A deep learning-based special disease high-risk population grading early warning method and system

Abstract

The application discloses a kind of based on deep learning's special disease high-risk population grading early warning method and system, it is related to health risk early warning technical field, including the non-equal interval timing management sample being constructed by combining adjacent event actual time interval, review follow-up delay, upload delay and equipment non-upload state, and data credibility factor and special disease risk state tensor are generated;Further construct personnel similar graph and special disease event graph, respectively extract personnel transmission risk vector and event conduction risk vector by comorbidity stratification edge gate control, personnel maximum group enhancement and event type gate control;Risk vector is input into deep learning model to predict risk probability, and combined with risk acceleration and data credibility to generate correction early warning score.Can improve the comprehensive identification ability of special disease high-risk population grading early warning to non-equal interval management process, data credibility state, similar crowd risk migration and special disease event conduction relationship, to improve the timeliness and accuracy of high-risk object early warning.

Description

Technical Field

[0001] This invention relates to the field of health risk early warning technology, and in particular to a method and system for classifying and early warning of high-risk groups for specific diseases based on deep learning. Background Technology

[0002] Currently, with the development of medical informatization and the continuous management system for chronic diseases, the management methods for specific populations such as diabetes, hypertension, COPD, and cardiovascular and cerebrovascular diseases are gradually evolving from offline manual registration, regular follow-ups, and single-indicator threshold judgments to digital management methods based on electronic health records, follow-up records, examination and test results, medication records, and vital sign monitoring data. Existing specific population management platforms can typically collect basic patient information, medical records, follow-up results, follow-up records, and some monitoring data, and identify individuals requiring follow-up, re-examination, or special attention based on preset management rules or statistical models. Some systems are also beginning to introduce machine learning or deep learning methods to extract features and predict risks from historical health data to assist management institutions in conducting population stratification, health interventions, and follow-up resource allocation.

[0003] However, existing methods often focus on individual risk prediction for single managed individuals or rely solely on single-graph propagation based on patient similarity relationships. They struggle to simultaneously express the risk migration relationships between similar managed individuals, as well as the temporal transmission relationships between disease-specific management events such as examinations, medication administration, follow-up visits, equipment uploads, medical insurance settlement, and comorbidity labeling. For graded early warning systems for high-risk groups of specific diseases, risk is not solely determined by a single indicator value but also by the comorbidity status of similar populations, the similarity of management behaviors, and the chain-like changes in events. If the model cannot distinguish between the risk of transmission based on similar individuals and the risk of event transmission, it can easily lead to a single basis for early warning, insufficient interpretability of graded results, and affect the generation of subsequent follow-up reminders, key reviews, and immediate early warning tasks. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a method and system for classifying and warning high-risk groups of specific diseases based on deep learning, which solves the problems of insufficient characterization of non-equal interval management process and data reliability status, and insufficient joint modeling of the relationship between risk migration of similar groups and transmission of specific disease events in existing methods.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides a method for tiered early warning of high-risk groups for specific diseases based on deep learning, comprising,

[0008] Acquire multi-source management data of target disease management objects, generate disease population management data by association according to the unique identifier of the management object, preprocess the disease population management data, generate effective management event sequences sorted by time, and generate non-equal interval time series management samples by combining the actual time interval of events and management delay characteristics.

[0009] Data credibility factors are generated based on the data missing status, data upload delay status, and follow-up completion status in the non-equal interval time series management samples, and the disease risk status tensor is encoded. Based on the disease risk status tensor, personnel similarity graph, disease event graph, and object-event association matrix are constructed.

[0010] Based on the personnel similarity graph, we perform comorbidity hierarchical edge gating and personnel maximum clique enhancement to extract personnel transmission risk vectors. Based on the disease-specific event graph and object-event association matrix, we perform event type gating and temporal association aggregation to extract event transmission risk vectors.

[0011] By fusing the personnel transmission risk vector, the event transmission risk vector, and the disease-specific risk state tensor into a deep learning risk prediction model, the risk probability within the current warning period can be obtained.

[0012] The risk acceleration is calculated by combining the risk probability of adjacent early warning cycles and the actual time interval between adjacent effective management events, and a revised early warning score is generated by combining the data credibility factor. The graded early warning level is determined based on the revised early warning score, and special disease population management tasks are generated according to the graded early warning level.

[0013] As a preferred embodiment of the deep learning-based method for tiered early warning of high-risk groups for specific diseases described in this invention, the multi-source management data includes: basic medical records, historical medical records, examination and testing data, medication execution data, follow-up plan data, actual follow-up data, follow-up plan data, actual follow-up data, medical insurance settlement data, vital sign monitoring data, device upload data, and comorbidity label data of the target specific disease management object; the specific disease population management data is a set of multi-source management data formed by associating the multi-source management data corresponding to each target specific disease management object according to the unique identifier of the management object.

[0014] As a preferred embodiment of the deep learning-based method for classifying and warning high-risk groups for specific diseases according to the present invention, the steps of generating a time-ordered sequence of effective management events and generating non-equidistant time-series management samples by combining the actual time intervals of the events and management delay features are as follows:

[0015] The preprocessed disease-specific population management data are grouped according to the unique identifier of the managed object and sorted by the time of event occurrence to generate a valid management event sequence corresponding to each managed object;

[0016] Valid management events at the same management time point are grouped into an event set, the actual interval between adjacent management time points is calculated, and management delay characteristics are generated based on events related to review, follow-up, indicator upload, equipment upload, and indicator anomaly, to obtain non-equal interval time series management samples.

[0017] As a preferred embodiment of the deep learning-based method for tiered early warning of high-risk populations for specific diseases described in this invention, the steps for generating data credibility factors based on data missing status, data upload delay status, and follow-up completion status in non-equal interval time-series management samples, and encoding them to obtain a disease risk state tensor, are as follows: generating data credibility factors corresponding to each management time point; converting the numerical event values ​​of each effective management event in the non-equal interval time-series management samples into model input numerical features, converting the non-numerical event information of each effective management event into event type vectors, and aggregating the model input numerical features and event type vectors under the same management time point to generate an event state expression; concatenating the event state expression with non-equal interval time features, management delay features, and data credibility factors to generate a disease risk state tensor.

[0018] As a preferred embodiment of the deep learning-based method for classifying and warning high-risk groups of specific diseases according to the present invention, the steps for constructing a personnel similarity graph, a specific disease event graph, and an object-event association matrix based on the specific disease risk state tensor are as follows:

[0019] Based on the disease risk state tensor, management profile features of each target disease management object are extracted, management object profile vectors are generated, and personnel similarity edges are established according to the similarity between the management object profile vectors to construct a personnel similarity graph.

[0020] Disease-specific management events are extracted based on disease-specific risk state tensors, the temporal correlation between different disease-specific management events is statistically analyzed, a disease-specific event graph is constructed, and an object-event correlation matrix is ​​constructed based on the correspondence between each management object and a disease-specific management event.

