Intelligent prevention and control method and system for inpatient venous thromboembolism

Abstract

This invention relates to the field of intelligent medical data analysis technology, specifically to a method and system for the intelligent prevention and control of venous thromboembolism in hospitalized patients. The method involves acquiring static medical record features and determining target test indicators; determining the target time window through a closed-loop adaptive mechanism driven by the fitting uncertainty of a Gaussian process regression model; constructing a weighted bipartite graph and removing confounding factors based on the topological overlap of pathological pathways to extract pure thrombosis baseline scores and pure bleeding baseline scores; extracting the first and second derivatives of time-series data and concatenating them with the pure baseline scores to construct a dynamic feature vector; generating a baseline-driven adaptive boundary in a two-dimensional phase plane; constructing a short-range predicted trajectory based on Taylor expansion and solving for the collision time; performing multi-objective game arbitration based on the time priority principle to generate prevention and control strategies; and achieving forward-looking joint decision-making in a dynamic conflicting scenario of thrombosis and bleeding risks.

Description

Technical Field

[0001] This invention relates to the field of intelligent medical data analysis technology, specifically to a method and system for the intelligent prevention and control of venous thromboembolism in hospitalized patients throughout the entire process. Background Technology

[0002] Venous thromboembolism (VTE) is a common and serious complication in hospitalized patients, encompassing two main clinical manifestations: deep vein thrombosis (DVT) and pulmonary embolism (PE). Currently, a relatively mature technical system has been established for the assessment and prevention of VTE risk in hospitalized patients. This system primarily relies on static risk scoring scales (such as the Padua score and Caprini score) to stratify patients' risk. By extracting static medical record characteristics from patients and assigning scores, discrete risk levels are output, and corresponding preventive measures are formulated accordingly. These methods can quantitatively assess patients' thrombotic and bleeding risks separately based on known fixed risk factors upon admission, providing a fundamental stratification basis for clinical decision-making and playing a crucial role in the initial risk screening of routine hospitalized patients.

[0003] However, the coagulation function, inflammation level, and related laboratory indicators of hospitalized patients fluctuate continuously with the progress of treatment and changes in their condition. Thrombosis risk and bleeding risk exhibit dynamic evolution over time, and information coupling occurs between them at the scoring level due to the sharing of some risk factors. Existing technologies treat risk assessment as a static numerical output problem based on fixed time points, lacking the ability to dynamically perceive the continuous changing trends of laboratory indicators and their evolutionary acceleration. They also lack a mechanism to quantitatively separate the redundant contributions of common risk factors between the two risk dimensions. Consequently, the system cannot proactively predict the future evolution of thrombosis risk and bleeding risk based on independent, purely quantitative risk assessment. In complex clinical scenarios where both risks are at high levels simultaneously, it is difficult to provide objective and quantitative joint decision-making basis, thus failing to meet the clinical needs of hospitalized patients for continuous tracking and timely response to the evolutionary trends of VTE over time. Summary of the Invention

[0004] To address the problems in related technologies, this invention provides a method and system for the intelligent prevention and control of venous thromboembolism in hospitalized patients throughout the entire process, thereby overcoming the aforementioned technical problems existing in the existing related technologies.

[0005] To address the aforementioned technical problems, this invention provides the following technical solution: a method for the intelligent prevention and control of venous thromboembolism in hospitalized patients, comprising the following steps:

[0006] Obtain static medical record characteristics of hospitalized patients and determine target test indicators;

[0007] The target time window is determined based on the biological half-life of the target test index, and the measured value of the target test index within the target time window is obtained to obtain the target time series data.

[0008] Based on the static medical record features, a weighted bipartite graph of venous thrombosis features and bleeding features is constructed. Within the bipartite graph topological space, confounding factors are removed by cross-graph correlation degree calculation, and pure thrombosis baseline score and pure bleeding baseline score are extracted.

[0009] The target time series data is subjected to regression processing to extract the first and second derivatives; and the first and second derivatives are used as time series evolution features, which are then concatenated across dimensions with the pure thrombosis baseline score and the pure bleeding baseline score to construct a dynamic feature vector.

[0010] In a two-dimensional phase plane, the pure thrombosis baseline score and the pure bleeding baseline score are used as nonlinear driving parameters to dynamically generate adaptive thrombosis risk boundaries and adaptive bleeding contraindication boundaries that vary with the baseline scores.

[0011] The dynamic feature vector is mapped onto the two-dimensional phase plane to generate a short-range predicted trajectory. The first predicted collision time when the short-range predicted trajectory breaks through the adaptive thrombosis risk boundary and the second predicted collision time when it breaks through the adaptive bleeding contraindication boundary are calculated respectively.

[0012] The first predicted collision time is compared with the second predicted collision time, and a multi-objective game arbitration is performed in combination with a preset safety time threshold. Based on the arbitration result, a full-process intelligent prevention and control strategy is generated and output.

[0013] This invention also includes an intelligent prevention and control system for venous thromboembolism in hospitalized patients, comprising:

[0014] The data acquisition module is used to acquire static medical record characteristics of hospitalized patients and determine target test indicators;

[0015] The time series data determination module is used to determine a target time window based on the biological half-life of the target test index, and to obtain the measurement value of the target test index within the target time window to obtain target time series data.

[0016] The bipartite graph static decoupling module is used to construct a weighted bipartite graph of venous thrombosis features and bleeding features based on the static medical record features. Within the bipartite graph topological space, confounding factors are removed by cross-graph correlation calculation, and pure thrombosis baseline score and pure bleeding baseline score are extracted.

[0017] The time-series dynamic vectorization module is used to perform regression processing on the target time-series data to extract the first and second derivatives; and uses the first and second derivatives as time-series evolution features, and concatenates them with the pure thrombosis baseline score and the pure bleeding baseline score across dimensions to construct a dynamic feature vector.

