Gastrointestinal tumor patient postoperative weakness risk correlation analysis method and system fused with co-disease map

Abstract

The invention discloses a gastrointestinal tumor patient postoperative weakness risk correlation analysis method and system fused with a co-disease map, and belongs to the field of medical care informatics, and the method comprises the following steps: obtaining clinical pathological characteristics, co-disease assessment, serum biomarkers and social and psychological assessment scale data of gastrointestinal tumor patients; obtaining statistical effect values of homocysteine and health attainment, and executing isomorphic mapping; identifying a chained intermediary conduction path and generating a coupling strength factor, and performing weighted correction on the initialized weight matrix; inputting the multi-source data as a risk source excitation signal into the atlas, executing correction calculation, and outputting a risk prediction score; a maximum weight subgraph search is performed to generate risk attribution analysis data. According to the method, the multi-dimensional heterogeneous association map is constructed to fuse the multi-source data of the patient and the medical priori knowledge, and the risk conduction is quantified through the path topology analysis and the matrix correction operation, so that the accurate prediction of the postoperative weakness risk and the interpretable attribution of the key pathogenic path can be realized.

Description

Technical Field

[0001] This invention relates to the field of healthcare informatics, and in particular to a method and system for analyzing the association between postoperative debilitating risk and comorbidity profiles in patients with gastrointestinal tumors. Background Technology

[0002] Gastrointestinal tumors are common malignant tumors in clinical practice, and surgical resection is the main treatment. However, most patients with gastrointestinal tumors are elderly and often have multiple underlying diseases, making them prone to decreased physiological reserves and weakened stress resistance after surgery, leading to postoperative weakness. Postoperative weakness not only increases the risk of complications but also seriously affects the patient's prognosis and quality of life. Therefore, accurately assessing the postoperative health status and risks of patients with gastrointestinal tumors and analyzing the correlation between their pathogenic factors is of great significance for developing personalized rehabilitation strategies.

[0003] Among related technologies, Chinese invention patent application CN119581002A discloses an intelligent postoperative gastrointestinal obstruction diagnosis and intervention decision-making system and method. This system includes a data preprocessing module, a multimodal data fusion module, a gastrointestinal function recovery model construction module, an adaptive progressive reinforcement learning (AHRL) optimization module, and a diagnosis and intervention recommendation module. The AHRL optimization module uses a five-layer progressive algorithm to optimize the intervention strategy layer by layer, from dynamic adjustment of feature weights, recursive state propagation, hierarchical cumulative optimization, symptom similarity matrix generation, to progressive weight back feedback. This achieves deep fusion of multimodal data on key symptoms and updates the intervention plan in real time according to changes in the patient's condition.

[0004] While the aforementioned technologies can dynamically predict and intervene in specific postoperative complications using multimodal data and deep learning models, these methods primarily rely on black-box computation of neural networks, focusing on the predictive accuracy of numerical results. They struggle to intuitively reveal the complex pathological mechanisms and causal transmission pathways between clinical pathology, biomarkers, and psychosocial factors. Furthermore, data fusion through simple feature vector concatenation ignores the potential topological relationships between comorbidities and different risk factors, and fails to fully utilize existing evidence-based medicine statistical evidence as prior knowledge to guide risk assessment. This results in weak model interpretability, making it difficult to meet clinical needs for attribution analysis of pathogenic pathways. Summary of the Invention

[0005] To address the aforementioned issues, this invention provides a method and system for analyzing the risk of postoperative weakness in gastrointestinal tumor patients by integrating comorbidity maps. The method employs the construction of a multidimensional heterogeneous association map that integrates multi-source patient data with prior medical knowledge. Through path topology analysis and matrix correction operations, the risk transmission is quantified, enabling accurate prediction of postoperative weakness risk and interpretable attribution of key pathogenic pathways.

[0006] The above objectives can be achieved through the following approach: A method for postoperative frailty risk association analysis in gastrointestinal tumor patients integrating a comorbidity map includes: acquiring clinicopathological feature data, comorbidity assessment data, serum biomarker data, and psychosocial assessment scale data for gastrointestinal tumor patients; performing normalization processing on the psychosocial assessment scale data to construct a multidimensional heterogeneous association map, which includes patient nodes, comorbidity nodes, biomarker nodes, and psychosocial feature nodes; acquiring medical historical literature data for feature extraction to construct a prior knowledge base; obtaining the standardized mean difference between homocysteine ​​and frailty, and the ratio of health literacy to frailty through the prior knowledge base; performing statistical effect value isomorphic mapping on the multidimensional heterogeneous association map, mapping the standardized mean difference to the connection weights of the biomarker nodes, and mapping the ratio to the connection weights of the psychosocial feature nodes, generating an initial weight matrix; The process involves performing path topology analysis on the multidimensional heterogeneous association graph to identify chain-like mediating pathways. A coupling strength factor is generated based on these pathways, and the initial weight matrix is ​​weighted and corrected using this factor to generate a target adjacency matrix. A health literacy score is extracted from the sociopsychological assessment scale data to generate a global moderating coefficient. The clinical pathological feature data, comorbidity assessment data, and serum biomarker data are used as risk source activation signals input into the multidimensional heterogeneous association graph. The global moderating coefficient is used to correct the target adjacency matrix, and a postoperative frailty risk prediction score is output for the target patient. If the postoperative frailty risk prediction score exceeds a preset safety threshold, a maximum weight subgraph search is performed in the target adjacency matrix to identify key pathogenic pathways and extract comorbidity node combinations and mediating feature nodes, generating risk attribution analysis data.

[0007] Optionally, the construction of the multidimensional heterogeneous association graph includes: acquiring electronic medical records and reading pathology examination records and laboratory reports; generating structured clinical pathological feature data and serum biomarker data; and reading past medical history records to generate comorbidity assessment data; performing missing item removal and score summarization on the sociopsychological assessment scale data; performing interval scaling on the maximum and minimum values ​​to generate sociopsychological assessment scale values; instantiating and generating patient nodes, comorbidity nodes, biomarker nodes, and sociopsychological feature nodes; and establishing a connection edge data table for the patient nodes pointing to the comorbidity nodes, the biomarker nodes, and the sociopsychological feature nodes to construct the multidimensional heterogeneous association graph.

[0008] Optionally, obtaining the standardized mean difference between homocysteine ​​and frailty, and the ratio of health literacy to frailty through the prior knowledge base includes: using homocysteine ​​and health literacy as exposure factors and frailty as an outcome variable, constructing a literature retrieval instruction, identifying and extracting mean data, standard deviation data, and risk-related data from the medical historical literature data; performing a unified conversion on the mean data and standard deviation data using concentration units, and performing alignment correction on the risk-related data in the direction of effect size, constructing a structured feature dataset; performing statistical heterogeneity testing and pooled effect size calculation on the structured feature dataset, aggregating and generating the standardized mean difference and ratio, and outputting them to the query interface of the prior knowledge base.

[0009] Optionally, generating the initial weight matrix includes: establishing an empty adjacency matrix with all zero values ​​based on the multidimensional heterogeneous association graph; resolving the row and column intersection positions of the patient node, the biomarker node, and the sociopsychological feature node in the empty adjacency matrix; writing the standardized mean difference into the connection weight storage unit at the row and column intersection position corresponding to the biomarker node; writing the ratio into the connection weight storage unit at the row and column intersection position corresponding to the sociopsychological feature node; and outputting the initial weight matrix in memory.