[0021] As a preferred embodiment of the deep learning-based method for classifying and warning high-risk groups of specific diseases according to the present invention, the extraction of the personnel transmission risk vector includes using a personnel similarity graph, utilizing event state expression, event type vectors corresponding to comorbidity labels, and management delay features to perform comorbidity hierarchical gating on personnel similarity edges to obtain the gating personnel similarity edge weights;

[0022] Based on the identification of the maximum clique structure formed by personnel similarity edges in the personnel similarity graph, and based on the maximum clique structure, the weight of the personnel similarity edges after gating is enhanced. Then, attention aggregation is performed on the disease risk state vectors of adjacent managed objects to obtain the personnel transmission risk vector.

[0023] As a preferred embodiment of the deep learning-based method for classifying and warning high-risk groups of specific diseases according to the present invention, the extraction of event transmission risk vector includes: based on the specific disease event graph, gating the event association edges using the event type and temporal correlation of specific disease management events to obtain enhanced event association edge weights, and aggregating them to generate event transmission vectors.

[0024] Based on the object-event correlation matrix, the event transmission vector is mapped to the corresponding target disease management object to obtain the event transmission risk vector.

[0025] As a preferred embodiment of the deep learning-based high-risk population classification and early warning method for specific diseases described in this invention, the steps for obtaining the risk probability within the current early warning period are as follows:

[0026] The risk vectors of personnel transmission, event transmission, disease-specific risk status, non-equidistant time intervals, and data credibility factors are mapped to the same feature dimension and fused to generate a fused associated risk input. The fused associated risk input is then input into the non-equidistant time-encoded Transformer model in the order of management time points. The time-series risk representation is extracted through a self-attention mechanism with time interval decay, and the risk probability within the current warning period is output based on the time-series risk representation at the current management time point.

[0027] As a preferred embodiment of the deep learning-based tiered early warning method for high-risk groups of specific diseases described in this invention, the steps of calculating risk acceleration by combining the risk probability of adjacent early warning periods and the actual time interval between adjacent effective management events, and generating a corrected early warning score by combining data credibility factors, are as follows:

[0028] Based on the current risk probability, the historical risk probability corresponding to adjacent management time points, and the actual time interval between adjacent management time points, calculate the unit time risk change rate of the current stage and the previous stage, and generate a risk acceleration index based on the difference between the two.

[0029] A revised early warning score is generated based on the current risk probability, risk acceleration index, and data credibility factor.

[0030] Secondly, this invention provides a deep learning-based method for tiered early warning of high-risk populations for specific diseases, including:

[0031] The data collection module is used to acquire and associate multi-source management data of the target disease management object;

[0032] The event generation module is used to preprocess data on the management of specific disease populations and generate effective management event sequences;

[0033] The state encoding module is used to generate data credibility factors based on non-equal interval time-series management samples, and to encode effective management events, time features, management delay features and data credibility factors to generate a disease-specific risk state tensor.

[0034] The graph structure construction module is used to construct personnel similarity graphs, disease event graphs, and object-event association matrices based on disease risk state tensors.

[0035] The associated risk extraction module is used to extract personnel transmission risk vectors based on personnel similarity graphs and to extract event transmission risk vectors based on disease-specific event graphs and object-event association matrices.

[0036] The early warning and response module is used to input the personnel transmission risk vector, event transmission risk vector, and disease-specific risk state tensor into the deep learning risk prediction model to generate risk probabilities, correct early warning scores, classify early warning levels, and corresponding disease-specific population management tasks.

[0037] The beneficial effects of this invention are as follows: By organizing multi-source management data of target diseases into non-equal interval time-series management samples and introducing a data credibility factor, this invention enables anomalies in the management process, such as overdue follow-up, incomplete follow-up, delayed upload, and failure to upload data, to participate in risk assessment, avoiding the delay in early warning caused by relying solely on static risk probabilities. At the same time, by extracting personnel transmission risk vectors and event transmission risk vectors from personnel similarity maps and disease event maps, respectively, and combining them with risk acceleration to generate a corrected early warning score, this invention can identify high-risk individuals with rapidly rising risks or declining data credibility earlier, improving the accuracy of tiered early warning and the pertinence of management task allocation. Attached Figure Description

[0038] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0039] Figure 1 This is a flowchart of the deep learning-based method for classifying and warning high-risk groups for specific diseases in Example 1.

[0040] Figure 2 This is a structural diagram of the deep learning-based high-risk population classification and early warning system for specific diseases in Example 1.

[0041] Figure 3 This is a schematic diagram of the dual-graph structure relationship in Example 1.

[0042] Figure 4 This is a schematic diagram of the risk scoring and graded treatment process in Example 1. Detailed Implementation

[0043] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0044] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0045] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0046] Example 1, referring to Figure 1 , Figure 3 and Figure 4 This is the first embodiment of the present invention, which provides a method for classifying and early warning of high-risk groups for specific diseases based on deep learning, including the following steps:

[0047] S1. Obtain multi-source management data of the target disease management object, generate disease population management data by association according to the unique identifier of the management object, preprocess the disease population management data, generate a valid management event sequence sorted by time, and generate non-equal interval time series management samples by combining the actual time interval of the events and management delay characteristics.

[0048] S1.1: When the current early warning task is triggered, the disease-specific population management platform reads the target disease code, management institution code, and early warning statistical period, and extracts management objects that are consistent with the target disease code and whose management status is valid from the disease-specific management object library to form a set of target disease-specific management objects; reads the unique identifier of the management object and uses the unique identifier of the management object as the unified association primary key of the management object in this early warning task; for management objects that have not generated a unique identifier, a unique identifier is generated based on their registration identity credential code, management institution code, and target disease code and then written into the disease-specific management object library.

[0049] S1.2: Based on the unique identifier of the managed object, query the cross-system object mapping table to obtain the source system object identifier of the managed object in the electronic health record system, follow-up management system, examination and testing system, medication management system, medical insurance settlement system, vital sign monitoring system, and wearable device access system; when no corresponding source system object identifier exists in any data source system, skip the data reading of that data source system and record the corresponding missing data source mark; according to the source system object identifier and the early warning statistical cycle, read the original management records related to the target disease management from each data source system, including basic record records, historical medical records, examination and testing records, medication execution records, follow-up plan records, actual follow-up records, follow-up plan records, actual follow-up records, medical insurance settlement records, vital sign monitoring records, device upload records, and comorbidity label records.

[0050] S1.3: The disease-specific population management platform writes a unique identifier for the managed object, the target disease code, the data source system code, the source system object identifier, the source record number, the business occurrence time, the platform collection time, and the data category code for each original management record. It also aggregates multiple original management records belonging to the same managed object into the same data set according to the unique identifier of the managed object. The platform summarizes the data sets corresponding to each managed object to generate the target disease-specific population management data.

