diagnosis is performed, basic information and diagnosis and treatment information are abstracted into a symptom description text, and are input into a decision tree algorithm together with a data field of blood examination; the data fields of the blood test comprise: "red blood cell count", "lymphocyte count", "C-reactive protein", "hemoglobin", "white blood cell", etc.

The method comprises the steps of firstly performing pre-training once by using a BERT pre-training model based on a symptom description text of a patient, and generating a corresponding 512-dimensional feature vector after extracting a text feature vector by using the BERT pre-training model.

And carrying out data normalization processing on the data field of the blood examination to obtain a blood characteristic vector.

And the decision tree is iteratively trained by taking the 512-dimensional feature vector and the blood feature vector as training data.

The decision tree is composed of nodes and directed edges, with interior nodes representing a feature or attribute and leaf nodes representing a classification. When the decision tree is used for classification, the instances are distributed into the classes of the leaf nodes, and the class to which the leaf nodes belong is the classification of the node. In this embodiment, the data after the text feature extraction and the data field normalization are combined together to obtain a training set a. Then, the training set a is randomly cut into a data set train and a data set valid according to the ratio of 3.

Wherein the data set train is used for training only. The dataset valid is only used to verify model accuracy and is not trained.

When the decision tree is executed, firstly, feature selection is carried out, and features with strong classification capability are selected. If the result of classification using a certain feature is not very different from the result of random classification, the feature is said to be not capable of classification. Throwing away such features empirically has little impact on the accuracy of decision tree learning. The criterion of selecting the features is to find out local optimal features as judgment for segmentation, and the segmentation rule is more appropriate as the divided partitioned data is purer depending on the degree of order (purity) of the categories in the segmented node data set. CART (kini coefficient) was used as an evaluation index of the classified purity.

All training samples are then placed at the root node. Then, all training samples are divided into a plurality of subsets according to an optimal characteristic, and the subsets are ensured to have the best classification effect. If the classification effect has reached our goal, we directly construct leaf nodes and partition the subsets into corresponding leaf nodes. And if the classification effect does not achieve the preset classification purpose, continuously dividing the subset.

Finally, 1000 steps are trained based on the data set train, and the accuracy of the current model (F1 index) is verified every 200 steps using the data set valid. If the accuracy of the current model is improved from the last (previous 200 steps), the weight of the current model is saved.

After 1000 steps are trained, the algorithm is automatically ended, and an optimal precision model is output.

And finally, according to the trained decision tree, taking the data field of the blood examination of the patient and the symptom description text of the patient as input, and outputting the corresponding disease classification probability for diagnosis.

It should be understood that the above examples are only for clarity of illustration and are not intended to limit the embodiments. Other variations and modifications will be apparent to persons skilled in the art in light of the above description. And are neither required nor exhaustive of all embodiments. And obvious variations or modifications derived therefrom are intended to be within the scope of the invention.

Claims

1. A monitoring and early warning system for respiratory infectious diseases is characterized by comprising a community monitoring and moving terminal subsystem, a hospital terminal individual monitoring subsystem and a group monitoring and early warning subsystem;

a source data layer for acquiring source data;

the basic resource layer is used for providing data encryption service and calling MySQL, docker and Hyperridge Fabric according to data encryption requirements;

2. The respiratory infectious disease monitoring and early warning system according to claim 1, wherein the community monitoring subsystem realizes self-filling of basic information of community residents, community health monitoring groups and major monitoring industry workers, including information of basic information of personnel, basic diseases, epidemiological history, symptoms and the like;

3. The respiratory infectious disease monitoring and early warning system according to claim 2, wherein the individual monitoring system at the hospital end is integrated with the HIS system and the PACS system in the hospital, acquires diagnosis and treatment information of the patient, feeds back which diseases the patient may suffer from and gives diagnosis and protection suggestions according to the knowledge graph and the rule engine, and the system adopts a multi-mode model to provide multi-stage prediction early warning information for the doctor so as to realize the targets of early triage and early protection;

4. A monitoring and pre-warning system for respiratory infectious diseases according to claim 3, wherein the group monitoring and pre-warning subsystem comprises visualization, statistical analysis, alarm management, etc.;

5. A monitoring and pre-warning system for respiratory infectious disease as claimed in claim 4, wherein the representative model of knowledge-graph intellectual representation learning comprises a distance model, a single layer neural network model, a bilinear model, a nerve tensor model, a matrix decomposition model and a translation model;

the neural tensor model is used for linking entities under different dimensions and representing complex semantic links among the entities;

6. The system of claim 5, wherein the knowledge-based profile fusion comprises knowledge fusion

7. A monitoring and pre-warning system for respiratory infectious diseases according to claim 6, wherein the knowledge graph storage includes a table structure-based storage and a graph structure-based storage;

8. A respiratory infectious disease monitoring and early warning system according to claim 7, wherein the data service layer is provided with a rules engine, and the rules engine comprises

reverse linking, starting from the conclusion of the rules engine assumptions, searches for sub-goals that can satisfy the assumptions if they cannot be directly satisfied.

9. A monitoring and early warning method for respiratory infectious diseases is characterized by comprising the following steps:

step 1, collecting basic information and diagnosis and treatment information of community residents, community health monitoring groups and workers in key monitoring industries at a community end; the individual monitoring subsystem at the medical end acquires basic information and diagnosis and treatment information of a patient; the diagnosis and treatment information comprises

and 4, the group monitoring subsystem centrally displays the processed basic information and the diagnosis and treatment information.

10. The method of claim 9, wherein the displaying process in step 4 is specifically performed as follows

11. The method of claim 9, wherein the step 2 of diagnosing comprises the steps of:

extracting features of the symptom description text to generate a corresponding 512-dimensional feature vector;

the corresponding disease classification probability is diagnosed.