[0010] Optionally, generating the target adjacency matrix includes: traversing the topology through the multidimensional heterogeneous association graph; based on the sociopsychological feature nodes, locking the sleep disorder node to the depressive mood node and then to the patient node as a directed connection sequence, and marking it as a chain-mediated transmission path; reading the statistical regression coefficients of the chain-mediated transmission path from the prior knowledge base, performing continuous multiplication, calculating and instantiating them as coupling strength factors; based on the initial weight matrix, locating the direct connection elements for the sleep disorder node and the patient weakness node, accumulating the coupling strength factors to the direct connection elements, and outputting the target adjacency matrix.

[0011] Optionally, the calculation and instantiation into a coupling strength factor includes: retrieving statistical values ​​from the prior knowledge base using the node connection relationship of the chain-mediated transmission path as the index key; performing floating-point multiplication on the statistical values; and assigning the result to the coupling strength factor.

[0012] Optionally, the output of the postoperative frailty risk prediction score includes: extracting a health literacy score based on the sociopsychological assessment scale data, converting the health literacy score into a global adjustment coefficient using a negative exponential function; vectorizing and concatenating the clinicopathological feature data, the comorbidity assessment data, and the serum biomarker data to construct a risk source activation signal vector, and mapping it to the multidimensional heterogeneous association map; performing a scalar product operation on the target adjacency matrix using the global adjustment coefficient to complete the correction, and performing matrix multiplication with the risk source activation signal vector, and outputting the postoperative frailty risk prediction score via an activation function mapping.

[0013] Optionally, constructing the risk source activation signal vector includes: extracting effective numerical components based on the clinicopathological feature data, the comorbidity assessment data, and the serum biomarker data; and performing a first-to-last serial splicing of the effective numerical components to generate the risk source activation signal vector.

[0014] Optionally, the generation of risk attribution analysis data includes: performing a maximum weight subgraph search based on the target adjacency matrix, using the patient node as the backtracking starting point, and locking the node connection sequence with the highest cumulative weight value; marking the node connection sequence as a key pathogenic path and identifying comorbid node combinations and mediating feature nodes; extracting cumulative weight values ​​based on the key pathogenic path and performing structured encapsulation with the comorbid node combinations and the mediating feature nodes to generate risk attribution analysis data.

[0015] Based on the same inventive concept, this invention also provides a system for analyzing the risk of postoperative frailty in gastrointestinal tumor patients by integrating a comorbidity map. The system includes: a map construction module for acquiring clinicopathological feature data, comorbidity assessment data, serum biomarker data, and psychosocial assessment scale data for gastrointestinal tumor patients; performing normalization processing on the psychosocial assessment scale data to construct a multidimensional heterogeneous association map, which includes patient nodes, comorbidity nodes, biomarker nodes, and psychosocial feature nodes; a priori knowledge base module for acquiring medical historical literature data and extracting features to construct a priori knowledge base; using the priori knowledge base to obtain the standardized mean difference between homocysteine ​​and frailty, and the ratio of health literacy to frailty; and a weight mapping module for performing statistical effect value isomorphic mapping on the multidimensional heterogeneous association map, mapping the standardized mean difference to the connection weights of the biomarker nodes, and mapping the ratios to the connection weights of the psychosocial feature nodes. The system generates an initial weight matrix; an adjacency generation module performs path topology analysis on the multidimensional heterogeneous association graph, identifies chain-like mediating transmission paths, generates coupling strength factors based on these paths, and uses these coupling strength factors to perform weighted correction on the initial weight matrix to generate a target adjacency matrix; a scoring calculation module extracts health literacy scores from the sociopsychological assessment scale data to generate a global moderating coefficient, inputs the clinical pathological feature data, comorbidity assessment data, and serum biomarker data as risk source activation signals into the multidimensional heterogeneous association graph, performs correction calculations on the target adjacency matrix using the global moderating coefficient, and outputs a postoperative frailty risk prediction score for the target patient; and an attribution analysis module performs a maximum weight subgraph search in the target adjacency matrix if the postoperative frailty risk prediction score exceeds a preset safety threshold, identifies key pathogenic paths, extracts comorbidity node combinations and mediating feature nodes, and generates risk attribution analysis data.

[0016] Compared with the prior art, the present invention has the following advantages: 1. It enables comprehensive integration and holistic assessment of risk factors. By constructing a multidimensional heterogeneous correlation map encompassing clinicopathology, comorbidities, biomarkers, and psychosocial characteristics, it breaks down the data silos of independent analysis of risk factors in traditional models. It treats the patient's health status as a complex, interconnected network, thereby revealing the comprehensive sources of postoperative frailty risk more comprehensively and deeply, and improving the overall holistic nature of risk assessment.

[0017] 2. Enhanced the accuracy and scientific rigor of risk prediction. This method not only relies on individual patient data but also incorporates a priori knowledge base extracted from a vast amount of medical literature, quantifying and integrating population-level statistical evidence such as standardized mean difference and odds ratio into the model weights.

[0018] 3. Provides highly interpretable risk attribution and decision support. Unlike traditional "black box" prediction models, after a high-risk warning, it can backtrack and pinpoint key pathogenic pathways in complex interconnected networks through a maximum weight subgraph search algorithm.

[0019] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures pointed out in the description, claims and drawings. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a flowchart illustrating the method for analyzing the risk of postoperative debilitation in gastrointestinal tumor patients by fusing comorbidity maps according to an embodiment of the present invention.

[0022] Figure 2 This is a comparison chart of the risk transmission path weights before and after correction in an embodiment of the present invention.

[0023] Figure 3 This is an attribution decomposition diagram of the postoperative frailty risk score according to an embodiment of the present invention.

[0024] Figure 4 This is a schematic diagram of the structure of the gastrointestinal tumor patient postoperative debilitation risk association analysis system based on the fusion of comorbidity maps according to an embodiment of the present invention. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] Reference Figure 1One embodiment of the present invention proposes a method for postoperative weakness risk association analysis in gastrointestinal tumor patients by integrating comorbidity maps. The method involves constructing a multidimensional heterogeneous association map that integrates multi-source patient data and medical prior knowledge. By quantifying risk transmission through path topology analysis and matrix correction operations, it is possible to achieve accurate prediction of postoperative weakness risk and interpretable attribution of key pathogenic pathways.

[0027] The method described in this embodiment specifically includes: For patients with gastrointestinal tumors, clinical pathological feature data, comorbidity assessment data, serum biomarker data, and psychosocial assessment scale data are obtained. The psychosocial assessment scale data is normalized and a multidimensional heterogeneous association map is constructed. The multidimensional heterogeneous association map includes patient nodes, comorbidity nodes, biomarker nodes, and psychosocial feature nodes. Medical historical literature data is acquired and features are extracted to construct a prior knowledge base. The prior knowledge base is used to obtain the standardized mean difference between homocysteine ​​and frailty, as well as the ratio of health literacy to frailty. Perform statistical effect value isomorphism mapping on the multidimensional heterogeneous association map, map the standardized mean difference to the connection weight of the biomarker node, map the ratio to the connection weight of the sociopsychological feature node, and generate an initial weight matrix; For the multidimensional heterogeneous association graph, perform path topology parsing to identify chain-like mediation transmission paths, generate coupling strength factors based on the chain-like mediation transmission paths, and use the coupling strength factors to perform weighted correction on the initial weight matrix to generate the target adjacency matrix. Based on the social and psychological assessment scale data, health literacy scores are extracted to generate a global regulation coefficient. The clinical pathological feature data, the comorbidity assessment data, and the serum biomarker data are used as risk source activation signals and input into the multidimensional heterogeneous association map. The global regulation coefficient is used to perform correction calculations on the target adjacency matrix, and a postoperative frailty risk prediction score is output for the target patient. If the postoperative weakness risk prediction score exceeds the preset safety threshold, a maximum weight subgraph search is performed in the target adjacency matrix to lock the key pathogenic path and extract the combination of comorbid nodes and mediating feature nodes to generate risk attribution analysis data.