[0018] The phase plane collision prediction module is used to dynamically generate adaptive thrombosis risk boundaries and adaptive bleeding contraindication boundaries that vary with the baseline scores in a two-dimensional phase plane, using the pure thrombosis baseline score and the pure bleeding baseline score as nonlinear driving parameters.

[0019] The prediction and extrapolation module is used to map the dynamic feature vector onto the two-dimensional phase plane to generate a short-range predicted trajectory, and to calculate the first predicted collision time when the short-range predicted trajectory breaks through the adaptive thrombosis risk boundary and the second predicted collision time when it breaks through the adaptive bleeding contraindication boundary.

[0020] The multi-objective game output module is used to compare the first predicted collision time with the second predicted collision time, and perform multi-objective game arbitration in combination with a preset safety time threshold. Based on the arbitration result, it generates and outputs a full-process intelligent prevention and control strategy.

[0021] By employing the above technical solution, the present invention provides a method and system for the intelligent prevention and control of venous thromboembolism in hospitalized patients throughout the entire process, which has at least the following beneficial effects:

[0022] 1. By introducing a closed-loop time window adaptive optimization mechanism, the data acquisition range is dynamically adjusted using the fitting uncertainty of the Gaussian process regression model as a feedback signal. This breaks away from the limitations of a fixed sampling period, enabling the acquisition window of time series data to adaptively match the biodynamic characteristics of different test indicators, thereby improving the quality and representativeness of time series data.

[0023] 2. By constructing a weighted bipartite graph and calculating the dynamic stripping coefficient based on topological overlap, quantitative decoupling of confounding factors in thrombosis risk and bleeding risk was achieved, avoiding baseline score bias caused by repeated weighting of confounding factors, so that the baseline scores of pure thrombosis and pure bleeding can more accurately reflect their respective independent pathological contributions.

[0024] 3. By cross-dimensionally splicing static baseline scores with dynamic temporal evolution features, and using baseline scores to drive the generation of adaptive safety boundaries in a two-dimensional phase plane, risk assessment is transformed from a static numerical comparison into a dynamic trajectory and boundary collision prediction problem, enabling the system to proactively predict risk evolution trends.

[0025] 4. By using the time priority principle to achieve multi-objective game arbitration, in scenarios where the risks of thrombosis and bleeding increase simultaneously, the priority direction of prevention and control strategies is automatically determined based on which risk will break through the safety boundary first, providing an objective and quantitative basis for joint decision-making. Attached Figure Description

[0026] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0027] Figure 1 A flowchart of the intelligent prevention and control method for venous thromboembolism in hospitalized patients provided by the present invention;

[0028] Figure 2 This is a schematic diagram of the modules of the intelligent prevention and control system for venous thromboembolism in hospitalized patients provided by the present invention. Detailed Implementation

[0029] 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, and 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.

[0030] Exemplary method:

[0031] In existing techniques for managing venous thromboembolism (VTE) risk in hospitalized patients, mainstream approaches primarily rely on static risk scoring scales (such as the Padua score and Caprini score) for one-time risk stratification. These methods calculate a static score based on fixed risk factors upon patient admission and determine preventative measures accordingly. However, hospitalized patients' coagulation function, inflammation levels, and related laboratory indicators (such as D-dimer concentration) fluctuate continuously with treatment progress and changes in their condition, and static scores cannot capture this dynamic evolution. Furthermore, thrombosis and bleeding risks often share some risk factors (such as advanced age and abnormal liver or kidney function). Existing methods, when assessing the two risks separately, repeatedly weight these shared factors, leading to varying degrees of overestimation or underestimation of both scores, making it difficult to accurately reflect the independent contribution of each risk. More critically, when both thrombosis and bleeding risks are at high levels simultaneously, existing technologies cannot provide an objective, quantitative joint decision-making framework to determine which risk is more urgent and should be prioritized.

[0032] If the above problems are not addressed, clinical prevention and control will face the following dilemmas: Prevention and control strategies based on static scores will become disconnected from the patient's real-time pathological status, potentially leading to thrombotic events in high-risk patients before scores are updated, or maintaining excessive anticoagulation even when the risk of thrombosis has decreased but the risk of bleeding is increasing, thus increasing the probability of bleeding events. In complex scenarios of "dual high risk" of thrombosis and bleeding, the lack of a systematic arbitration mechanism will make clinical decisions highly dependent on individual experience. Different doctors may make drastically different prevention and control choices for the same patient, affecting the consistency and traceability of prevention and control protocols.

[0033] In response, this application proposes a method for the intelligent prevention and control of venous thromboembolism in hospitalized patients throughout the entire process, such as... Figure 1 As shown, it includes:

[0034] S1. Obtain the static medical record characteristics of hospitalized patients and determine the target test indicators;

[0035] It should be noted that this step forms the data foundation layer of the entire prevention and control system. Its role is to completely extract all static baseline information related to VTE risk assessment from the patient's electronic medical record system and to clarify the target test indicators for subsequent time-series dynamic tracking. Static medical record characteristics refer to the patient's inherent attributes and clinical status information that do not change rapidly over time within a single assessment period, including but not limited to: age, gender, body mass index (BMI), type and duration of surgery, history and stage of malignant tumors, previous VTE history, current immobilization status (e.g., prolonged bed rest, lower limb plaster cast immobilization), central venous catheterization status, comorbid underlying diseases (e.g., heart failure, chronic kidney disease, liver dysfunction), and current medication regimen (e.g., type and dosage of anticoagulants, antiplatelet drugs, nonsteroidal anti-inflammatory drugs). These characteristics will serve as input nodes for the weighted bipartite graph in the subsequent S3 step, driving the removal of confounding factors and extraction of pure baseline scores. Target test indicators refer to quantitative laboratory tests used clinically to reflect the dynamic balance of the patient's coagulation-fibrinolysis system, typically represented by D-dimer concentration. As a fibrin degradation product, the sustained increase in D-dimer plasma concentration is a sensitive marker of enhanced thrombus formation and dissolution in vivo. Furthermore, target indicators can be expanded to include prothrombin time (PT), international normalized ratio (INR), activated partial thromboplastin time (APTT), fibrinogen (FIB), etc., with the specific selection determined by the clinical scenario and prevention objectives.