[0051] It should be noted that each record in the target disease population management data is identified by a unique identifier of the managed object, and the correspondence with the original business system record is maintained through the data source system code and the source record number.

[0052] S1.4: Based on the management data of the target disease population, the multi-source management data of each managed object is standardized according to the unique identifier of the managed object. This includes field standardization, timestamp unification, unit conversion, duplicate record merging, data quality anomaly marking, and privacy desensitization. Among them, field standardization is used to map synonymous fields in different data source systems to unified field names; timestamp unification is used to convert consultation time, examination time, medication time, follow-up time, re-examination time, and upload time into a unified time format; unit conversion is used to convert similar indicators into a unified unit of measurement; duplicate record merging is used to retain the latest valid records under the same managed object, the same data source, and the same event time; data quality anomaly marking is used to mark records with missing fields, incorrect formats, or values ​​that obviously do not meet the data collection requirements; and privacy desensitization is used to remove or encrypt direct identity information that is not related to risk warning.

[0053] S1.5: Group the preprocessed disease-specific population management data according to the unique identifier of the managed object, sort them according to the time of event occurrence, and generate a valid management event sequence corresponding to each managed object;

[0054] Specifically, the preprocessed disease-specific population management data is grouped according to the unique identifier of the managed object. This ensures that the basic records, historical medical records, examination and test records, medication execution records, follow-up plan records, actual follow-up records, follow-up plan records, actual follow-up records, medical insurance settlement records, vital sign monitoring records, device upload records, and comorbidity label records of the same managed object from different data source systems are grouped into the same object data group. Each piece of management data is converted into a valid management event. The valid management events in the same object data group are sorted from earliest to latest according to the event occurrence time. When multiple valid management events have the same occurrence time, they are arranged in a fixed order according to the event type and data source system code to ensure that the event sequence of the same managed object remains consistent when repeated. After sorting, the valid management event sequence corresponding to that managed object is generated.

[0055] It should be noted that effective event management includes at least the unique identifier of the managed object, the event occurrence time, the event type, the event value, and the data source. The unique identifier of the managed object is used to identify the managed object to which the event belongs. The event occurrence time is used to determine the sequence of events in the management process. The event type is used to distinguish different management matters such as basic records, medical visits, examinations and tests, medication administration, follow-up visits, medical insurance settlement, vital sign monitoring, equipment uploads, and comorbidity labels. The event value is used to store the standardized data content corresponding to the event, and the data source is used to record the business system from which the event originated.

[0056] S1.6: Merge valid management events at the same management time point into an event set, calculate the actual interval between adjacent management time points, and generate management delay characteristics based on events related to review, follow-up, indicator upload, equipment upload, and indicator anomaly to obtain non-equal interval time-series management samples.

[0057] Specifically, each valid management event corresponding to a managed object is read one by one according to its unique identifier, and the occurrence time of the valid management event is used as the management time point. When multiple valid management events exist for the same managed object at the same occurrence time, that occurrence time is taken as the same management time point, and all valid management events at that time point are grouped into the same event set. The event sets of the same managed object are arranged from earliest to latest according to the management time points, and the actual time interval between the current management time point and the previous management time point is calculated as the non-uniform interval time characteristic of the current management time point. For the first management time... At the designated point, the actual time interval is recorded as zero, and the actual time interval is uniformly converted into days. After obtaining the non-equal interval time characteristics, review plan events, actual review events, follow-up plan events, and actual follow-up events are extracted from the effective management event sequence. The review delay time and follow-up delay time are generated based on the difference between the planned time and the actual execution time. When the actual execution time is not later than the planned time, the corresponding delay time is recorded as zero. When there is no corresponding actual execution event as of the current management time point and the current management time point is later than the planned time, the difference between the current management time point and the planned time is used as the corresponding delay time.

[0058] Furthermore, for inspection and testing events, vital sign monitoring events, and equipment upload events, the indicator upload delay time is generated based on the difference between the event occurrence time and the platform data collection time; when the platform data collection time is not later than the event occurrence time, the indicator upload delay time is recorded as zero. For equipment upload events, if there is no equipment upload event at the current management time point, the duration of equipment not uploaded is generated based on the difference between the occurrence time of the most recent equipment upload event and the current management time point; if there is an equipment upload event at the current management time point, or if there was no previous equipment upload event for the managed object, the duration of equipment not uploaded is recorded as zero.

[0059] The set of valid management events corresponding to each management time point, non-equal interval time characteristics, review delay time, follow-up delay time, indicator upload delay time, duration of non-upload by device, and missing data source marker are combined to generate the non-equal interval time series management sample corresponding to that management time point; then, the non-equal interval time series management samples of the same management object are arranged from earliest to latest according to the management time point to form the non-equal interval time series management sample set of the target disease population.

[0060] S2. Generate data credibility factors based on the data missing status, data upload delay status, and follow-up completion status in the non-equal interval time series management samples, and encode them to obtain the disease risk status tensor. Construct personnel similarity graph, disease event graph, and object-event association matrix based on the disease risk status tensor.

[0061] S2.1: Generate data reliability factors corresponding to each management time point;

[0062] Specifically, based on the unique identifier of each managed object, the system sequentially reads the set of valid management events, data source missing markers, indicator upload delay time, equipment non-upload duration, planned review events, actual review events, planned follow-up events, and actual follow-up events corresponding to each management time point. It then determines the data missing status, data upload delay status, and review / follow-up completion status for the current management time point. The data missing status characterizes the degree of absence of event occurrence time, event type, event value, data source, and inspection / testing events, vital sign monitoring events, and equipment upload events. The data upload delay status characterizes the impact of indicator upload delays and equipment non-upload on the real-time performance of the samples. The review / follow-up completion status characterizes the actual completion degree of overdue review and follow-up plans. The data missing status, data upload delay status, and review / follow-up completion status are converted into normalized statistical values ​​and weighted according to preset weights to generate a data reliability factor corresponding to the current management time point. This data reliability factor is then bound to the corresponding non-equal interval time-series management sample.

[0063] It should be noted that the preset weights include the weight for missing data, the weight for upload delay, and the weight for completion of follow-up visits. The sum of the three is 1, and they are determined based on historical disease management samples before the system goes online and written into the disease population management platform, and remain fixed within the same early warning cycle.

[0064] S2.2: Convert the numerical event values ​​of each effective management event in the non-equal interval time series management sample into model input numerical features, convert the non-numerical event information of each effective management event into event type vectors, and aggregate the model input numerical features and event type vectors under the same management time point to generate an event state expression; concatenate the event state expression with non-equal interval time features, management delay features and data credibility factors to generate a disease-specific risk state tensor.