[0028] Optionally, the construction of the multidimensional heterogeneous association map includes: It acquires electronic medical records and reads pathology examination records and laboratory reports, generates structured clinical pathological feature data and serum biomarker data, and reads past medical history records to generate comorbidity assessment data; By accessing the patient's electronic medical record database through the standard data interface of the Hospital Information System (HIS), for unstructured pathology examination records, such as "gastric antral adenocarcinoma, moderately differentiated, infiltrating to the serosal layer," named entity recognition (NER) technology in natural language processing is used to extract tumor location and differentiation degree, converting it into structured clinicopathological feature data. For structured laboratory report records, biochemical indicator values ​​within a specific time window are directly extracted using SQL queries, particularly homocysteine ​​(Hcy) concentration values, as serum biomarker data. For past medical history records, the patient's long-term medical orders and discharge diagnosis lists are reviewed, and underlying diseases such as hypertension, diabetes, and coronary heart disease are matched according to the International Classification of Diseases (ICD) codes. Cumulative scores are calculated based on the Charlesson Comorbidity Index (CCI) scoring rules to generate comorbidity assessment data.

[0029] The missing items of the social psychological assessment scale data are removed and the scores are summarized. The maximum and minimum values ​​are scaled to generate the social psychological assessment scale values. When collecting data from patients' psychosocial assessment scales, data quality control was first implemented. If the number of missing items on a scale exceeded a certain proportion of the total number of items, the sample was removed; if the missing items did not exceed the standard, the missing values ​​were imputed using the sample mean. Subsequently, the scores for each item on the scale were summed to obtain the raw total score. To eliminate differences in the units of measurement between different scales and ensure comparability of all feature data in graph calculations, the raw total score was normalized using a maximum-minimum range scaling method, calculated as follows: , in, The value represents the normalized sociopsychological assessment scale value, which maps the original score to the dimensionless interval [0,1]. The closer the value is to 1, the stronger the characteristic. This indicates the patient's raw total score on a specific scale, derived from the direct summation of questionnaire items. This represents the theoretical minimum possible score for the scale, determined by the scale's design specifications. For example, if the scale has 10 items, and each item has a minimum score of 1 point, then... . This represents the theoretically maximum possible score for the scale, which is also determined by the scale design specifications. For example, if the scale has 10 items, and each item has a maximum score of 5, then... . It represents the range of values ​​on the scale and is used to measure the dispersion of data.

[0030] For example, suppose we use the "Health Literacy Scale for Patients with Chronic Diseases" to assess a patient with gastrointestinal cancer. This scale contains 12 items, each scored out of 5. Therefore, according to the scale definition, its theoretical minimum score is... Score, theoretical maximum score Points. If a patient Zhang San's original total score... The calculation process for performing interval scaling on the fraction is as follows: Finally, the value 0.75 was stored in the database as Zhang San's attribute value at the socio-psychological characteristic node of "health literacy".

[0031] Patient nodes, comorbidity nodes, biomarker nodes, and sociopsychological feature nodes are instantiated and generated. A connection edge data table is established for the patient nodes pointing to the comorbidity nodes, the biomarker nodes, and the sociopsychological feature nodes to construct a multidimensional heterogeneous association graph.

[0032] First, the schema of the graph is defined, declaring four types of node labels: Patient, Comorbidity, Biomarker, and Psychosocial. For each patient, a Patient node is instantiated, with attributes including patient ID and age. For each extracted comorbidity, an existing Comorbidity node is instantiated or matched. For each extracted biomarker, a Biomarker node is instantiated. For each processed scale value, a Psychosocial node is instantiated. Next, a connection edge data table is constructed, containing "source node ID," "target node ID," and "edge type" fields. Based on the attribution relationships between data points, directed edges are created from the Patient node to the other three types of nodes, logically connecting isolated data points into a multidimensional heterogeneous relational graph.

[0033] For example, for patient Zhang San, he is identified as having "type 2 diabetes", abnormal serum homocysteine ​​concentration, and a health literacy score of 0.75. The graph construction engine will perform the following operations: create a node (:Patient{id:“P001”,name:“Zhang San”}); match or create a node (:Comorbidity{name:“Type 2 diabetes”}); match or create a node (:Biomarker{name:“Homocysteine”}); match or create a node (:Psychosocial{name:“Health literacy”}). Establish an edge (P001)-[:HAS_COMORBIDITY]->(Type 2 diabetes); establish an edge (P001)-[:HAS_BIOMARKER{value:“18μmol / L”}]->(Homocysteine); establish an edge (P001)-[:HAS_PSYCHOSOCIAL{normalized_score:0.75}]->(Health literacy). All of Zhang San's multimodal data were fused into a local map structure centered on himself and radiating outwards to connect risk factors.

[0034] Optionally, obtaining the standardized mean difference between homocysteine ​​and frailty, and the ratio of health literacy to frailty, through the prior knowledge base includes: Using homocysteine ​​and health literacy as exposure factors and frailty as an outcome variable, a literature retrieval instruction was constructed to identify and extract mean data, standard deviation data, and risk association data from the medical historical literature data. Logical combinations of search queries are constructed based on medical subject heading lists, such as "(homocysteine ​​OR Hcy) AND (weakness OR Frailty)". Historical medical literature is traversed, including local mirrors or API interfaces of databases such as PubMed, Web of Science, and Wanfang Medical. An entity relation extraction model based on BERT architecture is used to locate statistical tables describing "case group" and "control group" in the literature. The following key values ​​are identified and extracted: for continuous variables, sample size, mean, and standard deviation for both groups; for binary or rating variables, the number of events or risk association data for both groups.

[0035] The mean data and the standard deviation data are converted to a unified unit of concentration measurement, and the risk-related data are aligned and corrected in the direction of effect size to construct a structured feature dataset. Because different publication dates or regions vary, the units of measurement for homocysteine ​​concentration may differ. A built-in unit conversion rule library identifies the original units of the extracted data. If the original unit is... Automatically multiply it by a conversion factor to convert it. This achieves standardized conversion of concentration units. For risk-related data, alignment correction of effect size direction is required. This is because higher scores on some scales represent better health, while higher scores on others represent greater risk. Based on variable attributes, all effect values ​​reported in the literature are uniformly converted to "risk intensity relative to the occurrence of frailty." If a literature reports "for every 1 point increase in health literacy, the risk of frailty decreases," then its original state of OR value being less than 1 is maintained; if the reporting logic is reversed, then the reciprocal operation is performed on the OR value to unify the direction.

