Patient transition image construction and evaluation system oriented to dynamic timeline feature extraction

Abstract

The application relates to the technical field of medical data processing, and discloses a patient conversion image construction and evaluation system facing dynamic time axis feature extraction, wherein the system comprises an asynchronous data acquisition and semantic segmentation module, a global event anchor point generation module, a state embedding generation module, a causal network snapshot generation module, a cross-chain competition quantification module, a conversion image generation module, an incremental update module, and a dynamic priority evaluation and result mapping module. The system detects semantic segmentation boundaries of each data source based on an information entropy change rate, generates a dynamic semantic time period that is uniform across sources, generates a time period state node vector through a modal adaptive encoder, performs Shapley value distribution on shared intermediate indicator nodes and constructs a cross-chain coupling influence coefficient tensor, and combines constraint optimization to solve and output a dynamic priority sequence of multiple disease conversion paths and a competition state trend drift early warning evaluation result.

Description

Technical Field

[0001] This invention relates to the field of medical data processing technology, and more specifically, to a patient conversion profile construction and evaluation system oriented towards dynamic timeline feature extraction. Background Technology

[0002] In the long-term cross-departmental collaborative management of patients with multiple coexisting diseases, departments such as endocrinology, nephrology, and cardiology generate asynchronous and heterogeneous medical data streams according to their respective treatment plans. The collection cycle and data mode of each data source are independent of each other, and there are observation gaps during the treatment intervals.

[0003] Existing technologies typically extract time-series features from data for each disease dimension separately, generate single-disease-dimensional transformation profiles, and then present the multi-dimensional profiles side by side.

[0004] The above approach has two shortcomings: First, the data from different departments are naturally asynchronous on the timeline, and extracting features by independent time series will lose information on the collaborative changes across disease dimensions, making it difficult to form a unified dynamic timeline that reflects the overall evolution of the patient's condition. Second, patients with multiple coexisting diseases share some intermediate indicator nodes in their multiple disease causal chains. The transmission and advancement of one causal chain will inhibit or accelerate the transmission of another chain. Parallel assessment of independent profiles ignores the competition and coupling between chains, leading to inaccurate assessment of conversion risk and making it difficult to dynamically determine the most urgent conversion path. Summary of the Invention

[0005] This invention provides a patient conversion profile construction and evaluation system for dynamic time axis feature extraction, which solves the technical problems in related technologies such as the difficulty in unifying the time alignment of asynchronous medical data across departments for patients with multiple coexisting diseases, the inability to quantify the competitive relationship between causal chains across diseases, and the inability to dynamically evaluate the priority of multiple disease conversion paths.

[0006] This invention discloses a patient conversion profile construction and evaluation system for dynamic time axis feature extraction, including an asynchronous data acquisition and semantic segmentation module, configured to acquire asynchronous medical data streams generated by various departments of patients with multiple coexisting diseases, detect semantic segmentation boundary points for each data source based on the local information entropy change rate, and generate a set of local segmentation results for each source. The global event anchor point generation module is configured to calculate the cross-source boundary correlation matrix based on the local segmentation result set of each source, merge the cross-source boundary points that meet the correlation threshold into global event anchor points, and define a unified dynamic semantic time period across sources. The state embedding generation module is configured to perform modality adaptive coding on heterogeneous indicators of each disease dimension within each dynamic semantic time period to generate modality-independent state node vectors. The causal network snapshot generation module is configured to perform causal direction determination based on transfer entropy for cross-disease dimension indicator pairs within each dynamic semantic time period, generate multi-disease causal network snapshots, and identify cross-disease shared intermediate indicator nodes. The cross-chain competition quantification module is configured to use the Shapley value allocation method to calculate the competitive allocation share of each disease causal chain for shared intermediate indicator nodes, and generate a shared node competition allocation matrix and a cross-chain coupling influence coefficient tensor. The patient conversion profile generation module is configured to connect the state node vectors of each dynamic semantic time period, the snapshot of the multi-disease causal network, and the shared node competition allocation matrix according to the global event anchor sequence to generate a patient conversion profile. The incremental update module is configured to locally recalculate the competition allocation for newly added dynamic semantic time periods when new department data arrives, generating a cross-chain competition situation vector. The dynamic priority assessment and result mapping module is configured to perform constraint optimization based on the cross-chain competitive situation vector and the cross-chain coupling influence coefficient tensor, and output the dynamic priority ranking of multiple disease transformation paths and the competitive situation drift early warning assessment results.

[0007] Furthermore, in the asynchronous data acquisition and semantic segmentation module, the data source... ordered timestamp sequence By time point Calculate local information entropy by the distribution of index values ​​within a local observation window centered on the local center. The rate of change of information entropy between adjacent observation time points is: in, For time points The rate of change of information entropy relative to the previous observation point; when Greater than the history of this data source When the preset quantile value of the distribution is used, The semantic segmentation boundary points marked as the data source are denoted as the quantile parameters corresponding to the preset quantiles. ,and .

[0008] Furthermore, in the global event anchor point generation module, the cross-source boundary correlation matrix elements By time proximity factor Covariance factors of cross-source indicators Sure: in, For data source No. The timestamps of the boundary points For data source No. The timestamps of the boundary points The time decay coefficient and , The mutual information value between available indicators from two data sources within a local time window centered at two boundary points, after range normalization. ; for satisfying Perform a merge operation on cross-source boundary point pairs. The correlation threshold and The cross-source boundary points that meet the conditions are aggregated into global event anchor points. The anchor point timestamp is the weighted median time obtained by weighting the timestamps of each boundary point in the aggregation point set according to the correlation.

[0009] Furthermore, in the state embedding generation module, the modal adaptive encoder consists of a continuous numerical index coding branch, a time-series segment index coding branch, an image evaluation summary index coding branch, and an attention-weighted fusion layer; the loss function of the modal adaptive encoder is: in, For modal branch index, The total number of modal branches, For the first The reconstruction error weights of each modal branch and , For the first Reconstruction error of each modal branch For cross-modal consistency constraint weights and , Pick Subtract the mean of the pairwise cosine similarity of the intermediate representation vectors of different modal branches within the same dynamic semantic time period.

[0010] Furthermore, the state embedding generation module is also configured to perform implicit state transition estimation for dynamic semantic time periods lacking observation data during the treatment interval. Based on the observation values ​​of each available index at the two most recent global event anchor points before and after the treatment interval, the continuous index is estimated using a monotonic constraint interpolation algorithm in the standardized space, and the categorical index is estimated using a recent effective encoding value preservation strategy. The estimation results are combined to generate a virtual state node vector to fill the corresponding dynamic semantic time period position.