[0036] For example, in this embodiment, a 68-year-old male patient is used as the example throughout the text: This patient was admitted to the hospital for total hip replacement surgery due to a left femoral neck fracture. Postoperatively, he was in a bed-immobilization state, with a history of hypertension and mild renal insufficiency, and was currently taking aspirin orally. The system extracts the above-mentioned static medical record features from the electronic medical record and determines D-dimer as the target test indicator, whose biological half-life... It takes about 8 hours.

[0037] S2. Determine the target time window based on the biological half-life of the target test index, and obtain the measured value of the target test index within the target time window to obtain target time series data;

[0038] It should be noted that this step uses the fitting uncertainty of the first Gaussian process regression (GPR) model as the error signal, and dynamically adjusts the size of the data acquisition time window through an adaptive iterative expansion mechanism. This closed-loop time window adaptive optimization mechanism enables the system to automatically determine the required minimum sufficient time window for different test indicators and different patient data distributions, while meeting the model accuracy requirements. This avoids data sparsity and fitting distortion caused by an excessively small time window, and avoids computational redundancy and interference from outdated data caused by an excessively large time window. The biological half-life serves as the physical anchor point for the time window, ensuring the rationality of the initial search range in a biodynamic sense.

[0039] Specifically, the target time window is determined based on the biological half-life of the target test indicator, and the measured values ​​of the target test indicator within the target time window are obtained to acquire the target time series data, including:

[0040] S21. An initial time window is calculated by multiplying the initial search constant by the biological half-life of the target test index, and the measured values ​​within the initial time window are obtained as the first time-series data; initial time window The calculation formula is:

[0041] ,

[0042] in, This is the initial search constant. The biological half-life (in hours) of the target test indicator. Initial search constant. The setting follows basic biokinetic principles: within a biological half-life, the decay or change in the index concentration is approximately 50% of the initial value, typically taken as... to This means that the initial time window covers a period of 2 to 4 biological half-lives, theoretically encompassing the full dynamic range of indicator concentration decay from peak to 6.25% to 25%. The system then traces back from the current moment to the past. The time length is used to obtain the measurement values ​​of all available target inspection indicators within the time window, which are then arranged in chronological order to form the first time series data.

[0043] For example, in this embodiment, the biological half-life of the D-dimer h is the initial search constant. Then the initial time window h. The system backtracks D-dimer detection records from the past 24 hours, assuming the following four measurements are obtained as the first time-series data: .

[0044] S22. Input the first time series data into a preset first Gaussian process regression model for fitting to obtain the fitting uncertainty;

[0045] Specifically, the first Gaussian process regression (GPR) model performs Bayesian posterior inference on the first time series data, generating the posterior mean and posterior prediction variance of the trend of the test index within the time window. The GPR model is chosen for fitting because it not only outputs the mean prediction, but also simultaneously outputs the posterior prediction variance at each prediction time. This variance naturally quantifies the confidence level of the model's fit to the trend under the current data conditions.

[0046] The first GPR model uses a radial basis function (RBF) kernel as the covariance function, and its expression is as follows:

[0047] ,

[0048] in, and For any two time points, This refers to the signal variance (the fluctuation range of the control function value). The length scale parameter controls the rate of correlation decay between adjacent data points. The RBF kernel is chosen based on its infinitely differentiable smoothness, making it suitable for examining biodynamic scenarios where indicator concentrations change continuously over time. Model hyperparameters and Automatic optimization is achieved by maximizing the logarithmic marginal likelihood function, eliminating the need for manual pre-setting. In alternative implementations, the kernel function can also be a Matérn kernel or a rational quadratic kernel to suit different time-series data smoothness characteristics.

[0049] The fitting uncertainty extracted in this step Defined as the model in the current time window The mean of the posterior prediction variances at all prediction times is calculated using the following formula:

[0050] ,

[0051] in, This represents the total number of predicted times that are uniformly distributed within the time window. For the first The posterior prediction variance at each prediction time. When the number of data points is insufficient or the distribution is sparse, the posterior variance of the GPR model is too large. A high value indicates insufficient fit quality; conversely, when the data is sufficient and evenly distributed, The value is trending towards stability at a low level.

[0052] S23. When the fitting uncertainty is greater than the preset quality threshold, the initial search constant is increased by a preset step size to expand the initial time window, and the measurement value in the corresponding time window is re-acquired for fitting until the fitting uncertainty obtained by refitting is not greater than the preset quality threshold. The time window that satisfies the preset quality threshold is determined as the target time window.

[0053] It should be noted that this step will determine the fitting uncertainty. Compared with the preset quality threshold Comparison: If This indicates that the amount or distribution of data within the current time window is insufficient to support high-quality trend fitting, and the preset step size should be used. Increase the search constant, and indirectly expand the time window by multiplying the search constant by the biological half-life, i.e.:

[0054] ,

[0055] ,

[0056] in, For the first The search constant during round iteration, The preset search constant step size, Biological half-life (in hours) of the target test indicator. This is an indicator function that takes the value 1 when the condition within the parentheses is true, and 0 otherwise. This ensures that the expansion of the time window is always anchored to the biological half-life, with each expansion increment being [value missing]. For test indicators with different half-lives, the same The actual time expansion resulting from the step size automatically adapts to the biodynamic timescale of the metric, reflecting the adaptability of the search process. Specifically, the expansion operation is only performed when the fitting uncertainty exceeds a quality threshold; once... When the indicator function outputs 0, the search constant stops increasing, and the current... This is determined as the target time window.