[0036] For example, suppose the mean homocysteine ​​data extracted from a document is as follows: The unit was identified as being related to... If there is a discrepancy, call the conversion algorithm: The converted value of 14.78 is stored in the "homocysteine ​​concentration" field of the structured feature dataset to ensure dimensional consistency during statistical merging.

[0037] Perform statistical heterogeneity tests and pooled effect size calculations on the structured feature dataset, respectively aggregate and generate standardized mean difference and odds ratio, and output them to the query interface of the prior knowledge base.

[0038] First, calculate the heterogeneity statistic. .like This indicates that the differences in research results among the various studies are small, and a fixed-effects model is automatically selected; if This indicates significant heterogeneity, and a random-effects model is automatically selected. Subsequently, the data are weighted and pooled using the selected model. For homocysteine, due to potential slight differences in the measurement tools used in different studies, the standardized mean difference (SMD) is used as the final pooled effect size. SMD eliminates the influence of dimensions and reflects the degree of overlap between the distributions of the two groups of data. The formula for calculating the SMD for a single study is as follows: , in, The standardized mean difference represents how many times the difference in homocysteine ​​levels between the weakened and non-weak groups is greater than the pooled standard deviation. A larger value indicates a stronger ability of the biomarker to distinguish between weakened and non-weak groups. This represents the extracted mean data of the weakened group. This represents the mean data of the extracted non-weak groups. The sample size of the weak group is shown. This represents the sample size of the non-weak group. This represents the standard deviation data for the weak group. This represents the standard deviation data for the non-weak group. The denominator part... This represents the pooled standard deviation, used to measure the overall dispersion of two sets of data. After calculating the SMD for each article, the inverse variance method is used to calculate a weighted average of all SMDs to obtain the final aggregated standardized mean difference. Similarly, for risk association data of health literacy, the log-odds ratio (logOR) of each article is weighted and pooled, and then the exponent is taken to obtain the final odds ratio.

[0039] For example, suppose we need to calculate the SMD of homocysteine ​​based on data from a certain literature review, and the extracted data is as follows. Debilitated group: Mean Standard deviation Sample size Non-weak group: mean Standard deviation Sample size Substitute the values ​​into the formula to calculate the square of the combined standard deviation: Therefore, the combined standard deviation is Calculate SMD: The calculation results indicate that, in this study, the homocysteine ​​levels in debilitated patients were one standard deviation higher than those in non-debilitated patients. This 1.0 and its weights are used in a weighted merging process, and the resulting aggregated value will be output to the prior knowledge base as the weighting benchmark for graph initialization.

[0040] Optionally, generating the initial weight matrix includes: Based on the multidimensional heterogeneous association graph, an empty adjacency matrix with all zero values ​​is established, and the row and column intersection positions of the patient node, the biomarker node, and the sociopsychological feature node in the empty adjacency matrix are analyzed. First, iterate through all node objects in the multidimensional heterogeneous association graph to obtain the total number of nodes. Subsequently, a contiguous block of memory is allocated, and a dimension of is initialized. A two-dimensional array, i.e., an empty adjacency matrix, is used, with all elements preset to the floating-point number 0.0. To determine the correspondence between graph nodes and matrix rows and columns, a node index mapping table is constructed. This mapping table uses a hash table structure, with the node's unique identifier as the key and values ​​ranging from 0 to 1. The integer sequence is the value. The index of the patient node can be quickly parsed by querying this mapping table. and the index of biomarker nodes. Index of socio-psychological characteristic nodes .thus, and This refers to the intersection of the row and column positions of a node in the matrix.

[0041] For example, assuming the constructed multidimensional heterogeneous association graph contains 10 nodes, establish a The matrix consists of all zeros. The patient node "Zhang San" has the ID "Patient_01" and is assigned an index of 0 in the mapping table; the homocysteine ​​node has the ID "Bio_Hcy" and is assigned an index of 3 in the mapping table; the health literacy node has the ID "Psy_HealthLit" and is assigned an index of 7 in the mapping table. At this point, the row and column intersection of the patient and homocysteine ​​nodes is (0,3). Assuming that the rows represent the source nodes and the columns represent the target nodes, the row and column intersection of the patient and health literacy nodes is (0,7).

[0042] At the row-column intersection position corresponding to the biomarker node, the standardized mean difference is written into the connection weight storage unit; The standardized mean difference (SMD) between homocysteine ​​and frailty is retrieved from the prior knowledge base. Since the SMD reflects the fold difference in biomarker levels between the two groups, this value directly characterizes the contribution of the biomarker to the risk of frailty. The row and column intersection positions corresponding to the patient node and the biomarker node in the empty adjacency matrix are located, and the element value at this position is updated to the absolute value of the SMD. The matrix assignment logic can be expressed by the following formula: , in, Indicates the first element in the initial weight matrix. Line 1 The column's connection weight storage unit. This is the row index of the patient node in the matrix. This is the column index of the biomarker node in the matrix. This represents the absolute value of the standardized mean difference between the biomarker and the association with frailty, obtained from a prior knowledge base. The physical meaning of taking the absolute value is that, regardless of whether the biomarker is a positive or negative indicator, its deviation from the normal value is considered as an input of risk intensity.

[0043] At the row and column intersection position corresponding to the socio-psychological feature node, the ratio is written into the connection weight storage unit, and the initial weight matrix is ​​output in memory.

[0044] Read the ratio (OR) of health literacy to frailty, which reflects the probability multiple of the outcome event occurring when the exposure factor is present. Locate the row and column intersection corresponding to the patient node and the psychosocial characteristic node, and update the element value at that position with the OR value. If the prior knowledge base stores logarithmically transformed data... If the value is zero, it is written directly; otherwise, the original OR value is written. The method of directly writing the original OR value is used to maintain the intuitiveness of the "isomorphic mapping". Finally, after the weights of all relevant nodes have been written, the original all-zero matrix in memory is transformed into a sparse matrix, i.e., the initialized weight matrix. This matrix mathematically characterizes the initial topological structure of the multidimensional heterogeneous association graph and the prior strength of each risk factor.

[0045] For example, suppose the data obtained from the prior knowledge base is as follows: the standardized mean difference (SMD) between homocysteine ​​and frailty is 0.82; the ratio of health literacy to frailty is 2.15. A write operation is performed, locking position (0,3) and changing the default value from 0.0 to 0.82; locking position (0,7) and changing the default value from 0.0 to 2.15; other positions in the matrix, such as (0,1) and (2,5), remain at 0.0. At this point, a value containing the above values ​​is output in memory. Initialize the weight matrix, which will be passed to the path topology parser for weighted correction.

[0046] Optionally, generating the target adjacency matrix includes: By traversing the topological structure through the multidimensional heterogeneous association graph, based on the socio-psychological feature nodes, the sleep disorder node pointing to the depressive mood node and then to the patient node is locked as a directed connection sequence and marked as a chain-mediated transmission path. The search starts at the "sleep disorder node" under the socio-psychological feature node category in a multidimensional heterogeneous association graph, and ends at the "patient node." Path constraints are set: the path length must be 2, and intermediate nodes must have the semantic label of "depressive mood" or "psychological distress." When the algorithm traverses a directed connection sequence that conforms to the topological structure "sleep disorder node -> depressive mood node > patient node," it locks the connection, assigns a unique path ID, and marks it in memory as a chain-like mediation path. This aims to identify potential causal chains where socio-psychological factors influence physical health through mediating variables.