[0011] Furthermore, in the causal network snapshot generation module, the indicators... and indicators Calculate the transfer entropy separately and ,when Time determination to There is a causal relationship between directions. Threshold for determining causal direction and Perform range normalization on all original causal strengths within the same dynamic semantic time period to obtain edge weights; generate multi-disease causal network snapshots. ,in For the set of indicator nodes, For a set of causal directed edges, This is the normalized set of edge weights; indicator nodes that are simultaneously located on the internal paths of two or more different disease causal chains are identified as cross-disease shared intermediate indicator nodes.

[0012] Furthermore, in the cross-chain competition quantization module, the cross-chain coupling influence coefficient tensor The element is defined as: in, and Indexing causal chains for different diseases For the first A causal chain with shared intermediate indicator nodes Market share of competition For the first The causal chain enters the shared intermediate indicator node. The weight of the incoming edge at the end, To share intermediate indicator nodes Along the first The weight of the outgoing edge from the direction of the causal chain to the nearest converging node. As an indicator function, when the sink node is simultaneously located at the _th Article and Section When the value is taken on the downstream path of the causal chain, it is... Otherwise, the value is ;when season .

[0013] Furthermore, in the incremental update module, the cross-chain competition situation vector The The components are: in, This is the set of shared intermediate indicator nodes for the current dynamic semantic time period. For the first A causal chain with shared intermediate indicator nodes Market share of competition To share intermediate indicator nodes In the Path weights on the causal chain transmission path and The node is at the 1st The weights of incoming and outgoing edges on a causal chain.

[0014] Furthermore, in the dynamic priority evaluation and result mapping module, the actual achievable propagation progress of each causal chain is considered. Given the decision variables, solve the linear programming problem: in, For the first The first constraint represents the actual reachability of the causal chain; the second constraint reflects the upper limit of the competitive capacity of each shared intermediate index node; and the third constraint incorporates the cross-chain coupling effect into the upper bound of the reachability of each chain. The solution obtained... Generate a dynamic priority sequence of multi-disease transformation pathways by sorting from high to low; accelerate the change in competitive share. The causal chain generates drift warning tags. To preset the warning threshold and .

[0015] This invention provides a method for constructing and evaluating patient conversion profiles based on dynamic timeline feature extraction, comprising: Step 1: Obtain asynchronous medical data streams generated by various departments for patients with multiple coexisting diseases; detect semantic segmentation boundary points for each data source based on the local information entropy change rate; and generate a set of local segmentation results for each source. Step 2: Calculate the cross-source boundary correlation matrix based on the local segmentation results set of each source, merge the cross-source boundary points that meet the correlation threshold into global event anchor points, and delineate a unified dynamic semantic time period across sources; Step 3: Perform modality adaptive coding on the heterogeneous indicators of each disease dimension within each dynamic semantic time period to generate modality-independent state node vectors; Step 4: Perform causal direction determination based on transfer entropy for cross-disease dimension indicator pairs within each dynamic semantic time period, generate multi-disease causal network snapshots and identify cross-disease shared intermediate indicator nodes; Step 5: Use the Shapley value allocation method to calculate the competitive allocation share of each disease causal chain for the shared intermediate index nodes, and generate the shared node competitive allocation matrix and the cross-chain coupling influence coefficient tensor. Step 6: Connect the state node vectors of each dynamic semantic time period, the multi-disease causal network snapshots, and the shared node competition allocation matrix according to the global event anchor sequence to generate a patient conversion profile; Step 7: When the data from the new department arrives, locally recalculate the competition allocation for the newly added dynamic semantic time period to generate a cross-chain competition situation vector; Step 8: Perform constraint optimization based on the cross-chain competitive situation vector and the cross-chain coupling influence coefficient tensor, and output the dynamic priority ranking of multi-disease transformation paths and the early warning assessment results of competitive situation drift.

[0016] The present invention has the following beneficial effects: Firstly, by detecting the semantic segmentation boundary within each data source using the information entropy change rate, and generating global event anchors using the temporal proximity factor and the cross-source index covariance factor, asynchronous observations across departments are mapped to a unified dynamic semantic time period. This allows the synergistic changes generated by different disease dimensions at similar clinical stages to be expressed within the same dynamic semantic time period, avoiding the situation where cross-disease synergistic information is scattered at different time positions when analyzing independent time axes in parallel.

[0017] Secondly, by using a modal adaptive encoder, continuous numerical indicators, time-series segment indicators, and image assessment summary indicators are mapped to a shared state embedding space. For the treatment interval, a monotonic constraint interpolation and a recent effective value preservation strategy are used to generate virtual state node vectors, so that the missing observation dynamic semantic time period and the actual observation dynamic semantic time period adopt the same data expression form, and the state evolution process on the unified dynamic time axis can be kept continuous.

[0018] Third, by performing Shapley value allocation on shared intermediate indicator nodes across diseases and constructing a cross-chain coupling influence coefficient tensor, the patient conversion profile contains the competitive and coupling relationships of each disease causal chain at the shared intermediate indicator nodes, which can support the determination of the most urgent conversion path.

[0019] Fourth, when new department data arrives, only the newly added dynamic semantic time period and its associated shared intermediate indicator nodes are locally recalculated. Combined with the change rate, change acceleration and constraint optimization solution results of the cross-chain competitive situation vector, a dynamic priority sequence of multi-disease conversion paths and competitive situation drift early warning assessment results are generated. This allows the path priority to be updated synchronously with the current competitive situation, avoiding the situation where it is difficult to distinguish the urgency of conversion when only static parallel profiles are used for assessment. Attached Figure Description

[0020] Figure 1 This is a flowchart of the patient conversion profile construction and evaluation method based on dynamic time axis feature extraction provided in this embodiment of the invention; Figure 2This is a schematic diagram of the causal chain structure and shared nodes of multiple coexisting diseases provided in the embodiments of the present invention; Figure 3 This is a schematic diagram of the asynchronous data processing and dynamic semantic time period division flowchart provided in the embodiments of the present invention; Figure 4 This is a schematic diagram comparing the rate of change of information entropy of semantic segmentation boundary points of various data sources provided in the embodiments of the present invention; Figure 5 This is a schematic diagram of the normalized edge weight distribution of the segment 4 multi-disease causal network provided in an embodiment of the present invention; Figure 6 This is a schematic diagram of the cross-chain competition allocation matrix of the segment 4 shared intermediate index node provided in an embodiment of the present invention; Figure 7 This is a schematic diagram illustrating the evolution trend of the cross-segment competition situation vector of the three causal chains provided in the embodiment of the present invention; Figure 8 This is a schematic diagram of the comprehensive result of dynamic priority evaluation of the three causal chains in segment 4 provided in an embodiment of the present invention; Figure 9 This is a schematic diagram illustrating the time-series directed graph structure of the transformed portrait provided in an embodiment of the present invention. Detailed Implementation