[0065] Specifically, based on the non-uniformly spaced time-series management samples and their corresponding data reliability factors obtained in the previous step, the set of valid management events at each management time point is read one by one according to the unique identifier of the management object and the management time point. For numerical event values ​​in valid management events that have completed business standardization, normalization is performed according to the mean and standard deviation of the corresponding fields in the historical training samples to obtain the numerical features of the model input. For non-numerical event information in valid management events, an event coding table is established using the label coding method. The event coding table is generated based on the event type, data category, unified field name, inspection item name, medication category, re-examination event, follow-up event, medical insurance settlement event, data source, and comorbidity label in the historical training samples, and a unique integer code is assigned to each event value according to a fixed sorting. For event values ​​that do not appear in the event coding table during the real-time early warning process, an unknown event code is written, and for missing events... Event values ​​are entered into missing event codes. Then, an embedding lookup method is used to input unique integer codes into a pre-trained embedding matrix, converting them into fixed-dimensional event type vectors. When multiple valid management events exist at the same management time point, the model input numerical features and event type vectors corresponding to each valid management event at that time point are averaged and aggregated to form the event state expression for that management time point. Subsequently, the event state expression is concatenated with non-equal interval time features, review delay time, follow-up delay time, indicator upload delay time, device non-upload duration, and data reliability factor to obtain the disease risk state vector for the corresponding management time point. Then, the disease risk state vectors of the same managed object are arranged from earliest to latest according to the management time point to form the disease risk state sequence for that managed object. Finally, the disease risk state sequences of each managed object are combined according to a unified feature dimension to generate a disease risk state tensor.

[0066] S2.3: Extract management profile features of each target disease management object based on the disease risk state tensor, generate management object profile vectors, and establish personnel similarity edges based on the similarity between the management object profile vectors to construct a personnel similarity graph;

[0067] Specifically, based on the disease risk state tensor, the disease risk state sequence corresponding to each managed object is read one by one according to the unique identifier of the managed object, and management profile features that can characterize the overall disease management status of the managed object are extracted from the disease risk state sequence. These features include basic state features, indicator state features, management behavior features, time delay features, data credibility features, and comorbidity label features. Then, the management profile features are concatenated according to a fixed feature order to generate the management object profile vector corresponding to each managed object. For the profile vectors between different managed objects, the Gaussian kernel similarity calculation method is used to calculate the personnel similarity between each pair of managed objects, and the K nearest neighbor graph construction method is used to select the K management objects with the highest similarity as adjacent objects for each managed object. The target disease managed object is used as a node in the personnel graph, and personnel similarity edges are established between the selected adjacent objects. The corresponding Gaussian kernel similarity is used as the edge weight of the personnel similarity edge, thereby constructing a personnel similarity graph.

[0068] It should be noted that the nodes in the personnel similarity graph represent target disease management objects, the edges represent the similarity management status relationship between management objects, and the edge weights represent the degree of similarity.

[0069] S2.4: Extract disease management events based on disease risk state tensors, statistically analyze the temporal correlation between different disease management events, construct a disease event graph, and construct an object-event correlation matrix based on the correspondence between each management object and disease management event.

[0070] Specifically, based on the disease-specific risk state tensor, the disease-specific risk state sequence of each managed object is read, and disease-specific management events are extracted from the set of effective management events corresponding to each management time point. These events include examination and testing events, vital sign monitoring events, medication execution events, follow-up events, follow-up events, indicator upload events, equipment failure to upload events, medical insurance settlement events, and comorbidity labeling events. Different types of disease-specific management events are used as event nodes to form a set of event nodes. According to the chronological order of effective management events for the same managed object, the event transition frequency statistics method is used to count the number of times other disease-specific management events continue to occur after the occurrence of any disease-specific management event. Then, the conditional transition probability method is used to determine the different diseases. The temporal correlation strength between management events is assessed. When a temporal correlation exists between two disease-specific management events, directed event edges are established between the corresponding event nodes, and the temporal correlation strength is used as the edge weight of the directed event edges, thereby constructing a disease-specific event graph. Simultaneously, based on the actual disease-specific management events occurring for each management object at each management time point, an object-event correlation matrix is ​​established using the object-event correlation statistical method. An exponential time decay weighting method is used to assign higher correlation weights to historical events closer to the current management time point and lower correlation weights to historical events farther from the current management time point, thereby obtaining the correlation strength between each management object and each disease-specific management event at different management time points.

[0071] Furthermore, when establishing the object-event association matrix using the object-event association statistical method, the disease-specific population management platform first uses the already formed set of disease-specific management event nodes as column indexes and the non-equal interval time-series management samples formed by each managed object at each management time point as row indexes to establish an initial object-event association matrix, and initially sets all elements in the matrix to zero; then, according to the unique identifier of the managed object, it reads the valid management event sequence of the same managed object one by one, extracts the currently occurring and historically occurring disease-specific management events of the managed object at each management time point, and matches the extracted event types with the set of disease-specific management event nodes. If a managed object is at the current management time point... If a valid management event corresponds to an event node, the associated value is written to the corresponding row of the managed object and the corresponding column of the event node. For similar events occurring multiple times within the same management time point, the frequency is first counted or merged before the corresponding associated value is written. For events that have occurred at historical management time points but may still affect the current state, a decay weighting is applied based on the time distance between the historical event and the current management time point, so that events closer to the current management time point have higher associated values ​​and events farther away have lower associated values. When a managed object does not have an event corresponding to a certain event node in the current or historical management records, the corresponding matrix element remains zero. Through the above processing, each row in the matrix can represent the association between a managed object and various disease-specific management events at a certain management time point, and each column can represent the occurrence and impact of a disease-specific management event in different managed objects and different management time points.

[0072] It should be noted that the disease-specific event diagram is used to represent the temporal co-occurrence or transmission relationship between disease-specific management events such as follow-up examinations, follow-ups, medication, indicator uploads, equipment failures to upload, medical insurance settlements, and comorbidity labels; the object-event association matrix is ​​used to map the event transmission characteristics back to specific management objects.

[0073] S3. Based on the personnel similarity graph, perform comorbidity hierarchical edge gating and personnel maximum clique enhancement to extract personnel transmission risk vectors. Based on the disease-specific event graph and object-event association matrix, perform event type gating and temporal association aggregation to extract event transmission risk vectors.

[0074] S3.1: Based on the personnel similarity graph, use the event state expression, the event type vector corresponding to the co-disease label and the management delay feature to perform co-disease hierarchical gating on the personnel similarity edges to obtain the gating personnel similarity edge weights;

[0075] Specifically, based on the disease risk state tensor and the personnel similarity graph, the disease risk state vector of the i-th managed object at the t-th management time point is read, and the event state expression, the event type vector corresponding to the comorbidity label, and the management delay feature corresponding to the management time point are extracted from it; for any personnel similarity edge (i,j) in the personnel similarity graph, the event state similarity, comorbidity label similarity, and management delay similarity between the i-th and j-th managed objects are calculated respectively; based on the obtained three types of similarity, the comorbidity hierarchical personnel edge gating coefficient is generated; the personnel edge gating coefficient is applied to the original personnel similarity edge weights in the personnel similarity graph to obtain the gated personnel similarity edge weights.