[0047] The statistical regression coefficients of the chain-mediated transmission path are read from the prior knowledge base, and continuous multiplication is performed to calculate and instantiate them as coupling strength factors. Using the edge relationships in the locked chain-like mediation transmission path as an index, a query is initiated to the prior knowledge base. The prior knowledge base stores path coefficients derived from large-scale epidemiological surveys or meta-analyses, typically standardized regression coefficients. Extract the coefficients of each connection segment along the path, and calculate the coupling strength factor using the product method principle in mediation effect analysis. The calculation formula is as follows: , in, This represents the coupling strength factor, which is the risk intensity of sleep disorders indirectly causing patient deterioration by leading to depressive mood. This represents the statistical regression coefficient of the first path segment, specifically the coefficient of influence of sleep disorders on depressive mood. This value is derived from prior knowledge base information on "sleep disorders". The regression analysis results for "depression" typically range from [value range missing]. between. This represents the statistical regression coefficient of the second path segment, specifically the coefficient of influence of depressive mood on weakness. This value is derived from prior knowledge base information on "depression". The regression analysis results of "weakness". This represents a mathematical multiplication operation. According to statistical principles, the transmission strength of an intermediate path is equal to the product of the coefficients of each link in the path.

[0048] For example, suppose the locked chain-mediated transmission path is: sleep disorder -> depressive mood -> patient. The regression coefficient of sleep disorder leading to depressive mood is retrieved from the prior knowledge base. The regression coefficient of depressive mood leading to weakness Then perform continuous product calculation: The value 0.24 is instantiated as the coupling strength factor corresponding to this path and temporarily stored in the calculation cache.

[0049] Based on the initial weight matrix, the direct connection elements between the sleep disorder node and the patient weakness node are located, the coupling strength factor is accumulated to the direct connection elements, and the target adjacency matrix is ​​output.

[0050] To generate a target adjacency matrix that reflects the "total effect," the calculated indirect effects need to be superimposed on the direct effects. First, the row indices of the sleep disorder nodes in the initialized weight matrix are analyzed. And the column index of the patient node in the matrix. Use these two indices to locate the directly connected elements in the matrix. This element represents the risk weight of sleep disorders directly leading to weakness; the initial value may be derived from the SMD or OR values ​​of a univariate analysis. Subsequently, a weighted adjustment operation is performed, as shown in the following formula: , in, This represents the element value at the corresponding position in the corrected target adjacency matrix, which is the total risk transmission intensity after combining the direct path and the intermediate path. This represents the original direct connection weights stored in the initial weight matrix. This represents the calculated coupling strength factor. Through this weighted correction, the values ​​in the matrix are no longer merely static statistical correlations, but incorporate dynamic causal transmission logic. After traversing all identified paths and completing the correction, the final target adjacency matrix is ​​output. Figure 2 As shown, the risk weight of the sleep disorder node increased from 0.35 to 0.59 after being superimposed with the coupling strength factor transmitted through depressive mood. The direction and length of the arrow quantify the risk gain magnitude of this hidden path.

[0051] For example, suppose that in the initial weight matrix: the sleep disorder node is located in the 5th row; the patient node is located in the 1st column; and the location... The value of the directly connected element at the location is This represents the standardized risk value for how sleep disorders directly lead to weakness. The calculated coupling strength factor is read as follows: Perform cumulative correction: . The target adjacency matrix The position value has been updated to 0.59. This change means that although the direct risk of sleep disorder alone is only 0.35, considering its high likelihood of triggering depression, its actual overall pathogenic impact on frailty has increased to 0.59. The final output matrix containing this correction value is the target adjacency matrix.

[0052] Optionally, the calculation and instantiation into a coupling strength factor includes: Using the node connection relationship of the chain-like mediator transmission path as the index key, retrieve statistical values ​​from the prior knowledge base; First, a topological decomposition is performed on the identified chain-like mediation path, assuming that the path contains... If there are 10 nodes, the path is decomposed into 100 nodes. Each pair of adjacent nodes is connected. The start and end node IDs of each connection pair are extracted, and a combined string or tuple is constructed as the index key. This is then connected to a relational database table in the prior knowledge base, which stores numerous regression coefficients between pairs of variables derived from meta-analysis of historical medical literature. The generated index key is used to perform an exact match search in the "source node-target node" composite primary key column of the database, extracting the corresponding statistical values. If a connection segment has multiple records in the knowledge base, a single, definitive statistical value is returned according to the established rules.

[0053] For example, suppose the locked chain of mediation pathways is: [Node_Sleep_Disorder]->[Node_Depression]->[Node_Frailty], meaning sleep disorder points to depressive mood, which in turn points to frailty. This can be decomposed into two connections, generating the following two index keys: Index key A: "Source:Sleep_Disorder|Target:Depression"; Index key B: "Source:Depression|Target:Frailty". Using these two keys to search the prior knowledge base, a successful match is found with a statistical value of 0.42 for index key A and 0.55 for index key B.

[0054] Perform a floating-point multiplication operation on the statistical value and assign the result to the coupling strength factor.

[0055] After obtaining the statistical values ​​of each connection segment on the path, the floating-point arithmetic unit is invoked to perform continuous multiplication. The mathematical principle of this operation is based on the mediation effect calculation rule in structural equation modeling (SEM), that is, the total effect of an indirect path is equal to the product of the coefficients of all direct paths on that path, as shown in the following formula: , in, The coupling strength factor quantifies the indirect influence of a risk source transmitted to its endpoint through a specific intermediary chain. The larger the value, the more critical the intermediary path is in the pathogenesis mechanism. This represents the total number of connecting edges in a chain-like mediation path, for example, the number of edges in a path with 3 nodes. . Indicates the first in the path The statistical values ​​corresponding to each connection edge are the standardized regression coefficients retrieved from the prior knowledge base. This represents a cumulative multiplication operation. The final product result, after being truncated for precision, is assigned to a coupling strength factor variable instantiated in memory as a gain parameter for matrix correction.

[0056] For example, the retrieved statistical values ​​are as follows: and Perform floating-point multiplication: This means that the indirect impact of "sleep disorder" on "weakness" through the mediating factor of "depressive mood" is 0.2310. This value is assigned to the coupling strength factor object corresponding to this path, preparing to accumulate it into the direct connection weights of the target adjacency matrix.

[0057] Optionally, the output postoperative frailty risk prediction score includes: Health literacy scores are extracted based on the data from the aforementioned social psychological assessment scale, and then converted into global adjustment coefficients using a negative exponential function. The normalized health literacy score is read. Since health literacy is considered a patient's "resilience" or "protective ability" against disease risk, a higher score indicates a stronger suppression of the likelihood of deterioration. To reduce computational complexity and improve clinical interpretability, an inverse proportional function is preferred as an implementation of this negative exponential function characteristic. The global adjustment coefficient is calculated using the inverse proportional function, and the formula is as follows: , in, This represents the global adjustment coefficient, which is a "attenuation factor" in the risk transmission process, and its value typically ranges from [value range missing]. Between these values, the smaller the value, the stronger the inhibitory effect of the patient's health literacy on risk. This serves as a baseline constant, ensuring that the adjustment coefficient is 1 when health literacy is 0. This indicates a health literacy score. This parameter represents the literacy efficacy gain, used to adjust the sensitivity of health literacy to risk suppression. It is derived from regression fitting of historical data or expert experience setting. For example, a value of 1.0 indicates the standard suppression strength.