[0021] Example 1 In long-term cross-departmental collaborative management scenarios for patients with multiple coexisting diseases, each department generates asynchronous and heterogeneous medical data streams according to its own treatment plan. For example, the endocrinology department generates blood glucose monitoring sequences, the nephrology department generates renal function index sequences, and the cardiology department generates electrocardiogram and imaging assessment data. These data belong to different collection periods and modal types. When it is necessary to assess the conversion risk of patients with multiple coexisting diseases, the existing methods usually extract time-series features from the data of each disease dimension separately and generate conversion profiles for each disease dimension, and then present the multi-dimensional profiles side by side. This method has two shortcomings: First, the data of each department are naturally asynchronous on the time axis and there are observation gaps in the treatment intervals. Extracting features by independent time series will lose information on the collaborative changes across disease dimensions, making it difficult to form a unified dynamic time axis that reflects the overall evolution of the patient's condition. Second, the multiple disease causal chains of patients with multiple coexisting diseases share some intermediate indicator nodes. The transmission of one causal chain will inhibit or accelerate the transmission of another chain. The parallel assessment of independent profiles ignores the competition and coupling relationship between chains, resulting in inaccurate conversion risk assessment and difficulty in dynamically determining the most urgent conversion path.

[0022] This embodiment provides a method for constructing and evaluating patient conversion profiles based on dynamic timeline feature extraction. It works collaboratively through asynchronous data acquisition and semantic segmentation modules, global event anchor generation modules, state embedding generation modules, causal network snapshot generation modules, cross-chain competition quantification modules, conversion profile generation modules, incremental update modules, and dynamic priority evaluation and result mapping modules to dynamically quantify the competitive relationships of causal chains of multiple diseases and determine the priority of conversion paths.

[0023] The asynchronous data acquisition and semantic segmentation module is configured to acquire asynchronous medical data streams generated by various departments for patients with multiple coexisting diseases, maintain independent timestamp sequences according to data sources, detect the internal semantic segmentation boundaries of each data source, and generate a set of local segmentation results for each source.

[0024] Acquire asynchronous medical data streams from various departments for patients with multiple coexisting conditions. Each data source corresponds to an independent timestamp sequence. Data sources include blood glucose monitoring data streams from the endocrinology department, renal function index data streams from the nephrology department, and electrocardiogram and imaging assessment data streams from the cardiology department. The acquisition period and data modality of each data source are independent of each other.

[0025] Data preprocessing was performed on the raw indicators from each data source. Z-score standardization was performed on continuous numerical indicators, and one-hot encoding was performed on non-numerical indicators such as image assessment grading, ECG conclusion labels, and medication categories. All timestamps were uniformly converted to the same time unit. The standardized and encoded indicator sequences were used for subsequent information entropy calculation, mutual information calculation, modal adaptive coding, and transfer entropy analysis.

[0026] For data source ordered timestamp sequence By time point Calculate the local information entropy based on the distribution of preprocessed index values ​​within the local observation window centered on the center. ,in, For data source identification, For data source The Each observation time point For the index of observation time points, For data source The total number of observation time points, For time points The Shannon entropy corresponds to the local observation window. Further calculation of the rate of change of information entropy between adjacent observation time points: in, For time points The rate of change of information entropy relative to the previous observation point, and , The unit is consistent with the unified time unit. Greater than the history of this data source When the preset quantile value of the distribution is used, The semantic segmentation boundary points marked as the data source are denoted as the quantile parameters corresponding to the preset quantiles. ,and After performing the above processing on each data source, a set of local segmentation results for each source is generated. Each set contains all semantic segmentation boundary points within that data source and their corresponding information entropy change rate values.

[0027] The global event anchor point generation module is configured to calculate the boundary correlation matrix based on the local segmentation result set of each source, merge cross-source boundary points that meet the correlation conditions into global event anchor points, and define a unified dynamic semantic time period across sources.

[0028] For any two different data sources and Calculate the cross-source boundary correlation matrix based on the segmentation boundary points. Matrix elements Indicates data source The Boundary points and data sources The The correlation degree between boundary points. The correlation degree is jointly determined by the temporal proximity factor and the cross-source index covariance factor.

[0029] Temporal proximity factor Defined as: in, The temporal proximity factor between two cross-source boundary points. Represents an exponential function. For data source No. The timestamps of the boundary points For data source No. The timestamps of the boundary points The unit is consistent with the unified time unit. Let be the time decay coefficient, and Its unit is a reciprocal of the unified unit of time. A larger value indicates greater sensitivity to time intervals.

[0030] Cross-source index covariance factor Within a local time window centered on two boundary points, the mutual information values ​​between available indicators from two data sources are selected. Continuous indicators are processed using standardized results, while categorical indicators are processed using coded results. Range normalization is performed on the mutual information values ​​to obtain the dimensionless cross-source indicator covariance factor. ,and The elements of the boundary correlation matrix are calculated as follows: in, For data source The Boundary points and data sources The The degree of correlation between boundary points For time proximity factor, This represents the normalized cross-source index covariance factor. Because... and ,therefore .

[0031] To satisfy The cross-source boundary point pairs are merged, where, This is the correlation threshold, and Cross-source boundary points that are temporally adjacent and meet a threshold condition are aggregated into a single global event anchor point. The anchor point timestamp is the weighted median time obtained by weighting the timestamps of all boundary points in the aggregation set according to the sum of the correlation between that boundary point and other boundary points in the same aggregation set. All global event anchor points are arranged in chronological order to generate a global event anchor point sequence. The overall patient observation timeline is divided into cross-source unified dynamic semantic time periods using adjacent global event anchor points as boundaries; the first and last dynamic semantic time periods are retained before the first anchor point and after the last anchor point, respectively.

[0032] The state embedding generation module is configured to perform modality adaptive coding on heterogeneous indicators of each disease dimension within each dynamic semantic time period, generate modality-independent time period state node vectors, and generate virtual state node vectors for treatment intervals.

[0033] For each dynamic semantic time period, identify the subset of available indicators for each disease dimension within that time period. Due to different data collection cycles across departments, only some indicators for certain disease dimensions may have valid observations within some dynamic semantic time periods.

[0034] The modality adaptive encoder takes a preprocessed subset of available indicators from each disease dimension as input and maps heterogeneous modal data to a shared state embedding space. The modality adaptive encoder consists of continuous numerical indicator encoding branches, time-series segment indicator encoding branches, image assessment summary indicator encoding branches, and an attention-weighted fusion layer. Continuous numerical indicators are processed through a fully connected encoding branch, time-series segment indicators through a recurrent encoding branch, and image assessment summary indicators are processed using encoded structured matrices through a convolutional encoding branch. The outputs of each branch are fused by the attention-weighted fusion layer to generate a fixed-dimensional modality-independent time-time state node vector. ,in, For dynamic semantic time period index, For the first A state node vector for a dynamic semantic time period. To share the embedded spatial dimension.