[0076] Calculate event state similarity The expression is:

[0077]

[0078] In the formula, and They represent the first The and the first The number of managed objects in the first The expression of event status at each management time point The kernel width parameter represents the similarity of event states;

[0079] Calculate comorbidity label similarity The expression is:

[0080]

[0081] In the formula, and They represent the first The and the first The event type vector corresponding to the comorbidity tags of each managed object This represents a very small constant to prevent the denominator from being zero;

[0082] Calculate management delay similarity The expression is:

[0083]

[0084] In the formula, and They represent the first The and the first The management delay characteristics of each managed object Kernel width parameter representing administrative delay similarity; administrative delay features It is composed of normalized re-examination delay time, follow-up delay time, indicator upload delay time, and device non-upload duration;

[0085] Generate the side gating coefficient for stratified individuals with comorbidities. The expression is:

[0086]

[0087] In the formula, , and These represent the gating weight parameters corresponding to event state similarity, comorbidity label similarity, and management delay similarity, respectively. This indicates that the personnel side gate is offset. This is a logical function used to map the linear weighted result in the personnel side gating formula to the interval between 0 and 1;

[0088] similarity edge weights of people after gating The expression is:

[0089]

[0090] In the formula, This represents the original edge weights of the personnel similarity graph.

[0091] It should be noted that, , and The values ​​of all are in the range [0,1], and satisfy . The values ​​are determined based on the fact that the similarity of comorbidity combinations, event status, and management delay have all been normalized to a comparable scale. Using non-negative normalized weights can ensure that the three types of similar factors maintain an interpretable weighted contribution in personnel side gating. For example, the three can be initially set to 1 / 3, and then adjusted and determined based on the early warning accuracy of the validation set of historical labeled samples.

[0092] S3.2: Identify the maximum clique structure formed by personnel similarity edges based on the personnel similarity graph, enhance the weight of personnel similarity edges after gating based on the maximum clique structure, and then perform attention aggregation on the disease risk state vectors of adjacent managed objects to obtain the personnel transmission risk vector;

[0093] Specifically, based on the weights of similar edges after gating, similar edges with weights greater than zero in the similarity graph are identified as support edges, forming an undirected personnel support graph. In this graph, the set of management objects where support edges exist between any two managed objects is defined as a clique structure, and the clique structure that cannot accommodate any more managed objects while maintaining pairwise connections is defined as a maximal clique structure. The Bron-Kerbosch maximal clique search algorithm is used to search the undirected personnel support graph to obtain the set of maximal cliques. The cluster strength of a similar edge is determined based on whether the managed objects at both ends of the similar edge belong to the same maximal clique, and the cluster strength is normalized to its maximum value. The normalized cluster strength is used to enhance the weights of the similar edges after gating, giving them higher propagation weights. Based on the enhanced weights, the neighborhood of each managed object is determined, and the personnel propagation attention weight of adjacent managed object j to the i-th managed object is calculated to generate a personnel propagation risk vector.

[0094] Cluster strength The expression is:

[0095]

[0096] In the formula, Indicates the largest group of people. Indicates that it contains nodes. and nodes The largest group of personnel This indicates the number of managed objects within the largest clique; when a similar edge does not belong to any largest clique, Take zero;

[0097] The weights of similar edges for gated personnel are enhanced, as expressed by:

[0098]

[0099] In the formula, This represents the weight of the enhanced personnel similarity edges. This represents the normalized cluster strength of individuals. This represents the cluster enhancement coefficient, with a value range of [0,1]. The value is determined based on the normalized cluster strength, which also ranges from [0,1], making the enhancement term... The value of is restricted to [1,2], thereby enhancing the stability of personnel edges in similar population clusters without allowing cluster factors to excessively dominate the personnel transmission risk vector. An example value is 0.5.

[0100] Generate a vector of personnel transmission risk. The expression is:

[0101]

[0102]

[0103] In the formula, Indicates adjacent management objects For the managed objects The weight of personnel's attention in communication. This represents the attention parameter vector for the personnel channel. This represents the characteristic transformation matrix of the personnel passage. and They represent the first The and the first The number of managed objects in the first A vector of disease risk status at each management time point This represents a non-linear activation function, for example, one of ReLU, LeakyReLU, or sigmoid, and the same type of activation function is used consistently during the model training phase. Represents the managed object The people's neighborhood, For normalized exponential functions, It is a linear rectified function with a leakage term, used to preserve the weak response of the negative input in the attention score calculation, avoiding the loss of weak correlation information in the graph attention network caused by ordinary ReLU setting all negative correlations to zero;

[0104] It should be noted that the personnel transmission risk vector is used to characterize the risk migration impact of similar managed objects on the current managed object.

[0105] S3.3: Based on the disease-specific event graph, the event association edges are gating by utilizing the event type and temporal relationship of disease-specific management events to obtain the enhanced event association edge weights, and event propagation vectors are generated accordingly.

[0106] Specifically, the process involves reading the event nodes, event-related edges, and corresponding temporal-sequence-related edge weights from the disease-specific event graph. Based on the event type of each disease-specific management event within the effective management event sequence, the event category corresponding to each disease-specific management event is determined. Based on the disease-specific risk state vector and object-event association matrix formed up to the current management time point, event node features are generated for each event node, ensuring that the event node features simultaneously include the event type, event category, and the status information of the managed object where the event occurred. Event edge gating coefficients are generated based on the proximity of event categories between the events at both ends of each directed event edge in the disease-specific event graph and the corresponding temporal-sequence-related edge weights. These gating coefficients are then used to adjust the original event-related edge weights, resulting in enhanced event-related edge weights. After obtaining the enhanced event-related edge weights, the adjacent events of each event node are used as aggregation objects. Attention aggregation is performed on the event node features of adjacent events through the event propagation channel to obtain the event propagation vector corresponding to each disease-specific management event.

[0107] Generate event node features corresponding to each event node. The expression is:

[0108]

[0109] In the formula, Indicates an event event type embedding vector, Indicates an event Embedding vector of the event category, Indicates the first Each managed object at the management time point With the event The strength of the association, Indicates the first Each managed object at the management time point The disease-specific risk state vector, This indicates that only data obtained up to the current management point in time will be used;

[0110] similarity of event categories The expression is:

[0111]

[0112] In the formula, and Representing events respectively and events Event category embedding vector;

[0113] Generate event side gating coefficient The expression is:

[0114]

[0115] In the formula, Indicates an event Pointing to events The edge weight of the directed event edge. and This represents the event-side gating weight parameter. Indicates event side-gating bias;

[0116] Enhanced event-related edge weights The expression is:

[0117]

[0118] Event propagation vector The expression is:

[0119]

[0120]

[0121] In the formula, Representing adjacent events Regarding the event Event transmission attention weight, This represents the event channel attention parameter vector. Represents the event channel feature transformation matrix. and Representing events respectively and events In the Event node characteristics at each management time point It is a normalized exponential function;

[0122] It should be noted that the event propagation vector is used to characterize the temporal propagation risk information obtained by the disease management event from its related events.