[0058] For example, suppose patient Zhang San's normalized health literacy score is Set the competency effectiveness gain parameter. Substitute into the formula to calculate: The value 0.5714 is stored as the global adjustment coefficient. This indicates that Zhang San's high health literacy can reduce the original pathological risk transmission intensity to approximately 57.14%.

[0059] The clinicopathological feature data, the comorbidity assessment data, and the serum biomarker data are vectorized and concatenated to construct a risk source activation signal vector, which is then mapped to the multidimensional heterogeneous association map. To drive the risk calculation in the atlas, the patient's current physiological state needs to be transformed into a computer-recognizable input vector. Clinical pathological feature data: For categorical variables, one-hot encoding or ordinal encoding is used to convert them into numerical values; comorbidity assessment data: The cumulative score of the Charlson Comorbidity Index (CCI) is used as the numerical value; Serum biomarker data: Normalized concentration values ​​are used. Following the index order of nodes in the atlas, the above effective numerical components are extracted, and sequential concatenation is performed to generate a vector with dimension [missing information]. The column vector is the risk source activation signal vector. Subsequently, the non-zero elements in this vector are mapped and assigned to the corresponding input nodes in the multidimensional heterogeneous correlation graph as the initial energy for risk propagation.

[0060] For example, suppose the graph has 5 nodes, indexed in the following order: [0: Patient, 1: Homocysteine, 2: Diabetes, 3: Hypertension, 4: TNM stage]. Patient Zhang San's data are: homocysteine ​​normalized value 0.8, has diabetes, no hypertension, and TNM stage normalized value 0.6. Then the constructed risk source activation signal vector... for: The first element is 0, and the remaining elements correspond to the intensity of each risk factor.

[0061] The target adjacency matrix is ​​modified by performing a scalar product operation using the global adjustment coefficient, and matrix multiplication is performed with the risk source excitation signal vector. The postoperative weakness risk prediction score is then output via activation function mapping.

[0062] Each element in the target adjacency matrix is ​​multiplied by a global adjustment coefficient to generate a new "adjusted matrix." Mathematically, this step is equivalent to reducing the risk transmission efficiency of all paths. Subsequently, matrix-vector multiplication is performed, multiplying the "adjusted matrix" by the "risk source trigger signal vector." According to linear algebra principles, this operation calculates the weighted sum of all input signals received by the patient node, i.e., the total risk impact value. Finally, to convert the variable-range impact value into a probability score between 0 and 1, a nonlinear mapping using the Sigmoid activation function is employed, calculated as follows: , in, This represents the final output score predicting postoperative weakness risk. This represents the total risk impact value of the patient node calculated by matrix multiplication. The bias term is used to calibrate the baseline risk rate of the model. It is the log odds of the base probability of a patient developing frailty when no risk factors are input, and is usually set based on historical prevalence. It is a natural constant.

[0063] For example, suppose that in the target adjacency matrix, the weight of homocysteine ​​pointing to the patient is 0.82. After correction by Zhang San's global adjustment coefficient (0.571), the effective weight becomes... Assuming we only consider the homocysteine ​​pathway, with an input signal of 0.8, then the risk shock value of polymerization is... Assuming bias term Substitute into the Sigmoid formula to calculate. The final output showed that Zhang San's postoperative weakness risk prediction score was 0.164, indicating that his probability of developing weakness after surgery was approximately 16.4%.

[0064] Optionally, the construction of the risk source excitation signal vector includes: Effective numerical components were extracted based on the clinicopathological feature data, the comorbidity assessment data, and the serum biomarker data. Numerical extraction operations are performed on different types of feature data to obtain effective numerical components that can be directly used in matrix operations. For serum biomarker data: the processed normalized concentration values ​​are directly read. For comorbidity assessment data: the cumulative total score of the Charlson Comorbidity Index (CCI) is directly extracted. For clinicopathological feature data: since the original data may be text labels, ordinal encoding or a risk weight mapping table is used to convert it into numerical values. For example, stage I is mapped to 1.0, stage II to 2.0, and so on, thereby quantifying the severity of pathological risk. After extraction, these values ​​are temporarily stored in a key-value pair list, where the key is the feature type ID and the value is the extracted floating-point number.

[0065] For example, suppose we are processing data from patient Li Si. Serum biomarkers: The normalized value of homocysteine ​​was read as 0.65. Comorbidity assessment: The Charlson Comorbidity Index score was read as 3.0. Clinicopathological features: The TNM stage was read as "Stage III", which was converted to a value of 3.0 according to the mapping rule. The extracted effective numerical component set is: {Hcy:0.65, CCI:3.0, TNM:3.0}.

[0066] The effective numerical components are serially spliced ​​end to end to generate a risk source excitation signal vector.

[0067] First, obtain the node index mapping table of the multidimensional heterogeneous association graph to determine the arrangement order and total number of all nodes in the graph. Then, initialize a string of length... The process involves creating a zero-based column vector. It iterates through the extracted effective numerical components and, based on the correspondence between feature types and graph nodes, fills the corresponding index positions in the zero-based column vector with the values. For nodes in the graph that exist but have no corresponding input data, their vector elements are kept at 0. Logically, this process is equivalent to sequentially concatenating the scattered numerical components according to their node index order, resulting in the vector representing the risk source's excitation signal. Mathematically, this vector represents the graph's... The initial state triggered by external risks at all times.

[0068] For example, suppose the multidimensional heterogeneous association graph has 5 nodes, and the internally maintained node index order is as follows: Index 0: Patient node, Index 1: Homocysteine ​​node, Index 2: Charlson comorbidity node, Index 3: TNM stage node, Index 4: Depressive mood node. Continuing with patient Li Si's data (Hcy: 0.65, CCI: 3.0, TNM: 3.0), a splicing operation is performed: Position 0: No input, enter 0.0; Position 1: Enter the homocysteine ​​value 0.65; Position 2: Enter the comorbidity score 3.0; Position 3: Enter the stage value 3.0; Position 4: No direct input, enter 0.0. The final generated risk source triggering signal vector... The vector will be fed into the matrix multiplication module and operated on with the target adjacency matrix after global adjustment coefficient correction.

[0069] Optionally, the generation of risk attribution analysis data includes: Based on the target adjacency matrix, with the patient node as the backtracking starting point, a maximum weight subgraph search is performed, and the node connection sequence with the highest accumulated weight value is locked. The process begins by identifying the column vector representing the patient node in the target adjacency matrix. Assuming the column index represents the target node and the row index represents the source node, the process iterates through all elements in this column vector, searching for the largest non-zero element. This element represents the direct predecessor node that contributes the most to the patient's deterioration risk. The row index of this direct predecessor node is recorded. The process then jumps to the column corresponding to that index and searches again for the largest non-zero element in the new column vector, finding the next higher-level pathogenic node. This backtracking process of "locating the maximum value - jumping to the column index - recording the node" is repeated until the backtracked node is the root node with an in-degree of zero, or the path length reaches a preset threshold, such as 3 levels. Finally, all node IDs recorded during the backtracking process are rearranged according to time logic, locking them into a node connection sequence. Mathematically, this sequence constitutes the backbone path of the maximum weight subgraph in the target adjacency matrix.