[0057] The search constant step size is typically between 0.5 and 1.0. This range is set based on an engineering trade-off between search accuracy and iterative efficiency. Each additional unit expands the time window by a full biological half-life. Within this time span, the theoretical change in the concentration of the test indicator is approximately 50% of the current value. This means that the newly added data in each iteration has a significant increase in information. Below 0.5, each expansion only adds data within half a half-life, resulting in limited information gain but a doubling of iterations. This forces the GPR model to repeatedly perform posterior inference, disproportionately increasing computational overhead. A value greater than 1.0 means that each expansion spans more than one full half-life, which can easily lead to the inclusion of too much old data at once, skipping the minimum sufficient time window where the fitting quality just meets the standard, and introducing unnecessary historical noise.

[0058] Among them, the preset quality threshold The threshold setting depends on clinical accuracy requirements and indicator characteristics, typically ranging from 0.05 to 0.15 of the normalized posterior variance of the GPR. If the threshold is set too low, the system will continuously expand the time window until a large amount of outdated data is included, introducing historical noise; if the threshold is set too high, it may terminate the iteration prematurely due to insufficient data, leading to distorted trend fitting in subsequent iterations. Preset search constant step size. The step size is typically between 0.5 and 1.0, and its setting needs to balance search efficiency and accuracy: too large a step size may skip the optimal time window, while too small a step size will increase the number of iterations and computational overhead.

[0059] For example, in this embodiment, take , , h, initial search constant In the first iteration, the initial time window h, GPR fitting uncertainty for 4 data points The system will increase the search constant to The time window has been expanded to h, one new measurement value was added. In the second iteration, The search constant increased to The time window has been expanded to h, two new measurements were added. In the third iteration, Once the quality threshold is met, the search constant stops increasing, and 32 hours is determined as the target time window.

[0060] S24. Obtain the measured values ​​of the target inspection index within the target time window to obtain the target time series data.

[0061] Specifically, after the target time window is determined, the system finally obtains the measurement values ​​of all target test indicators within the window, forming the target time series data. This target time series data will be simultaneously supplied to the confounding factor removal process in step S3 (providing the time window length parameter) and the second GPR continuous trend fitting process in step S4 (providing training data).

[0062] S3. Based on the static medical record features, construct a weighted bipartite graph of venous thrombosis features and bleeding features. Within the bipartite graph topological space, calculate the cross-graph correlation degree to remove confounding factors and extract the pure thrombosis baseline score and the pure bleeding baseline score.

[0063] It should be noted that in clinical practice, many static medical record features (such as advanced age, malignant tumors, prolonged bed rest, obesity, etc.) simultaneously contribute significantly to both thrombosis and bleeding risks. Accumulating scores in both scoring systems leads to a spurious positive correlation between the baseline scores for thrombosis and bleeding. This introduces a systematic bias into the adaptive boundary generation driven by these two scores in step S5, ultimately causing the game-theoretic arbitration in step S7 to lose its decision-making basis. This step, through topological decoupling and pure baseline extraction, provides information-independent driving parameters for all subsequent steps.

[0064] Specifically, a weighted bipartite graph of venous thrombosis and bleeding features is constructed based on the static medical record features. Within the bipartite graph topological space, confounding factors are removed by calculating cross-graph correlation degree, and the pure thrombosis baseline score and pure bleeding baseline score are extracted, including:

[0065] S31. Map the static medical record features to feature nodes;

[0066] Specifically, the system maps each static medical record feature extracted in step S1 to an independent node in graph theory. Each feature node carries the clinical attribute information of that feature (such as feature type, feature value, feature source, etc.), forming a common node set of the weighted bipartite graph.

[0067] For example, in this embodiment, the static medical record features of the 68-year-old male patient are mapped to the following feature nodes: Age = 68 years, Gender = Male, BMI = 26.5, Surgical Type = Hip Replacement, Immobilization Status = Bedridden, Malignancy = None, Previous VTE = None, Hypertension = Yes, Renal Insufficiency = Mild, Current Medication = Aspirin There are a total of 10 feature nodes.

[0068] S32. Calculate the first contribution weight of each feature node to the risk of venous thrombosis and the second contribution weight to the risk of bleeding, in order to construct a weighted bipartite graph.

[0069] Specifically, the weighted bipartite graph consists of two risk category nodes (venous thrombosis risk node). Bleeding risk points The system consists of a set of common feature nodes. Based on clinical evidence-based medicine literature or standardized scoring scales (such as the contribution of each feature to thrombosis risk in the Caprini scale, or the contribution of each feature to bleeding risk in the HAS-BLED score), the system calculates the pointer to each feature node. First contribution weight and pointing Second contribution weight The larger the weight value, the stronger the contribution of that feature to the corresponding risk.

[0070] For example, in this embodiment, the feature node "age = 68 years old" (2 points corresponds to 60-74 years old on the Caprini scale). (1.5 points for those over 65 years old in HAS-BLED); Feature node "Surgery type = hip replacement" (5 points correspond to major orthopedic surgery on the Caprini scale). (Surgery has a relatively low direct contribution to bleeding). Two weighted edges connect each feature node to two risk category nodes, forming a complete weighted bipartite graph.