[0035] The modal adaptive encoder employs a multi-task learning and training mode, with each encoding branch sharing the parameters of the attention-weighted fusion layer. The output layer has a dimension of [missing information]. The fully connected layer. The loss function is a weighted sum of the reconstruction errors of each branch and the cross-modal consistency constraints: in, This represents the total loss of the modal adaptive encoder. For modal branch index, The total number of modal branches, For the first The reconstruction error weights of each modal branch and , For the first The reconstruction error of each modal branch is calculated as the mean square error between the predicted output and the corresponding standardized or encoded input. For cross-modal consistency constraint weights and , Pick The mean cosine similarity of the intermediate representation vectors of different modal branches within the same dynamic semantic time period is subtracted. Parameter updates are performed using the Adam optimization algorithm.

[0036] For dynamic semantic time periods lacking observational data during treatment intervals, implicit state transition estimation is performed. Based on the observations of each available indicator at the two most recent global event anchor points before and after the treatment interval, the changes in each indicator are calculated. For continuous indicators, a monotonic constrained interpolation algorithm is used to generate estimated values ​​in the standardized space, with the generated estimated values ​​falling between the two observations before and after the treatment interval. For categorical indicators, a nearest-last-valid value preservation strategy is employed. When an indicator has only one-sided valid observations on either side of the treatment interval, the nearest valid value on that side is directly preserved. The interpolation results and preservation results are combined to generate an implicit state transition estimation vector, which is then used as a virtual state node vector to fill the corresponding dynamic semantic time period positions.

[0037] The causal network snapshot generation module is configured to perform causal direction determination on each disease dimension indicator pair within each dynamic semantic time period, generate multi-disease causal network snapshots, and identify cross-disease shared intermediate indicator nodes.

[0038] For each dynamic semantic time period This process acquires disease-dimensional indicator data accumulated within the current dynamic semantic time period and previous dynamic semantic time periods. For all cross-disease dimension indicator pairs, causal direction determination based on transfer entropy is performed. Continuous indicator sequences are Z-score standardized and discretized by quantile intervals before transfer entropy analysis; categorical indicators directly use the encoded results. For the indicators... and indicators Calculate separately and ,in To the indicators To the target The entropy of directional transition To the indicators To the target The entropy of directional transition. When At that time, the judgment to There is a causal relationship between directions, in which, The threshold for determining the causal direction, and The original causal strength in this direction is taken as... Range normalization is performed on all original causal strengths within the same dynamic semantic time period to obtain dimensionless edge weights. The range of edge weight values ​​is as follows: .

[0039] The causal direction determination result is compared with the state node vector. Associated storage, generating the first A snapshot of a multi-disease causal network over a dynamic semantic time period ,in, For the first A snapshot of a multi-disease causal network over a dynamic semantic time period. This refers to the set of indicator nodes involved in this dynamic semantic time period. For a set of causal directed edges, This is the set of normalized edge weights. This serves as a state attribute for a snapshot of the dynamic semantic time period.

[0040] Snapshot of multi-disease causal network In the process, all indicator nodes are traversed, and indicator nodes that are simultaneously located on the internal paths of two or more different disease causal chains are marked as cross-disease shared intermediate indicator nodes, generating a set of shared intermediate indicator nodes. ,in, For the first A set of shared intermediate indicator nodes for a dynamic semantic time period.

[0041] The cross-chain competition quantization module is configured to calculate the competitive allocation share of each disease causal chain for cross-disease shared intermediate index nodes in multi-disease causal network snapshots, and generate a shared node competition allocation matrix and a cross-chain coupling influence coefficient tensor.

[0042] Snapshot of multi-disease causal network Each shared intermediate indicator node The Shapley value allocation method is used to calculate the competitive allocation share of the node's activation state for each disease causal chain. The Shapley value allocation method prioritizes nodes participating in sharing intermediate indexes. The entire set of disease causal chains is used as the set of game participants, with each causal chain passing through a shared intermediate indicator node. The causal transmission strength is used as a marginal contribution measure, and the competitive allocation share of each causal chain to the shared intermediate index node is output.

[0043] The Shapley value allocation results of all shared intermediate index nodes are summarized to generate a shared node competition allocation matrix. ,in, For the first A shared node competition allocation matrix for a dynamic semantic time period. To share the number of intermediate indicator nodes. The number of causal chains in the disease matrix, matrix elements Indicates the first The first causal chain is related to the first The competitive share of shared intermediate indicator nodes, and For any shared intermediate indicator node satisfy ,in, Using the causal chain index, sum and traverse all. A disease causal chain.

[0044] right Based on the shared node competition allocation matrix Calculate the cross-chain coupling influence coefficient tensor Tensor elements are defined as: in, and Indexing causal chains for different diseases For the first The causal chain passes through shared intermediate indicator nodes For the The coupling effect coefficient of the first causal chain, a negative value indicates the first causal chain. The first causal chain is related to the first The inhibitory effect of a causal chain, with positive values ​​indicating an accelerating effect; For the first A causal chain with shared intermediate indicator nodes The share of competition is derived from the shared node competition allocation matrix. ; For the first The causal chain enters the shared intermediate indicator node. The weights of the incoming edges at the end are taken from the set of edge weights. ; To share intermediate indicator nodes Along the first The causal chain direction leads to the nearest convergence node. The outgoing edge weights are taken from the edge weight set. ; To be simultaneously located in the first The first causal chain and the second The nearest converging node on the downstream path of the causal chain; This is an indicator function that takes the value when the aggregation node exists. If it does not exist, the value is taken as .when season .because , and The values ​​are all located in ,therefore The value is located in .

[0045] The patient conversion profile generation module is configured to connect the state node vectors of each dynamic semantic time period, the snapshot of the multi-disease causal network, and the shared node competition allocation matrix according to the global event anchor sequence to generate a patient conversion profile containing the causal competition relationship of multiple diseases.

[0046] Based on the chronological order of the global event anchor sequence, each dynamic semantic time period is... Corresponding state node vector Multi-disease causal network snapshot and shared node competition allocation matrix Ordered connections are established. Temporal associations are created between adjacent dynamic semantic time periods through directed edges. The attributes of these directed edges include differences in the state node vectors of preceding and following dynamic semantic time periods, as well as information on changes in the structure of the multi-disease causal network snapshot. The generated patient transformation profile is a temporal directed graph. Each node in the graph corresponds to a complete state description of a dynamic semantic time period. Node attributes include the state node vector, the multi-disease causal network snapshot topology, and the competition allocation share. Edge attributes include the state transition direction and state transition magnitude.