[0123] It should be noted that, and The values ​​of are all in the range [0,1], and satisfy . The values ​​are determined based on both event type similarity and event temporal correlation strength, both of which measure the effective propagation of event edges. Using non-negative normalized weights allows both event category proximity and historical co-occurrence relationships to participate in event edge selection. For example, [the following can be used as an example:] and The initial value was set to 0.5, and then adjusted based on the event propagation prediction effect in historical samples.

[0124] S3.4: Based on the object-event association matrix, the event transmission vector is mapped to the corresponding target disease management object to obtain the event transmission risk vector.

[0125] Specifically, the correlation strength between the target disease management object and each disease management event is read according to the unique identifier of the management object and the management time point. The event transmission vector of the corresponding disease management event is weighted and summarized according to the correlation strength to obtain the event transmission risk vector of the target disease management object at the current management time point.

[0126] S4. The risk vector of personnel transmission, the risk vector of event transmission, and the risk state tensor of specific diseases are fused and input into the deep learning risk prediction model to obtain the risk probability within the current warning period.

[0127] S4.1: Map the personnel transmission risk vector, event transmission risk vector, disease-specific risk status vector, non-equidistant time interval features, and data credibility factors to the same feature dimension and fuse them to generate a fused and associated risk input;

[0128] Specifically, based on the personnel transmission risk vector and event transmission risk vector corresponding to the current management time point obtained in the previous step, and combined with the disease-specific risk status vector, non-equidistant time interval characteristics, and data reliability factors, the first step is to construct... The number of managed objects in the first Risk input related to the integration of various management time points :

[0129]

[0130]

[0131] In the formula, This represents the vector mapping matrix of disease-specific risk states. This represents the vector mapping matrix of personnel transmission risk. This represents the event propagation risk vector mapping matrix. Represents the time interval mapping matrix. This represents the data credibility mapping parameter. Represents the disease-specific risk state vector. This represents the vector of risk of transmission among individuals. Represents the event propagation risk vector. Indicates the data credibility factor. This indicates the non-equal interval time encoding result corresponding to the current management time point. Indicates the first The number of managed objects in the first The actual time interval between each management time point and the previous management time point Indicates a unit of time scale;

[0132] It should be noted that, , , , and These are all trainable mapping parameters, used to uniformly map corresponding inputs to the same feature dimension before fusion.

[0133] S4.2: Input the fused and associated risk inputs into the non-equal interval time-encoded Transformer model in the order of management time points, extract the time-series risk representation through the self-attention mechanism with time interval decay, and output the risk probability within the current warning period based on the time-series risk representation of the current management time point.

[0134] Specifically, the fusion-related risk inputs for the same managed object are arranged from earliest to latest according to the management time point, forming a fusion-related risk input sequence. To ensure the consistency of the model input length, the O nearest management time points before the current management time point are selected as the input window. When the number of historical management time points for the managed object is less than O, blank inputs are added to the front of the sequence and an invalid mask is set. When the number of historical management time points exceeds O, the sequence of management time points closest to the current management time point is retained, thus forming a fixed-length fusion-related risk input sequence that can be input into the model. This sequence is then input into a pre-trained non-equidistant time-encoded Transformer model. The model outputs the hidden state corresponding to the current management time point based on the input content, actual time interval, and data reliability of each management time point, and obtains the risk probability within the current warning period through the risk probability output layer.

[0135] It should be noted that the non-uniform time-encoded Transformer model includes an input mapping layer, a time interval encoding layer, a time-decay self-attention layer, a feedforward network layer, a normalization layer, and a risk probability output layer. The input mapping layer converts the fused and correlated risk input into a latent vector of uniform dimension; the time interval encoding layer converts the actual time interval between adjacent management time points into a time interval representation and combines it with the latent vector of the corresponding management time point; the time-decay self-attention layer calculates the association weights between different management time points and attenuates the impact of the actual time interval between different management time points on distant historical management events. Specifically, the fused and correlated risk input sequence is mapped to a query matrix Q, a key matrix K, and a value matrix V, and the attention output with time interval decay is calculated as follows:

[0136]

[0137] In the formula, This represents a matrix indicating the time intervals between different management points for the same managed object. Indicates the time decay coefficient. Indicates the dimensions of the query vector and key vector;

[0138] The feedforward network layer performs a non-linear transformation on the latent vectors output by the self-attention layer; the normalization layer stabilizes the model training and inference processes; and the risk probability output layer converts the latent state corresponding to the current management time point into the risk probability within the current warning period.

[0139]

[0140] In the formula, Indicates the first The number of managed objects in the first The risk probability of the current early warning period corresponding to each management time point This represents the current hidden state output by the non-equal-interval time-encoded Transformer model. This represents the weight vector of the risk probability output layer. Indicates the bias term;

[0141] During the model training phase, historical disease-specific population management data is used as training data. This data undergoes the aforementioned multi-source data association, effective management event sequence generation, non-uniform interval time-series management sample construction, data credibility factor calculation, disease-specific risk state tensor construction, personnel transmission risk vector extraction, and event transmission risk vector extraction to form a historical fusion-related risk input sequence. Simultaneously, based on the actual management results after the end of the historical warning cycle, each historical sample is labeled with whether a disease-specific management abnormal event occurred, forming a training label. During training, the historical fusion-related risk input sequence is input into the non-uniform interval time-encoding Transformer model. The model outputs the predicted risk probability corresponding to the historical sample, and the training loss is calculated based on the difference between the predicted risk probability and the training label. Subsequently, the model parameters of the input mapping layer, time interval encoding layer, time decay self-attention layer, feedforward network layer, and risk probability output layer are updated through backpropagation until the predicted loss on the validation set meets the preset convergence condition. After training, the disease-specific population management platform solidifies the model parameters, resulting in a pre-trained non-uniform interval time-encoding Transformer model.

[0142] It should be further explained that, The value range can be set to [0, +∞). The basis for its value is that the time decay term should reduce the attention weight of historical events as the management time interval increases. Therefore, non-negative trainable parameters are used, and the parameters are determined by backpropagation based on historical disease management samples during the model training process.

[0143] The value range of can be set to the real number domain (−∞, +∞), and its value is based on the fact that the risk probability output layer belongs to the conventional fully connected logistic regression output layer, and the weight vector The values ​​are automatically learned through supervised training by historical fusion of risk inputs and abnormal labels in disease management, and can be constrained by regularization to avoid excessive values.

[0144] S5. Calculate the risk acceleration by combining the risk probability of adjacent early warning cycles and the actual time interval between adjacent effective management events, and generate a corrected early warning score by combining the data credibility factor. Determine the graded early warning level based on the corrected early warning score, and generate disease-specific population management tasks according to the graded early warning level.