[0071] S33. Based on the first contribution weight and the second contribution weight, calculate the cross-graph correlation ratio for each feature node; cross-graph correlation ratio Quantified the first The degree of information overlap between two risk dimensions for each feature node; the calculation formula is as follows:

[0072] ,

[0073] in, .when When the value approaches 1, it indicates that the contribution weights of this feature node to the two risk dimensions are highly similar, exhibiting strong co-origin characteristics, and is a typical candidate for confounding factors; when When the value approaches 0, it indicates that the feature node contributes to almost only a single risk dimension, has high information purity, and does not require stripping.

[0074] For example, in this embodiment, the feature node "age = 68 years old" The feature node "Surgery Type = Hip Replacement" The feature node "Current medication = Aspirin" , , .

[0075] S34. When the cross-graph correlation ratio is greater than the preset confounding threshold, the corresponding feature node is determined to be a confounding factor.

[0076] Among them, the preset confounding threshold The threshold is typically set between 0.40 and 0.60. This threshold is based on the statistical definition of confounding factors: when the weighting of a feature's contribution across two risk dimensions exceeds 40% to 60%, its "double inclusion" in the scoring system is sufficient to significantly interfere with the independence of the baseline score. If the threshold is set too low (e.g., below 0.30), many features with only weak cross-contributions will be excessively stripped away, resulting in severe loss of baseline score information; if the threshold is set too high (e.g., above 0.70), truly strong confounding factors may be missed, leading to insufficient decoupling.

[0077] For example, in this embodiment, take The feature node "age = 68 years old" It was determined to be a confounding factor; the feature node "hypertension" was identified as such. The feature node "Surgery Type = Hip Replacement" was identified as a confounding factor. It was determined to be a non-confounding factor (pure thrombosis risk characteristic); the characteristic node "Current medication = Aspirin" It was determined to be a non-confounding factor (pure bleeding risk characteristic).

[0078] S35. The initial thrombosis score and the initial bleeding score are calculated based on the static medical record features; initial thrombosis score Compared with the initial bleeding score The weighted sum of the first and second contribution weights of all feature nodes is calculated using the following formula:

[0079] ,

[0080] ,

[0081] in, This represents the total number of feature nodes.

[0082] Exemplarily, in this embodiment, Score (total contribution of each feature to the risk of thrombosis). Score (total contribution of each characteristic to bleeding risk).

[0083] S36. For feature nodes identified as confounding factors, retrieve the pathological pathway sequences of venous thrombosis and bleeding respectively from the preset medical causal knowledge graph.

[0084] The pre-defined medical causal knowledge graph is a directed acyclic graph (DAG) that encodes the causal relationships between disease, symptoms, and mechanisms. Its nodes cover disease entities, pathological mechanisms, molecular targets, etc., and directed edges represent causal driving relationships. This knowledge graph is pre-constructed and stored in the system based on publicly available pathophysiological causal evidence in clinical practice guidelines, evidence-based medicine literature, and expert consensus. It is ready before the evaluation process is executed.

[0085] Specifically, this step involves retrieving the pathological pathway sequences from each confounding factor to the "venous thrombosis" endpoint node and the "bleeding event" endpoint node in the knowledge graph. That is, starting from the confounding factor, the path set of all reachable intermediate nodes to the endpoint node is traversed along the directed edges.

[0086] For example, in this embodiment, the pathological pathways of the confounding factor "age = 68 years old" in the knowledge graph are: the pathway pointing to thrombosis - "advanced age → vascular endothelial degeneration → enhanced activation of coagulation factors → thrombosis", the pathway pointing to bleeding - "advanced age → increased vascular fragility → microvascular rupture → bleeding event" and "advanced age → decreased liver metabolic function → accumulation of anticoagulant drugs → bleeding event".

[0087] S37. Calculate the topological overlap between the pathological pathway sequences, and generate a dynamic stripping coefficient to characterize the independence weight based on the topological overlap.

[0088] It's important to note that topological overlap refers to the degree to which a confounding factor simultaneously drives both thrombotic and bleeding risks through the same pathological mechanism chain. If the intermediate nodes of two pathological pathways highly overlap (high topological overlap), it indicates that the factor's bidirectional contribution originates from a bifurcation of the same pathological mechanism, resulting in high information redundancy and requiring significant disentanglement. Conversely, if the two pathways have almost no overlapping intermediate nodes (low topological overlap), it suggests that although the same factor appears in both scores, their driving mechanisms are completely independent, requiring less disentanglement. Topological overlap The calculation formula is as follows:

[0089] ,

[0090] in, For the first A set of pathway nodes pointing to the endpoint of a thrombus, containing confounding factors. Let be the set of path nodes pointing to the end of the bleeding. Indicates the number of elements in the set. The calculation of topological overlap is essentially an application of Jaccard similarity to causal knowledge graph pathways.

[0091] Furthermore, based on topological overlap, the dynamic stripping coefficient The formula for generating it is as follows:

[0092] ,

[0093] in, Characterizing the first Independence weights of each confounding factor: The smaller the value (i.e., the higher the topological overlap), the more homogeneous the bidirectional contributions of the factor are, and the greater the redundancy deduction should be. The larger the value (i.e., the lower the topological overlap), the more likely the factor is to affect two dimensions simultaneously, but their driving paths are independent, and the smaller the redundancy deduction should be.

[0094] For example, in this embodiment, the set of thrombotic pathway nodes with the confounding factor "age = 68 years old" is... Endothelial degeneration and enhanced activation of coagulation factors. Bleeding pathway node set Increased vascular fragility, microvascular rupture, decreased liver metabolic function, and accumulation of anticoagulant drugs. The intersection of the two sets is an empty set (the intermediate mechanism nodes of the two pathological pathways do not overlap at all), and the union has 6 nodes. , This indicates that while advanced age affects both thrombosis and bleeding, the driving mechanisms are completely independent and should not be redundantly deducted. The thrombotic pathway of the confounding factor "hypertension". Endothelial injury, platelet activation bleeding pathway Endothelial damage and weakened vessel walls The intersection is Endothelial injury , and set as Endothelial damage, platelet activation, and weakened vessel walls Therefore , .