[0047] The incremental update module is configured to perform local segmentation and anchor matching when new department data arrives, and to recalculate the competition allocation and cross-chain coupling influence coefficient tensor locally for the newly added dynamic semantic time period, generating an incrementally updated cross-chain competition situation vector.

[0048] When new departmental data arrives, local semantic segmentation is performed only on the data sources involved in the new data to obtain the newly added segmentation boundary points within those data sources. The correlation between the newly added segmentation boundary points and the end anchor points of existing global event anchor point sequences is calculated and matched to determine whether the newly added segmentation boundary points trigger the generation of new global event anchor points or are merged into existing global event anchor points.

[0049] Based on the matching results, a new dynamic semantic time period is defined, and the index of the newly added current dynamic semantic time period is recorded as... Modal adaptive coding, transition entropy causal direction determination, and multi-disease causal network snapshot generation are performed on the newly added dynamic semantic time period to obtain the state node vector of the current dynamic semantic time period. Multi-disease causal network snapshot and shared intermediate indicator node set The newly added dynamic semantic time period multi-disease causal network snapshots are connected to the end nodes of the existing patient transformation profiles via directed edges.

[0050] Only the Shapley value allocation and cross-chain coupling influence coefficient tensors are locally recalculated for the newly added dynamic semantic time period and the shared intermediate index nodes involved, to obtain the updated shared node competition allocation matrix. and cross-chain coupling influence coefficient tensor Based on the updated shared node competition allocation matrix, the current competition share value of each disease causal chain is extracted, and an incrementally updated cross-chain competition situation vector is generated. ,in, This is the incrementally updated cross-chain competitive situation vector. Let be the number of causal chains in the disease, and let be the vector of the th . The component represents the first... The overall competitive dominance of a causal chain within the current dynamic semantic time period is defined as: in, For the first The degree of overall competitive dominance of each causal chain within the current dynamic semantic time period. For causal chain index, To share intermediate indicator node indices and traverse the collection All intermediate indicator nodes are shared. For the updated shared node contention allocation matrix, the first... A causal chain with shared intermediate indicator nodes Market share of competition To share intermediate indicator nodes In the The path weights on the causal chain transmission path are taken from the snapshot of the node in the multi-disease causal network. Middle of the first The product of the weights of the incoming and outgoing edges in the direction of the chain is... ,in, and These are shared intermediate indicator nodes. In the The weights of both the incoming and outgoing edges on a causal chain are taken from the set of edge weights. .when When, calculate according to the above formula. When there is no shared intermediate index node in the current dynamic semantic time period or the denominator is zero, let Take the first The proportion of the path weight of a causal chain in the current dynamic semantic time period to the total path weight of all causal chains.

[0051] The dynamic priority assessment and result mapping module is configured to compare the trend of the incrementally updated cross-chain competitive situation vector with the historical competitive situation vector sequence, and combine constraint optimization solution and result mapping to output the dynamic priority ranking of multi-disease transformation paths and the competitive situation drift early warning assessment results.

[0052] The incrementally updated cross-chain competitive situation vector Vector sequence of historical competitive situation Perform trend comparisons, among which Indicates the first The first dynamic semantic time period The degree of competitive dominance in a causal chain. For dynamic semantic time period index, This represents the length of the historical competitive landscape vector sequence. For each disease causal chain... ,when season and ;when At that time, calculate the rate of change of competitive share and let ;when At that time, calculate the rate of change and acceleration of change of competitive share: in, For the first The rate of change of competitive share in the causal chain For the first The rate of change in the competitive share of the causal chain is accelerating. For the first incremental update The competitive share of each causal chain, and These are the first two dynamic semantic time periods, respectively. The competitive share of each causal chain, This represents the total number of dynamic semantic time periods that have been updated.

[0053] In the current dynamic semantic time period Based on the actual achievable transmission progress of each causal chain. Given decision variables, solve the constrained optimization problem: in, For the first The actual achievable propagation progress of the causal chain under current competitive conditions. For causal chain index, The total number of causal chains for the disease. In order to be with the first The index of each causal chain is different from other causal chain indices. To share intermediate indicator node indices and traverse the collection All intermediate indicator nodes are shared. For the updated shared node contention allocation matrix, the first... A causal chain with shared intermediate indicator nodes Market share of competition For the first The causal chain passes through shared intermediate indicator nodes For the The coupling effect coefficient of the causal chain, For the first incremental update The problem involves a competition for a share of causal chains. The first constraint ensures that the propagation progress of each chain is within an effective range. The second constraint reflects the upper limit of the competition capacity of each shared intermediate index node. The third constraint incorporates the coupling effect in the cross-chain coupling effect coefficient tensor into the upper bound of the achievable progress of each chain. This constraint-based optimization problem is a linear programming problem, which is solved using a linear programming algorithm.

[0054] The actual reachable propagation progress values ​​of each causal chain obtained from the constraint optimization solution. Map according to the corresponding causal chain identifier, and according to Sort from high to low to generate a dynamic priority sequence for multi-disease transformation paths. Each item in the dynamic priority sequence contains a causal chain identifier, a current dynamic semantic time period identifier, and a set of shared intermediate indicator node identifiers corresponding to that causal chain. For those satisfying... The causal chain generates drift warning tags, among which, To preset the warning threshold, and The drift warning label is combined with the corresponding causal chain identifier, the set of shared intermediate indicator node identifiers, and the dynamic semantic time period identifier to output the competitive situation drift warning assessment result.

[0055] This implementation first performs unified time unit conversion, Z-score standardization, and encoding on asynchronous heterogeneous medical data. Then, it detects the semantic segmentation boundaries within each data source using the information entropy change rate, and generates global event anchors using time proximity factors and cross-source index covariance factors. This allows asynchronous observations across departments to be mapped to a unified dynamic semantic time period. In this way, collaborative changes in different disease dimensions at similar clinical stages can be expressed within the same dynamic semantic time period, avoiding the situation where cross-disease collaborative information is scattered across different time locations when analyzing independent time axes side by side.

[0056] This implementation uses a modal adaptive encoder to map continuous numerical indicators, time-series segment indicators, and image assessment summary indicators to the same shared state embedding space. It also employs monotonic constraint interpolation and recent effective value preservation to generate virtual state node vectors for treatment intervals. Thus, the missing observation dynamic semantic time periods and the actual observed dynamic semantic time periods use a consistent data representation, ensuring the continuity of the state evolution process on the unified dynamic time axis.