[0145] S5.1: Based on the current risk probability, the historical risk probability corresponding to adjacent management time points, and the actual time interval between adjacent management time points, calculate the unit time risk change rate of the current stage and the previous stage, and generate a risk acceleration index based on the difference between the two.

[0146] Specifically, based on the current risk probability, the platform reads the historical risk probabilities of the same managed object at the previous and two previous management time points, as well as the actual time interval between adjacent management time points. For managed objects with at least three valid management time points, the platform first calculates the unit-time risk change rate for the current stage based on the difference between the current risk probability and the risk probability at the previous management time point, and the actual time interval between the current and previous management time points. Then, it calculates the unit-time risk change rate for the previous stage based on the difference between the risk probability at the previous and two previous management time points, and the actual time interval between the previous and two previous management time points. Subsequently, the difference between the unit-time risk change rate for the current stage and the unit-time risk change rate for the previous stage is used as the risk acceleration index to represent the degree of change in the rate of risk increase for the managed object. For managed objects with fewer than three valid management time points, if there are only two valid management time points, the unit-time risk change rate for the current stage is used as the risk acceleration index; if there is only one valid management time point, the risk acceleration index is recorded as zero.

[0147] S5.2: Generate a revised early warning score based on the current risk probability, risk acceleration index, and data credibility factor.

[0148] Specifically, based on the current risk probability, risk acceleration index, and data credibility factor, the current risk acceleration index is first normalized according to the range of risk acceleration values ​​in historical training samples, so that it is within a fusionable numerical range with the risk probability and data credibility. When the normalization result exceeds the preset value range, it is limited to that value range to maintain score stability. The current risk probability, the normalized risk acceleration index, and the data credibility decay term are then weighted and fused to generate a corrected warning score.

[0149] It should be noted that the current risk probability is used to represent the basic risk level of the managed object in the current warning period, the normalized risk acceleration index is used to represent the rate of risk deterioration, and the data credibility decay term is used to represent the impact of missing data, incomplete review, incomplete follow-up, upload delay, or failure of device to upload on the credibility of the warning.

[0150] The weighting coefficients used in weighted fusion are determined by historical labeled samples before the model goes live and remain fixed within the same deployment cycle;

[0151] Through this revised early warning score, the platform can simultaneously identify management targets with a high current risk probability, a rapid increase in risk, and those requiring early intervention due to declining data credibility.

[0152] S5.3: Based on the corrected warning score obtained in the previous step, read the corrected warning score of each target disease management object at the current management time point according to the unique identifier of the management object, and determine the graded warning level according to the preset grading threshold;

[0153] It should be noted that below the follow-up reminder threshold It was determined to be at the routine management level, reaching the follow-up reminder threshold but not the key review threshold. It has been determined to be at the follow-up reminder level, reaching the key review threshold but not the immediate warning threshold. The criteria are designated as key review levels, and those that reach the immediate warning threshold are designated as immediate warning levels.

[0154] Follow-up reminder threshold The range of values ​​is It can usually be determined within the middle and high percentile of the historical distribution of the early warning score. Its value is based on the score distribution of mild management risk subjects such as abnormal follow-up, overdue re-examination, and missing data in the historical disease population management sample, as well as the periodic handling capacity of the grassroots follow-up task.

[0155] Key review threshold The range of values ​​is The value can usually be determined within the higher quantile of the historical distribution of the warning score. The value is based on the score distribution of objects in the historical sample that have been reviewed and confirmed by management personnel to have a continuous increase in risk, incomplete follow-up review, or multiple management abnormalities, as well as the periodic handling ability of the review personnel.

[0156] Real-time warning threshold The range of values ​​is The value can usually be determined within the highest quantile interval of the historical distribution of the warning score. The value is based on the score distribution of objects that need immediate intervention or are transferred to the key management queue in the historical sample, the capacity of immediate disposal resources, and the recall rate requirements for high-risk objects.

[0157] S5.4: Generate corresponding disease-specific population management tasks based on the graded early warning level;

[0158] Among them, the regular management level corresponds to the generation of periodic management reminder tasks, the follow-up reminder level corresponds to the generation of follow-up reminders, review reminders or data re-transmission reminder tasks, the key review level corresponds to the generation of management personnel review tasks, and the immediate warning level corresponds to the generation of high-priority warning and handling tasks. Each management task records the unique identifier of the managed object, the target disease code, the graded warning level, the corrected warning score, the current risk probability, the risk acceleration index, the data credibility factor, and the task trigger time to ensure that the source of the task is traceable.

[0159] Reference Figure 2 This embodiment also provides a deep learning-based triage and early warning system for high-risk groups of specific diseases, including:

[0160] The data collection module is used to acquire and associate multi-source management data of the target disease management object;

[0161] The event generation module is used to preprocess data on the management of specific disease populations and generate effective management event sequences;

[0162] The state encoding module is used to generate data credibility factors based on non-equal interval time-series management samples, and to encode effective management events, time features, management delay features and data credibility factors to generate a disease-specific risk state tensor.

[0163] The graph structure construction module is used to construct personnel similarity graphs, disease event graphs, and object-event association matrices based on disease risk state tensors.

[0164] The associated risk extraction module is used to extract personnel transmission risk vectors based on personnel similarity graphs and to extract event transmission risk vectors based on disease-specific event graphs and object-event association matrices.

[0165] The early warning and response module is used to input the personnel transmission risk vector, event transmission risk vector, and disease-specific risk state tensor into the deep learning risk prediction model to generate risk probabilities, correct early warning scores, classify early warning levels, and corresponding disease-specific population management tasks.

[0166] In summary, this invention integrates multi-source data, including basic records, medical visits, examinations, medication, follow-ups, medical insurance, vital sign monitoring, and device uploads, through a unique identifier for the managed object, forming a continuous and traceable data chain for disease management. By jointly modeling non-uniformly spaced time-series samples, data reliability factors, personnel similarity graphs, and disease event graphs, it can simultaneously reflect changes in individual status, risk migration among similar populations, and the transmission relationship of management events. Furthermore, by combining deep learning risk prediction, risk acceleration, and corrective early warning scoring, it achieves hierarchical identification of high-risk individuals, task generation, and priority ranking of treatment, improving the real-time performance, accuracy, and closed-loop management capabilities of disease-specific population early warnings.