[0095] S38. Using the dynamic stripping coefficient, calculate the redundancy deduction scores of the confounding factors for thrombosis risk and bleeding risk, respectively, and subtract the corresponding redundancy deduction scores from the initial thrombosis score and the initial bleeding score to obtain the pure thrombosis baseline score and the pure bleeding baseline score. The calculation formula for the redundancy deduction score is as follows:

[0096] ,

[0097] ,

[0098] in, For the first The redundant contribution score that should be deducted from the thrombosis risk score for each confounding factor. For the first The redundant contribution score to be deducted from the bleeding risk score by each confounding factor. The combined effect of the two product factors ensures that the deduction is proportional to the degree of confounding of the factors and the degree of homology of the pathways.

[0099] Furthermore, the baseline score for pure thrombosis Baseline score compared to pure bleeding The calculation formula is:

[0100] ,

[0101] ,

[0102] in, This represents the set of feature nodes identified as confounding factors. Specifically, the pure baseline score after dynamic stripping eliminates redundant contributions caused by common pathological pathways, providing information-independent nonlinear driving parameters for adaptive boundary generation in the subsequent S5 step.

[0103] For example, in this embodiment, the confounding factor "age = 68 years old" is... Points (no deduction due to 0% topological overlap). Points. Confounding factor "hypertension" point, Points. Ultimately, point, point.

[0104] It should be noted that, in addition to the weighted bipartite graph and topological overlap methods mentioned above, there are alternative techniques for identifying and removing confounding factors: for example, statistical decoupling methods based on partial correlation analysis, causal inference methods based on structural equation modeling (SEM), or causal graph adjustment methods based on the backdoor criterion. This invention chooses a weighted bipartite graph combined with a medical causal knowledge graph because this approach simultaneously utilizes the statistical correlation strength of features (cross-graph correlation ratio) and domain knowledge of biomedical mechanisms (pathological pathway topological overlap), achieving a balance between decoupling accuracy and clinical interpretability.

[0105] S4. Perform regression processing on the target time series data to extract the first and second derivatives; and use the first and second derivatives as time series evolution features, and concatenate them with the pure thrombosis baseline score and the pure bleeding baseline score across dimensions to construct a dynamic feature vector.

[0106] It should be noted that this step transforms the discrete time-series data obtained in step S2 into physically meaningful kinematic parameters (position → velocity → acceleration) through GPR continuousization processing within step S4. This data is then fused across dimensions with the pure baseline score extracted in step S3 to construct a unified feature vector that includes both the static baseline base and the dynamic evolution rate. This vector serves as the complete input for phase plane boundary generation in step S5 and trajectory prediction in step S6, acting as a crucial bridge connecting static evaluation and dynamic prediction.

[0107] The construction of dynamic feature vectors includes the following steps:

[0108] S41. Based on the second Gaussian process regression model, the target time series data is subjected to regression processing and continuous trend fitting to generate a continuous mean function that characterizes the change of the target test index over time, and the posterior prediction variance output by the second Gaussian process regression model at the current time is extracted as the fitting confidence parameter.

[0109] Specifically, the second GPR model uses the same kernel function structure and hyperparameter optimization method as the first GPR model in step S22 (i.e., RBF kernel and automatic optimization of log-marginal likelihood, see step S22 for details). It performs Bayesian posterior inference on the target time series data obtained in step S2 to generate a continuous mean function. With posterior prediction variance Continuous mean function Interpolating discrete test index measurements into a continuous, smooth curve along the time axis ensures that subsequent differentiation operations have good mathematical definition. Fitting confidence parameters. Take the posterior prediction standard deviation at the current moment, which physically represents the uncertainty of the model's trend estimation at the current moment.

[0110] It should be noted that the second GPR model and the first GPR model in step S22 have a clear division of labor: the responsibility of the first GPR model is to evaluate whether the time window is "large enough" (outputting uncertainty as a closed-loop feedback signal), while the responsibility of the second GPR model is to perform fine continuous trend fitting within the determined target time window (outputting the mean function and variance for subsequent differentiation and vectorization).

[0111] For example, in this embodiment, the second GPR model performs posterior inference on seven D-dimer measurements within a 32-hour target time window and fits a continuous mean function. At the current moment place, , .

[0112] S42. Calculate the time derivative of the continuous mean function to extract the first and second derivatives at the current time step; the formulas for calculating the first and second derivatives are as follows:

[0113] ,

[0114] ,

[0115] in, The first derivative, in physical terms, represents the rate of deterioration of the test index at the current moment (unit: A positive value indicates that the indicator is rising (deteriorating), and a negative value indicates that the indicator is falling (improving). It is the second derivative, and its physical meaning is to worsen the acceleration (unit: Positive values ​​indicate that the rate of deterioration is accelerating (accelerating deterioration), while negative values ​​indicate that the rate of deterioration is slowing down (decelerating or converging). The core advantage of the GPR model lies in its mean function. The posterior distribution of derivatives of any order can be obtained directly by analytically differentiating the kernel function, without the need for numerical difference approximation, thus mathematically guaranteeing the accuracy of differentiation.

[0116] For example, in this embodiment, at the current time place, (D-dimer is rising). (The rate of increase is accelerating), indicating that the patient's coagulation-fibrinolysis system is deteriorating rapidly.