[0057] This implementation further generates multi-disease causal network snapshots for each dynamic semantic time period and performs Shapley value allocation on shared intermediate indicator nodes across diseases to obtain the competitive share of each causal chain for the shared intermediate indicator nodes. Then, the blocking and accelerating effects are quantified through cross-chain coupling influence coefficient tensors. Thus, the patient conversion profile not only includes the state changes of each disease dimension itself, but also the competitive and coupling relationships of each disease causal chain at the shared intermediate indicator nodes, which can support the determination of the most urgent conversion path.

[0058] This implementation method, upon the arrival of new departmental data, performs local recalculation only on the newly added dynamic semantic time period and its associated shared intermediate indicator nodes. It then uses the cross-chain competitive situation vector, rate of change, acceleration of change, and constraint optimization solution results to jointly generate a dynamic priority sequence for multi-disease conversion paths and a competitive situation drift early warning assessment result. Therefore, after a new observation is added, the path priority can be updated synchronously with the current competitive situation, avoiding the difficulty in distinguishing the urgency of conversion when relying solely on static parallel profiles for assessment.

[0059] The following is an example of an application of the present invention, such as... Figure 2-9 As shown, the implementation process is as follows: A tertiary hospital admitted a patient with multiple comorbidities (patient number PT-0471), who also suffered from type 2 diabetes, early diabetic nephropathy, and hypertensive heart disease. The endocrinology, nephrology, and cardiology departments conducted asynchronous follow-up monitoring of the patient according to their respective treatment plans. The endocrinology department collected fasting blood glucose (FBG), glycated hemoglobin (HbA1c), and insulin resistance index (HOMA-IR) weekly; the nephrology department collected serum creatinine (SCr), urine albumin-to-creatinine ratio (UACR), and estimated glomerular filtration rate (eGFR) every two weeks; and the cardiology department collected left ventricular ejection fraction (LVEF), QTc interval, and diastolic function classification (E / A ratio) monthly. The three causal chains for the disease are: Chain 1 (progression chain of diabetic nephropathy, FBG→HOMA-IR→SCr→eGFR), Chain 2 (progression chain of cardiorenal syndrome, SCr→UACR→LVEF→QTc), and Chain 3 (hyperglycemic cardiac injury chain, HbA1c→HOMA-IR→E / A ratio→QTc). SCr, HOMA-IR, and QTc are shared intermediate indicator nodes across chains. The assessment timeframe is the 8th week of QA in 20XX (current dynamic semantic time period). ).

[0060] Asynchronous data acquisition and semantic segmentation module; For patient PT-0471, the system retrieved asynchronous medical data streams from three data sources. The endocrinology data stream (source S1) had a timestamp sequence in days, the nephrology data stream (source S2) had a 14-day interval, and the cardiology data stream (source S3) had a 28-day interval; all were uniformly converted to daily timestamps. Z-score standardization was performed on continuous indicators such as FBG, HbA1c, SCr, eGFR, UACR, LVEF, HOMA-IR, and QTc. One-hot encoding was performed on diastolic function classification (E / A ratio classification) and ECG conclusion labels.

[0061] Taking source S1 (endocrinology department) as an example, the local information entropy is calculated within the observation window of day 28. The previous time point Information entropy is The time interval is 7 days, and the rate of change of information entropy is: This value exceeds the historical value of source S1. The 85th percentile threshold of the distribution ( On day 28, the semantic segmentation boundary point of S1 was marked.

[0062] Table 1. Semantic segmentation boundary point detection results for each data source. Global event anchor point generation module; Using the boundary point of S1 on day 28 ( The boundary point between day 30 and S2 (day ) Taking (day) as an example, calculate the cross-source boundary correlation degree. Time decay coefficient. (sky The time proximity factor is: Within a local time window centered on two boundary points, the mutual information between FBG and SCr, after range normalization, is obtained as follows: The boundary correlation degree is: correlation threshold , The boundary points of day 28 of S1 and day 30 of S2 are aggregated to form the global event anchor point GA. The anchor point timestamp, calculated using the median time weighted by relevance, is approximately day 29. Day 56 of S3 independently forms the global event anchor point GB (day 56). Therefore, the overall patient observation timeline is divided into four dynamic semantic time periods: Segment 1 (days 0 to 28), Segment 2 (days 29 to 55), Segment 3 (days 56 to 83), and Segment 4 (from day 84 onwards, the current segment). ).

[0063] Table 2 Global Event Anchor Points and Dynamic Semantic Time Period Division State embedding generation module; Within dynamic semantic time period 2 (days 29 to 55), the S3 cardiology data has not yet arrived (the cardiology data collection cycle is 28 days; the first data arrived on day 28, and the second on day 56). There are no valid observations for LVEF, QTc, and E / A classification within this period, representing a missing data point during the treatment interval. The system performs monotonic constrained interpolation on LVEF and QTc in the standardized space: using the observations at the end of segment 1 (day 28) and the beginning of segment 3 (day 56) as boundaries, it generates a sequence of estimated values ​​within segment 2, with all estimated values ​​located between the valid observations on both sides. The E / A classification is used as a categorical indicator, and the most recently valid encoded value is preserved (preserving the encoded value from day 28). The interpolation and preservation results are combined into a hidden state transition estimation vector, which is then used to fill the virtual state node vector of segment 2.

[0064] For segment 4 (current segment) All three data sources had valid observations. The modal adaptive encoder's continuous numerical branch processed the standardized values ​​of FBG, HbA1c, SCr, eGFR, UACR, LVEF, HOMA-IR, and QTc; the cyclic coding branch processed the 7-day fragment sequence of blood glucose monitoring; and the convolutional coding branch processed the structured matrix of echocardiography assessment. The outputs of the three branches were then fused with attention weighting to generate... ( ).

[0065] Table 3, Section 4: Standardized values ​​and modal branching of key input indicators Causal network snapshot generation module; Within section 4, the system performs transfer entropy calculations on all cross-disease dimension indicators. Taking FBG and SCr as examples, , The difference is 0.27, which exceeds the threshold for determining the causal direction. A causal relationship was determined to exist from FBG to SCr, with an initial causal strength of 0.27. After performing range normalization on the initial causal strengths of all causal directed edges within segment 4, a snapshot of the multi-disease causal network was generated. .

[0066] Traversal All indicator nodes in the chain are as follows: SCr is located on the internal paths of both chain 1 (FBG→HOMA-IR→SCr→eGFR) and chain 2 (SCr→UACR→LVEF→QTc); HOMA-IR is located on the internal paths of both chain 1 and chain 3 (HbA1c→HOMA-IR→E / A→QTc); and QTc is located at the end of both chain 2 and chain 3. Therefore, they share the same set of intermediate indicator nodes. .