[0070] For example, suppose the target adjacency matrix is The matrix has node indices of [0: Patient, 1: Homocysteine, 2: Diabetes, 3: Depressive Mood, 4: Sleep Disorder]. Step 1: Backtracking: Check column 0 corresponding to the "Patient" node. The weight value of row 3 is found to be 0.59, which is the maximum value in that column. Record the predecessor node as "Depressive Mood". Step 2: Backtracking: Jump to column 3 corresponding to "Depressive Mood". The weight value of row 4 is found to be 0.42, which is the maximum value in that column. Record the predecessor node as "Sleep Disorder". Termination Check: Check column 4 corresponding to "Sleep Disorder" and find all values ​​are 0, so the search terminates. The final node connection sequence is: [Sleep Disorder] -> [Depressive Mood] -> [Patient].

[0071] The node connection sequences are marked as key pathogenic pathways, and combinations of comorbid nodes and mediating feature nodes are identified; The locked node connection sequence is marked as a critical pathogenic path in memory. Then, each node object in this path, except for the patient node, is traversed, and its metadata's "label type" attribute is read. If the node label is "Comorbidity," it is identified and categorized as part of a comorbidity node group. If the node label is "Psychosocial" or "Biomarker," and the node is located in the middle of the path, it is identified as a mediator node. The identified node IDs and their role labels are stored in a structured dictionary or hash map for encapsulation.

[0072] For example, in the sequence [Sleep Disorder] -> [Depressive Mood] -> [Patient], querying node attributes reveals: Node "Sleep Disorder": labeled Psychosocial and located at the beginning of the path, it is marked as a risk source. Node "Depressive Mood": labeled Psychosocial and located in the middle of the path, it is identified as a mediating node. Node "Patient": marked as an ending node. If the path is [Diabetes] -> [Renal Failure] -> [Patient], then "Diabetes" is identified as a comorbidity node combination, and "Renal Failure" is identified as a mediating node.

[0073] Based on the key pathogenic pathways, cumulative weight values ​​are extracted and combined with the co-pathogenic nodes and the mediating feature nodes to perform structured encapsulation, generating risk attribution analysis data.

[0074] The cumulative weight of key pathogenic pathways is calculated, which is the weighted sum of the weights of all connecting edges along the path or the maximum single-edge weight. Then, a standardized data object, such as a JSON object, is constructed, structurally encapsulating the path ID, cumulative weight value, identified comorbidity name, and mediator name. This data object is the risk attribution analysis data, which can be directly parsed by the front-end visualization component for drawing a highlighted pathogenic pathway diagram. Figure 3 As shown, the marginal contribution of each risk source and mediating pathway to the patient's final frailty risk score was quantitatively analyzed. Although homocysteine ​​and comorbidity provided the basic risk, the key pathogenic pathway of "sleep disorder -> depressive mood" contributed the largest risk increment, while health literacy, as a moderating coefficient, produced a significant risk offsetting effect.

[0075] Based on the same inventive concept, this invention also provides a system for analyzing the risk of postoperative debilitation in gastrointestinal tumor patients by integrating comorbidity maps, such as... Figure 4 As shown, the system includes: The atlas construction module is used to obtain clinicopathological feature data, comorbidity assessment data, serum biomarker data, and psychosocial assessment scale data for patients with gastrointestinal tumors. The psychosocial assessment scale data is normalized and a multidimensional heterogeneous association atlas is constructed. The multidimensional heterogeneous association atlas includes patient nodes, comorbidity nodes, biomarker nodes, and psychosocial feature nodes. The prior knowledge base module is used to acquire medical historical literature data, extract features to construct a prior knowledge base, and obtain the standardized mean difference between homocysteine ​​and frailty, as well as the ratio of health literacy to frailty through the prior knowledge base. The weight mapping module is used to perform statistical effect value isomorphism mapping on the multidimensional heterogeneous association map, map the standardized mean difference to the connection weight of the biomarker node, map the ratio to the connection weight of the sociopsychological feature node, and generate an initial weight matrix. The adjacency generation module is used to perform path topology parsing on the multidimensional heterogeneous association graph, identify chain-like mediator transmission paths, generate coupling strength factors based on the chain-like mediator transmission paths, and use the coupling strength factors to perform weighted correction on the initial weight matrix to generate the target adjacency matrix. The scoring calculation module is used to extract health literacy scores based on the social psychological assessment scale data to generate a global adjustment coefficient, input the clinical pathological feature data, the comorbidity assessment data and the serum biomarker data as risk source activation signals into the multidimensional heterogeneous association map, use the global adjustment coefficient to perform correction calculations on the target adjacency matrix, and output a postoperative frailty risk prediction score for the target patient. The attribution analysis module is used to perform a maximum weight subgraph search in the target adjacency matrix if the postoperative weakness risk prediction score exceeds a preset safety threshold, to lock the key pathogenic path and extract the combination of comorbid nodes and mediating feature nodes, and generate risk attribution analysis data.

[0076] It should be noted that the functional division and information interaction between the various modules described above are logical, but in terms of physical implementation, they can be integrated on the same software platform or deployed in a distributed manner. The connections between them represent data flow and control flow, aiming to collaboratively achieve the objectives of this invention. The above descriptions are merely exemplary embodiments of this invention and should not be construed as limiting the scope of protection of this invention.

Claims

1. A method for analyzing the association between postoperative debilitating risk and comorbidity profiles in gastrointestinal tumor patients, characterized in that... The method includes: For patients with gastrointestinal tumors, clinical pathological feature data, comorbidity assessment data, serum biomarker data, and psychosocial assessment scale data are obtained. The psychosocial assessment scale data is normalized and a multidimensional heterogeneous association map is constructed. The multidimensional heterogeneous association map includes patient nodes, comorbidity nodes, biomarker nodes, and psychosocial feature nodes. Medical historical literature data is acquired and features are extracted to construct a prior knowledge base. The prior knowledge base is used to obtain the standardized mean difference between homocysteine ​​and frailty, as well as the ratio of health literacy to frailty. Perform statistical effect value isomorphism mapping on the multidimensional heterogeneous association map, map the standardized mean difference to the connection weight of the biomarker node, map the ratio to the connection weight of the sociopsychological feature node, and generate an initial weight matrix; For the multidimensional heterogeneous association graph, perform path topology parsing to identify chain-like mediation transmission paths, generate coupling strength factors based on the chain-like mediation transmission paths, and use the coupling strength factors to perform weighted correction on the initial weight matrix to generate the target adjacency matrix. Based on the social and psychological assessment scale data, health literacy scores are extracted to generate a global regulation coefficient. The clinical pathological feature data, the comorbidity assessment data, and the serum biomarker data are used as risk source activation signals and input into the multidimensional heterogeneous association map. The global regulation coefficient is used to perform correction calculations on the target adjacency matrix, and a postoperative frailty risk prediction score is output for the target patient. If the postoperative weakness risk prediction score exceeds the preset safety threshold, a maximum weight subgraph search is performed in the target adjacency matrix to lock the key pathogenic path and extract the combination of comorbid nodes and mediating feature nodes to generate risk attribution analysis data.