[0117] S43. The pure thrombosis baseline score, the first derivative, the second derivative, and the fitting confidence parameter are concatenated and spliced ​​to generate a thrombosis evolution sub-vector; the thrombosis evolution sub-vector Its composition is as follows: This subvector integrates the pure thrombosis baseline score (representing the static risk base) extracted in step S3 with the dynamic kinematic parameters (representing the current evolutionary state) extracted in this step into a unified mathematical object, where the physical meanings of each component are as follows: The static base height is the risk factor for thrombosis. Given the current rate of deterioration, To worsen the trend, The confidence level for the trend estimate.

[0118] S44. Concatenate the pure bleeding baseline score, the first derivative, the second derivative, and the fitting confidence parameter to generate a bleeding evolution sub-vector; the bleeding evolution sub-vector Its composition is as follows: ,

[0119] It should be noted that the thrombosis evolution subvector and the hemorrhage evolution subvector share the same first derivative. Second derivative With fit confidence parameters This is because the two sub-vectors track the temporal evolution of the same target test indicator (such as D-dimer), differing only in the static baseline scores they are paired with. An increase in D-dimer can reflect both enhanced thrombotic activity and, indirectly, compensatory activation of the post-thrombotic fibrinolytic system. Therefore, the same set of dynamic parameters has different clinical interpretations in the two risk dimensions. This difference in interpretation will be reflected in the adaptive boundary of step S5 through different baseline driving parameters.

[0120] S45. Merge the thrombosis evolution sub-vector and the bleeding evolution sub-vector to generate a dynamic feature vector.

[0121] Specifically, the dynamic feature vector is a seven-dimensional, cross-dimensional evolution vector that fully encodes the patient's current "static base + dynamic evolution" panoramic information. In step S6, each component of this vector is mapped onto a two-dimensional phase plane, serving as the initial conditions and dynamic parameters for trajectory prediction. Dynamic Feature Vector The complete structure is as follows:

[0122] ,

[0123] For example, in this embodiment, the dynamic feature vector is: .

[0124] S5. In the two-dimensional phase plane, using the pure thrombosis baseline score and the pure bleeding baseline score as nonlinear driving parameters, dynamically generate adaptive thrombosis risk boundaries and adaptive bleeding contraindication boundaries that change with the baseline scores.

[0125] It should be noted that this step establishes a two-dimensional abstract space with the rate of deterioration as the coordinate axis, and defines two absolute taboo boundaries in this space. These two boundaries are not fixed, but are dynamically driven by the pure baseline score extracted in step S3—the higher the baseline score, the closer the corresponding boundary shrinks towards the origin of the coordinate axis, meaning that the safety tolerance of this dimension is smaller and the fault tolerance space of the system is narrower. This mechanism has intuitive rationale in clinical practice: a patient with an extremely high baseline risk of thrombosis should be considered to be in a critical danger state even if the rate of deterioration of their test indicators is not yet significant, and the shrinking of the boundary is precisely a spatial expression of this high baseline risk.

[0126] The dynamic generation of adaptive thrombotic risk boundaries and adaptive bleeding contraindication boundaries that change with baseline scores includes the following steps:

[0127] S51. Establish a two-dimensional phase plane with the rate of bleeding deterioration and the rate of thrombosis as coordinate axes;

[0128] Specifically, the two-dimensional phase plane is related to the rate of thrombosis progression. for Axis, rate of worsening bleeding for Axis. Every point in the phase plane. This represents the rate of risk evolution for a patient at a given moment. (Original coordinate point) This indicates that the rate of deterioration in both dimensions is zero, meaning the patient is in a state of risk quiescence. Areas far from the origin represent states where the risk is rapidly deteriorating.

[0129] S52. Input the pure thrombosis baseline score into a first preset boundary generation function to generate a nonlinear adaptive thrombosis risk boundary that shrinks towards the origin as the score increases; generate the adaptive thrombosis risk boundary. The first preset boundary generation function is as follows:

[0130] ,

[0131] in, This is the maximum distance parameter of the boundary (the asymptotic upper limit of the boundary when the baseline score is extremely low). This is the boundary contraction sensitivity coefficient. The midpoint of the contraction (the boundary distance is reduced to the value when the score reaches this value) This function is based on a pure thrombosis baseline score. Sigmoid decay function for the independent variable: when When increasing from low to high values from The near-terminal value monotonically decreases and approaches 0. Physically, this means that the higher the baseline thrombosis risk, the lower the upper limit of the thrombosis progression rate that the system can tolerate, and the boundary shrinks towards the origin.

[0132] It should be noted that, The value is typically set to 0.3 to 0.5 (normalized deterioration rate unit). Since the daily fluctuation limit of the deterioration rate of D-dimer in healthy hospitalized patients is statistically within the range of 0.3 to 0.5 under normalized coordinate system, the maximum distance parameter is set within this range so that the safety tolerance for patients with low baseline risk just covers normal physiological fluctuations without triggering false alarms. The value is typically between 0.5 and 1.0. This coefficient controls the steepness of the transition of the boundary from the maximum distance to the origin. If the value is below 0.5, the boundary contraction is too gentle, resulting in no significant difference in the safety tolerance between medium- and high-risk patients and low-risk patients, and the system's sensitivity to distinguishing changes in baseline scores is insufficient. If the value is above 1.0, the boundary collapses sharply in the medium score range, which can easily lead to an excessive narrowing of the safety tolerance for medium-risk patients and generate an allergic warning. The value is usually set to 8 to 12 points (corresponding to the intermediate to high risk boundary). This value corresponds to the clinical boundary interval for the transition from intermediate to high risk in the Caprini scoring system. Setting the midpoint of contraction here ensures that the most sensitive contraction response at the boundary occurs precisely at the critical boundary of the clinical risk level transition, thus aligning with the clinical stratification criteria.

[0133] For example, in this embodiment, take , , The patient Substitute into the formula:

[0134] This indicates that for this high-risk patient, a thrombus progression rate of approximately 0.137 (normalized units) is considered to have breached the thrombus risk threshold. However, if the patient... Only 5 points (low risk), substituting into the formula:

[0135] The boundary is far from the origin, providing a large margin for error.