[0067] Table 4 shows the snapshot of the multi-disease causal network, including the main directed edges and normalized weights. Cross-chain competitive quantification module; For the shared intermediate index node SCr, the causal chains participating in its competition are chain 1 and chain 2. Using the normalized edge weights of each chain passing through SCr as a marginal contribution measure, a Shapley value allocation is performed: the transmission strength of chain 1 to SCr (incoming edge HOMA-IR→SCr weight is 0.67) is higher than that of chain 2 to SCr (SCr is the starting node of chain 2, outgoing edge SCr→UACR weight is 0.63), and the Shapley allocation result is that chain 1 has the highest share of competition. Chain 2 competition share The same allocation is performed for HOMA-IR (chain 1 competes with chain 3) and QTc (chain 2 competes with chain 3).

[0068] Table 5, Section 4: Shared Node Competition Allocation Matrix Using SCr as a shared node, calculate the cross-chain coupling influence coefficient of chain 1 on chain 2 via SCr. , (HOMA-IR→SCr normalized weights), there are no convergence nodes downstream of SCr for chains 1 and 2 (eGFR and UACR belong to two separate chains and do not converge), therefore... ,therefore: A negative value indicates that chain 1 inhibits chain 2 at the SCr node. The higher the share of competition for SCr by chain 1, the less SCr transmission resources chain 2 can obtain.

[0069] Transformation profile generation module; Based on the time sequence of global event anchor points GA (day 29) and GB (day 56), the state node vectors of segments 1 to 4 are... to Multi-disease causal network snapshot to and competitive allocation matrix to An ordered connection is performed to generate a time-series directed graph of the transformation profile of patient PT-0471. The virtual state node vector of segment 2 (filled by implicit state transition estimates) is connected to the actual state node vectors of segments 1 and 3 through directed edges. The edge attributes record the differences in the state node vectors (e.g., ...). The L2 norm is 0.34) and the snapshot structure changes (segment 2 adds a HOMA-IR→SCr causal edge).

[0070] Incremental update module; On day 84, the Nephrology Department received new observational data with SCr = 141 μmol / L, eGFR = 51 mL / min, and UACR = 203 mg / g, triggering an incremental update. The system performed local semantic segmentation only on the S2 data source and detected the information entropy change rate on day 84. NAT / day, exceeding S2 Quantile thresholds are used to generate new segmentation boundary points. The correlation between this boundary point and GB (day 56) is then determined. This triggers a new global event anchor GC (day 84), officially establishing the starting boundary of segment 4 as day 84 (the current dynamic semantic time period). ).

[0071] Perform modal adaptive coding on segment 4 to obtain And locally recalculate the Shapley value assignment and cross-chain coupling influence coefficient tensor. Updated shared node contention allocation matrix As shown in Table 5, taking Chain 1 as an example, the overall competitive advantage is calculated. : Path weights of each shared node :SCr HOMA-IR QTc is not on path 1, therefore .

[0072] Similarly, calculate , Generate cross-chain competitive situation vector .

[0073] Dynamic priority evaluation and result mapping module; Historical competitive situation vector is , ( ), calculate the rate of change and acceleration of each chain: Chain 1: , Chain 3: , Chain 2: , .

[0074] Based on the actual achievable transmission progress of each chain Given SCr as the decision variables, solve a linear programming problem. Constraints include: SCr node capacity constraints. HOMA-IR node capacity constraints QTc node capacity constraints ; Incorporation of coupling upper bound constraints Inhibition terms (such as the inhibition of chain 1 on chain 2) The solution result is: , , .

[0075] Table 6. Dynamic Priority Assessment Results of Segment 4 Multiple Disease Transformation Pathways Throughout the implementation process, the raw asynchronous heterogeneous medical data, after standardization and encoding preprocessing, firstly detects semantic segmentation boundaries within each data source using the information entropy change rate. Then, using temporal proximity factors and cross-source covariance factors, the boundary points of day 28 of S1 and day 30 of S2 are aggregated into a global event anchor point (GA), connecting the four dynamic semantic time periods into a unified time axis. Heterogeneous indicators from various disease dimensions within segment 4 are fused into a state node vector through modal adaptive encoding. Transition entropy analysis generates snapshots of causal networks for multiple diseases. Shared intermediate indicator nodes The competitive allocation matrix is ​​obtained through Shapley allocation. Cross-chain coupling influence coefficient tensor The inter-chain inhibition relationship was quantified (e.g.) The incremental update module only locally recalculates segment 4 when new nephrology data arrives, generating a competition situation vector. The final constraint optimization solution output chain 1 (diabetic nephropathy progression chain) can actually achieve the highest conduction progress ( This is identified as the most urgent transformation path, which is logically consistent with the numerical relationship of the competitive situation vector and the cross-chain inhibition structure.

[0076] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A patient transformational portrait construction and evaluation system oriented to dynamic time axis feature extraction, characterized in that, include: The asynchronous data acquisition and semantic segmentation module is configured to acquire asynchronous medical data streams generated by various departments for patients with multiple coexisting diseases, detect semantic segmentation boundary points for each data source based on the local information entropy change rate, and generate a set of local segmentation results for each source. The global event anchor point generation module is configured to calculate the cross-source boundary correlation matrix based on the local segmentation result set of each source, merge the cross-source boundary points that meet the correlation threshold into global event anchor points, and define a unified dynamic semantic time period across sources. The state embedding generation module is configured to perform modality adaptive coding on heterogeneous indicators of each disease dimension within each dynamic semantic time period to generate modality-independent state node vectors. The causal network snapshot generation module is configured to perform causal direction determination based on transfer entropy for cross-disease dimension indicator pairs within each dynamic semantic time period, generate multi-disease causal network snapshots, and identify cross-disease shared intermediate indicator nodes. The cross-chain competition quantification module is configured to use the Shapley value allocation method to calculate the competitive allocation share of each disease causal chain for shared intermediate indicator nodes, and generate a shared node competition allocation matrix and a cross-chain coupling influence coefficient tensor. The patient conversion profile generation module is configured to connect the state node vectors of each dynamic semantic time period, the snapshot of the multi-disease causal network, and the shared node competition allocation matrix according to the global event anchor sequence to generate a patient conversion profile. The incremental update module is configured to locally recalculate the competition allocation for newly added dynamic semantic time periods when new department data arrives, generating a cross-chain competition situation vector. The dynamic priority assessment and result mapping module is configured to perform constraint optimization based on the cross-chain competitive situation vector and the cross-chain coupling influence coefficient tensor, and output the dynamic priority ranking of multiple disease transformation paths and the competitive situation drift early warning assessment results.