2. The method for postoperative debilitating risk association analysis in gastrointestinal tumor patients based on a fusion comorbidity map as described in claim 1, characterized in that, The construction of the multidimensional heterogeneous association map includes: It acquires electronic medical records and reads pathology examination records and laboratory reports, generates structured clinical pathological feature data and serum biomarker data, and reads past medical history records to generate comorbidity assessment data; The missing items of the social psychological assessment scale data are removed and the scores are summarized. The maximum and minimum values ​​are scaled to generate the social psychological assessment scale values. Patient nodes, comorbidity nodes, biomarker nodes, and sociopsychological feature nodes are instantiated and generated. A connection edge data table is established for the patient nodes pointing to the comorbidity nodes, the biomarker nodes, and the sociopsychological feature nodes to construct a multidimensional heterogeneous association graph.

3. The method for postoperative debilitating risk association analysis in gastrointestinal tumor patients based on fused comorbidity maps according to claim 1, characterized in that, The process of obtaining the standardized mean difference between homocysteine ​​and frailty, and the ratio of health literacy to frailty, through the prior knowledge base includes: Using homocysteine ​​and health literacy as exposure factors and frailty as an outcome variable, a literature retrieval instruction was constructed to identify and extract mean data, standard deviation data, and risk association data from the medical historical literature data. The mean data and the standard deviation data are converted to a unified unit of concentration measurement, and the risk-related data are aligned and corrected in the direction of effect size to construct a structured feature dataset. Perform statistical heterogeneity tests and pooled effect size calculations on the structured feature dataset, respectively aggregate and generate standardized mean difference and odds ratio, and output them to the query interface of the prior knowledge base.

4. The method for postoperative debilitating risk association analysis in gastrointestinal tumor patients based on a fusion comorbidity map as described in claim 1, characterized in that, The generation of the initial weight matrix includes: Based on the multidimensional heterogeneous association graph, an empty adjacency matrix with all zero values ​​is established, and the row and column intersection positions of the patient node, the biomarker node, and the sociopsychological feature node in the empty adjacency matrix are analyzed. At the row-column intersection position corresponding to the biomarker node, the standardized mean difference is written into the connection weight storage unit; At the row and column intersection position corresponding to the socio-psychological feature node, the ratio is written into the connection weight storage unit, and the initial weight matrix is ​​output in memory.

5. The method for postoperative debilitating risk association analysis in gastrointestinal tumor patients based on a fusion comorbidity map according to claim 1, characterized in that, The generation of the target adjacency matrix includes: By traversing the topological structure through the multidimensional heterogeneous association graph, based on the socio-psychological feature nodes, the sleep disorder node pointing to the depressive mood node and then to the patient node is locked as a directed connection sequence and marked as a chain-mediated transmission path. The statistical regression coefficients of the chain-mediated transmission path are read from the prior knowledge base, and continuous multiplication is performed to calculate and instantiate them as coupling strength factors. Based on the initial weight matrix, the direct connection elements between the sleep disorder node and the patient weakness node are located, the coupling strength factor is accumulated to the direct connection elements, and the target adjacency matrix is ​​output.

6. The method for postoperative debilitating risk association analysis in gastrointestinal tumor patients based on fused comorbidity maps according to claim 5, characterized in that, The calculation and instantiation into the coupling strength factor includes: Using the node connection relationship of the chain-like mediator transmission path as the index key, retrieve statistical values ​​from the prior knowledge base; Perform a floating-point multiplication operation on the statistical value and assign the result to the coupling strength factor.

7. The method for postoperative debilitating risk association analysis in gastrointestinal tumor patients based on a fusion comorbidity map according to claim 1, characterized in that, The output postoperative frailty risk prediction score includes: Health literacy scores are extracted based on the data from the aforementioned social psychological assessment scale, and then converted into global adjustment coefficients using a negative exponential function. The clinicopathological feature data, the comorbidity assessment data, and the serum biomarker data are vectorized and concatenated to construct a risk source activation signal vector, which is then mapped to the multidimensional heterogeneous association map. The target adjacency matrix is ​​modified by performing a scalar product operation using the global adjustment coefficient, and matrix multiplication is performed with the risk source excitation signal vector. The postoperative weakness risk prediction score is then output via activation function mapping.

8. The method for postoperative debilitating risk association analysis in gastrointestinal tumor patients based on a fusion comorbidity map according to claim 7, characterized in that, The constructed risk source excitation signal vector includes: Effective numerical components were extracted based on the clinicopathological feature data, the comorbidity assessment data, and the serum biomarker data. The effective numerical components are serially spliced ​​end to end to generate a risk source excitation signal vector.

9. The method for postoperative debilitating risk association analysis in gastrointestinal tumor patients based on a fusion comorbidity map according to claim 1, characterized in that, The generated risk attribution analysis data includes: Based on the target adjacency matrix, with the patient node as the backtracking starting point, a maximum weight subgraph search is performed, and the node connection sequence with the highest accumulated weight value is locked. The node connection sequences are marked as key pathogenic pathways, and combinations of comorbid nodes and mediating feature nodes are identified; Based on the key pathogenic pathways, cumulative weight values ​​are extracted and combined with the co-pathogenic nodes and the mediating feature nodes to perform structured encapsulation, generating risk attribution analysis data.

10. A system for analyzing the risk of postoperative debilitating gastrointestinal tumor patients by fusing comorbidity maps, applied to the method for analyzing the risk of postoperative debilitating gastrointestinal tumor patients by fusing comorbidity maps as described in any one of claims 1-9, characterized in that, The system includes: The atlas construction module is used to obtain clinicopathological feature data, comorbidity assessment data, serum biomarker data, and psychosocial assessment scale data for patients with gastrointestinal tumors. The psychosocial assessment scale data is normalized and a multidimensional heterogeneous association atlas is constructed. The multidimensional heterogeneous association atlas includes patient nodes, comorbidity nodes, biomarker nodes, and psychosocial feature nodes. The prior knowledge base module is used to acquire medical historical literature data, extract features to construct a prior knowledge base, and obtain the standardized mean difference between homocysteine ​​and frailty, as well as the ratio of health literacy to frailty through the prior knowledge base. The weight mapping module is used to perform statistical effect value isomorphism mapping on the multidimensional heterogeneous association map, map the standardized mean difference to the connection weight of the biomarker node, map the ratio to the connection weight of the sociopsychological feature node, and generate an initial weight matrix. The adjacency generation module is used to perform path topology parsing on the multidimensional heterogeneous association graph, identify chain-like mediator transmission paths, generate coupling strength factors based on the chain-like mediator transmission paths, and use the coupling strength factors to perform weighted correction on the initial weight matrix to generate the target adjacency matrix. The scoring calculation module is used to extract health literacy scores based on the social psychological assessment scale data to generate a global adjustment coefficient, input the clinical pathological feature data, the comorbidity assessment data and the serum biomarker data as risk source activation signals into the multidimensional heterogeneous association map, use the global adjustment coefficient to perform correction calculations on the target adjacency matrix, and output a postoperative frailty risk prediction score for the target patient. The attribution analysis module is used to perform a maximum weight subgraph search in the target adjacency matrix if the postoperative weakness risk prediction score exceeds a preset safety threshold, to lock the key pathogenic path and extract the combination of comorbid nodes and mediating feature nodes, and generate risk attribution analysis data.

Images