[0167] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for tiered early warning of high-risk groups for specific diseases based on deep learning, characterized in that, include: Acquire multi-source management data of target disease management objects, generate disease population management data by association according to the unique identifier of the management object, preprocess the disease population management data, generate effective management event sequences sorted by time, and generate non-equal interval time series management samples by combining the actual time interval of events and management delay characteristics. Data credibility factors are generated based on the data missing status, data upload delay status, and follow-up completion status in the non-equal interval time series management samples, and the disease risk status tensor is encoded. Based on the disease risk status tensor, personnel similarity graph, disease event graph, and object-event association matrix are constructed. Based on the personnel similarity graph, we perform comorbidity hierarchical edge gating and personnel maximum clique enhancement to extract personnel transmission risk vectors. Based on the disease-specific event graph and object-event association matrix, we perform event type gating and temporal association aggregation to extract event transmission risk vectors. By fusing the personnel transmission risk vector, the event transmission risk vector, and the disease-specific risk state tensor into a deep learning risk prediction model, the risk probability within the current warning period can be obtained. The risk acceleration is calculated by combining the risk probability of adjacent early warning cycles and the actual time interval between adjacent effective management events, and a revised early warning score is generated by combining the data credibility factor. The graded early warning level is determined based on the revised early warning score, and special disease population management tasks are generated according to the graded early warning level.

2. The method for tiered early warning of high-risk groups for specific diseases based on deep learning as described in claim 1, characterized in that: The multi-source management data includes basic medical records, historical medical records, examination and testing data, medication execution data, follow-up plan data, actual follow-up data, follow-up plan data, actual follow-up data, medical insurance settlement data, vital sign monitoring data, device upload data, and comorbidity label data for the target disease management subjects. The disease-specific population management data is a set of multi-source management data formed by associating the multi-source management data corresponding to each target disease management subject according to the unique identifier of the management subject.

3. The method for graded early warning of high-risk groups for specific diseases based on deep learning as described in claim 2, characterized in that: The steps for generating a time-ordered valid management event sequence and combining the actual time intervals of the events with management delay characteristics to generate non-uniformly spaced time-series management samples are as follows: The preprocessed disease-specific population management data are grouped according to the unique identifier of the managed object and sorted by the time of event occurrence to generate a valid management event sequence corresponding to each managed object; Valid management events at the same management time point are grouped into an event set, the actual interval between adjacent management time points is calculated, and management delay characteristics are generated based on events related to review, follow-up, indicator upload, equipment upload, and indicator anomaly, to obtain non-equal interval time series management samples.

4. The method for graded early warning of high-risk groups for specific diseases based on deep learning as described in claim 3, characterized in that: The steps for generating data reliability factors based on data missing status, data upload delay status, and follow-up completion status in non-equal interval time-series management samples, and encoding them into a disease risk state tensor, are as follows: Generate data reliability factors corresponding to each management time point; convert the numerical event values ​​of each effective management event in the non-equal interval time-series management samples into model input numerical features; convert the non-numerical event information of each effective management event into event type vectors; aggregate the model input numerical features and event type vectors under the same management time point to generate an event state expression; concatenate the event state expression with the non-equal interval time features, management delay features, and data reliability factors to generate a disease risk state tensor.

5. The method for tiered early warning of high-risk groups for specific diseases based on deep learning as described in claim 4, characterized in that: The steps for constructing the personnel similarity graph, the disease event graph, and the object-event association matrix based on the disease risk state tensor are as follows: Based on the disease risk state tensor, management profile features of each target disease management object are extracted, management object profile vectors are generated, and personnel similarity edges are established according to the similarity between the management object profile vectors to construct a personnel similarity graph. Disease-specific management events are extracted based on disease-specific risk state tensors, the temporal correlation between different disease-specific management events is statistically analyzed, a disease-specific event graph is constructed, and an object-event correlation matrix is ​​constructed based on the correspondence between each management object and a disease-specific management event.

6. The method for tiered early warning of high-risk groups for specific diseases based on deep learning as described in claim 5, characterized in that: The extraction of the personnel transmission risk vector includes using a personnel similarity graph, utilizing event state expressions, event type vectors corresponding to comorbidity labels, and management delay features to perform comorbidity hierarchical gating on personnel similarity edges, and obtaining the gated personnel similarity edge weights; Based on the identification of the maximum clique structure formed by personnel similarity edges in the personnel similarity graph, and based on the maximum clique structure, the weight of the personnel similarity edges after gating is enhanced. Then, attention aggregation is performed on the disease risk state vectors of adjacent managed objects to obtain the personnel transmission risk vector.

7. The method for graded early warning of high-risk groups for specific diseases based on deep learning as described in claim 6, characterized in that: The extraction of the event transmission risk vector includes gating the event association edges based on the disease-specific event graph and the event type and temporal correlation of disease-specific management events to obtain enhanced event association edge weights, and then aggregating and generating event transmission vectors accordingly. Based on the object-event correlation matrix, the event transmission vector is mapped to the corresponding target disease management object to obtain the event transmission risk vector.

8. The method for tiered early warning of high-risk groups for specific diseases based on deep learning as described in claim 7, characterized in that: The steps to obtain the risk probability within the current early warning period are as follows: The risk vectors of personnel transmission, event transmission, disease-specific risk status, non-equidistant time intervals, and data credibility factors are mapped to the same feature dimension and fused to generate a fused associated risk input. The fused associated risk input is then input into the non-equidistant time-encoded Transformer model in the order of management time points. The time-series risk representation is extracted through a self-attention mechanism with time interval decay, and the risk probability within the current warning period is output based on the time-series risk representation at the current management time point.

9. The method for graded early warning of high-risk groups for specific diseases based on deep learning as described in claim 8, characterized in that: The steps for calculating risk acceleration by combining the risk probability of adjacent early warning periods and the actual time interval between adjacent effective management events, and generating a corrected early warning score by combining the data credibility factor, are as follows: Based on the current risk probability, the historical risk probability corresponding to adjacent management time points, and the actual time interval between adjacent management time points, calculate the unit time risk change rate of the current stage and the previous stage, and generate a risk acceleration index based on the difference between the two. A revised early warning score is generated based on the current risk probability, risk acceleration index, and data credibility factor.

10. A deep learning-based hierarchical early warning system for high-risk groups of specific diseases, based on the deep learning-based hierarchical early warning method for high-risk groups of specific diseases as described in any one of claims 1 to 9, characterized in that, include: The data collection module is used to acquire and associate multi-source management data of the target disease management object; The event generation module is used to preprocess data on the management of specific disease populations and generate effective management event sequences; The state encoding module is used to generate data credibility factors based on non-equal interval time-series management samples, and to encode effective management events, time features, management delay features and data credibility factors to generate a disease-specific risk state tensor. The graph structure construction module is used to construct personnel similarity graphs, disease event graphs, and object-event association matrices based on disease risk state tensors. The associated risk extraction module is used to extract personnel transmission risk vectors based on personnel similarity graphs and to extract event transmission risk vectors based on disease-specific event graphs and object-event association matrices. The early warning and response module is used to input the personnel transmission risk vector, event transmission risk vector, and disease-specific risk state tensor into the deep learning risk prediction model to generate risk probabilities, correct early warning scores, classify early warning levels, and corresponding disease-specific population management tasks.

Images