2. The patient conversion profile construction and evaluation system based on dynamic timeline feature extraction according to claim 1, characterized in that, The asynchronous data acquisition and semantic segmentation module is configured to acquire data from a data source with an ordered timestamp sequence , and calculate local information entropy of index value distribution in a local observation window centered on a time point , and the information entropy change rate between adjacent observation time points is ​ Wherein, is a time point The rate of change of information entropy relative to the previous observation time point; when is greater than the preset quantile value of the data source history distribution, mark as the semantic segmentation boundary point of the data source, the quantile parameter corresponding to the preset quantile is denoted as , and .

3. The patient conversion profile construction and evaluation system based on dynamic timeline feature extraction according to claim 1, characterized in that, In the global event anchor point generation module, the cross-source boundary correlation matrix elements By time proximity factor Covariance factors of cross-source indicators Sure: in, For data source No. The timestamps of the boundary points For data source No. The timestamps of the boundary points The time decay coefficient and , The mutual information value between available indicators from two data sources within a local time window centered at two boundary points, after range normalization. ; for satisfying Perform a merge operation on cross-source boundary point pairs. The correlation threshold and The cross-source boundary points that meet the conditions are aggregated into global event anchor points. The anchor point timestamp is the weighted median time obtained by weighting the timestamps of each boundary point in the aggregation point set according to the correlation.

4. The patient conversion profile construction and evaluation system based on dynamic timeline feature extraction according to claim 1, characterized in that, In the state embedding generation module, the modal adaptive encoder consists of a continuous numerical index coding branch, a time-series segment index coding branch, an image evaluation summary index coding branch, and an attention-weighted fusion layer; the loss function of the modal adaptive encoder is: in, For modal branch index, The total number of modal branches, For the first The reconstruction error weights of each modal branch and , For the first Reconstruction error of each modal branch For cross-modal consistency constraint weights and , Pick Subtract the mean of the pairwise cosine similarity of the intermediate representation vectors of different modal branches within the same dynamic semantic time period.

5. The patient conversion profile construction and evaluation system based on dynamic timeline feature extraction according to claim 1, characterized in that, The state embedding generation module is also configured to perform implicit state transition estimation for dynamic semantic time periods lacking observation data during treatment intervals. Based on the observation values ​​of each available index at the two most recent global event anchor points before and after the treatment interval, continuous indices are estimated using a monotonic constraint interpolation algorithm in the standardized space, and categorical indices are estimated using a recent effective encoding value preservation strategy. The estimation results are combined to generate virtual state node vectors to fill the corresponding dynamic semantic time period positions.

6. The patient conversion profile construction and evaluation system based on dynamic timeline feature extraction according to claim 1, characterized in that, In the causal network snapshot generation module, the indicators and indicators Calculate the transfer entropy separately and ,when Time determination to There is a causal relationship between directions. Threshold for determining causal direction and ; Perform range normalization on all original causal strengths within the same dynamic semantic time period to obtain edge weights; Generate multi-disease causal network snapshots ,in For the set of indicator nodes, For a set of causal directed edges, This is the normalized set of edge weights; indicator nodes that are simultaneously located on the internal paths of two or more different disease causal chains are identified as cross-disease shared intermediate indicator nodes.

7. The patient conversion profile construction and evaluation system based on dynamic timeline feature extraction according to claim 1, characterized in that, In the cross-chain competition quantization module, the cross-chain coupling influence coefficient tensor The element is defined as: in, and Indexing causal chains for different diseases For the first A causal chain with shared intermediate indicator nodes Market share of competition For the first The causal chain enters the shared intermediate indicator node. The weight of the incoming edge at the end, To share intermediate indicator nodes Along the first The weight of the outgoing edge from the direction of the causal chain to the nearest converging node. As an indicator function, when the sink node is simultaneously located at the _ Article and Section When the value is taken on the downstream path of the causal chain, it is... Otherwise, the value is ;when season .

8. The patient conversion profile construction and evaluation system based on dynamic timeline feature extraction according to claim 1, characterized in that, In the incremental update module, the cross-chain competition situation vector The The components are: in, This is the set of shared intermediate indicator nodes for the current dynamic semantic time period. For the first A causal chain with shared intermediate indicator nodes Market share of competition To share intermediate indicator nodes In the Path weights on the causal chain transmission path and The node is at the 1st The weights of incoming and outgoing edges on a causal chain.

9. The patient conversion profile construction and evaluation system based on dynamic timeline feature extraction according to claim 1, characterized in that, In the dynamic priority evaluation and result mapping module, the actual achievable propagation progress of each causal chain is used as the basis. Given the decision variables, solve the linear programming problem: in, For the first The first constraint represents the actual reachability of the causal chain; the second constraint reflects the upper limit of the competitive capacity of each shared intermediate index node; and the third constraint incorporates the cross-chain coupling effect into the upper bound of the reachability of each chain. The solution obtained... Generate a dynamic priority sequence of multi-disease transformation pathways by sorting from high to low; accelerate the change in competitive share. The causal chain generates drift warning tags. To preset the warning threshold and .

10. A method for constructing and evaluating patient conversion profiles based on dynamic timeline feature extraction, used to execute the patient conversion profile construction and evaluation system based on dynamic timeline feature extraction as described in any one of claims 1 to 9, characterized in that, include: Step 1: Obtain asynchronous medical data streams generated by various departments for patients with multiple coexisting diseases; detect semantic segmentation boundary points for each data source based on the local information entropy change rate; and generate a set of local segmentation results for each source. Step 2: Calculate the cross-source boundary correlation matrix based on the local segmentation results set of each source, merge the cross-source boundary points that meet the correlation threshold into global event anchor points, and delineate a unified dynamic semantic time period across sources; Step 3: Perform modality adaptive coding on the heterogeneous indicators of each disease dimension within each dynamic semantic time period to generate modality-independent state node vectors; Step 4: Perform causal direction determination based on transfer entropy for cross-disease dimension indicator pairs within each dynamic semantic time period, generate multi-disease causal network snapshots and identify cross-disease shared intermediate indicator nodes; Step 5: Use the Shapley value allocation method to calculate the competitive allocation share of each disease causal chain for the shared intermediate index nodes, and generate the shared node competitive allocation matrix and the cross-chain coupling influence coefficient tensor. Step 6: Connect the state node vectors of each dynamic semantic time period, the multi-disease causal network snapshots, and the shared node competition allocation matrix according to the global event anchor sequence to generate a patient conversion profile; Step 7: When the data from the new department arrives, locally recalculate the competition allocation for the newly added dynamic semantic time period to generate a cross-chain competition situation vector; Step 8: Perform constraint optimization based on the cross-chain competitive situation vector and the cross-chain coupling influence coefficient tensor, and output the dynamic priority ranking of multi-disease transformation paths and the early warning assessment results of competitive situation drift.

Images