[0136] S53. Input the pure bleeding baseline score into the second preset boundary generation function to generate a nonlinear adaptive bleeding taboo boundary that shrinks towards the origin as the score increases. Generate adaptive bleeding taboo boundary. The second preset boundary generation function is as follows:

[0137] ,

[0138] in, , as well as These represent the maximum distance controlling the bleeding boundary, contraction sensitivity, and contraction midpoint, respectively. Their physical meanings are consistent with the corresponding parameters in the thrombus boundary, but their specific values ​​differ. The value is typically between 0.25 and 0.45, slightly lower. The range of values ​​is determined because once a bleeding event occurs, the irreversibility of its clinical consequences is usually higher than that of a thrombotic event (thrombosis can be remedied by thrombolysis or thrombectomy, while the reversal window for serious bleeding events such as intracranial hemorrhage is extremely limited). Therefore, the system sets a more conservative safety tolerance for the bleeding dimension. The value is typically between 0.4 and 0.9, slightly lower. The range of values ​​is important because the total score range of the bleeding risk scoring system is usually narrower than that of the thrombosis scoring system (HAS-BLED typically has a maximum score of 9 points, while the Caprini scale can have a total score of over 20 points). Within a narrower scoring range, the same contraction sensitivity coefficient will result in a relatively more severe contraction at the boundary, therefore, it is necessary to appropriately reduce the value. To avoid excessive collapse of the bleeding boundary in the middle score range. Typically, a score of 6 to 10 is chosen, lower than the range for the contraction midpoint of the thrombosis boundary. This is because the clinical boundary between intermediate and high risk in the bleeding scoring system is lower than the corresponding boundary in the thrombosis scoring system (a score of 3 in the HAS-BLED score indicates high risk, while a score of 5 or higher in the Caprini score indicates high risk). Aligning the contraction midpoint with the clinical stratification criteria of their respective scoring systems ensures that the most sensitive contraction response occurs at the risk level transition zone of the corresponding scoring system. This design guarantees the symmetry of the two boundaries in terms of mathematical structure, providing a fair mathematical basis for game arbitration, while reflecting the inherent differences in clinical severity and scoring dimensions of the two risk dimensions through their independent parameter values.

[0139] S6. Map the dynamic feature vector onto the two-dimensional phase plane to generate a short-range predicted trajectory, and calculate the first predicted collision time when the short-range predicted trajectory breaks through the adaptive thrombosis risk boundary and the second predicted collision time when it breaks through the adaptive bleeding contraindication boundary.

[0140] It should be noted that this step does not treat the risk as a static value for threshold comparison, but rather as a moving point mass in the phase plane, following kinematic equations along a deterministic trajectory in phase space. The trajectory extrapolation equation constructed using a second-order Taylor expansion allows the system to utilize the current position extracted in step S4 (…). ),speed( ) and acceleration ( Using this as an initial condition, the future trajectory of the particle is predicted within a short time range, and the intersection time of this trajectory with the two adaptive boundaries generated in step S5 is solved. This allows for the prediction of when and which boundary will be breached before the risk actually worsens, thus achieving forward-looking early warning.

[0141] The calculation of the first predicted collision time when the short-range predicted trajectory breaks through the adaptive thrombus risk boundary and the second predicted collision time when it breaks through the adaptive bleeding contraindication boundary includes the following steps:

[0142] S61. The starting coordinate is the current state value in the dynamic feature vector.

[0143] Specifically, starting coordinates The dynamic feature vectors constructed in step S4 are taken from The normalized deterioration rate component in the phase plane. The current value corresponding to the rate of thrombus progression. The current value corresponding to the rate of worsening bleeding.

[0144] S62. Using the first and second derivatives in the dynamic feature vector, a trajectory extrapolation equation with a time parameter is constructed based on the second-order Taylor expansion, generating a short-range predicted trajectory with a curved shape in the two-dimensional phase plane; the trajectory extrapolation equation is constructed based on the classic second-order Taylor expansion and includes a look-ahead time parameter. ( (Unit: h)

[0145] ,

[0146] in, and These are the first derivative values ​​(velocities) of the rate of thrombosis and the rate of hemorrhage worsening at the current moment, respectively. and These are the second derivative values ​​(accelerations) of the rates of thrombosis and hemorrhage worsening at the current moment. A quadratic curve (parabolic shape) is plotted in the phase plane using the trajectory extrapolation equation, with its bending direction and curvature determined by the acceleration components.

[0147] S63. Solve the trajectory extrapolation equation and the first boundary equation corresponding to the adaptive thrombosis risk boundary simultaneously, and extract the positive minimum time solution as the first predicted collision time.

[0148] Specifically, extrapolating the trajectory equations The components and the thrombosis risk boundary equation are combined as follows:

[0149] ,

[0150] Solve the equation The positive minimum solution is used as the first predicted collision time. If the quadratic equation has no positive real roots, it means that under the current acceleration conditions, the trajectory will not break through the boundary. Pick .

[0151] For example, in this embodiment, take (Normalized value of current thrombus progression rate) / h, / h², Simultaneous equations Simplify to From the quadratic formula h. Therefore, h.

[0152] S64. Solve the trajectory extrapolation equation and the second boundary equation corresponding to the adaptive bleeding taboo boundary simultaneously, and extract the positive minimum time solution as the second predicted collision time.

[0153] Similarly, the extrapolation equation of the trajectory... The equations for the component and the contraindication boundary for bleeding are combined as follows:

[0154] ,

[0155] Solve the equation The positive minimum solution is used as the second predicted collision time. ;

[0156] For example, in this embodiment, take , / h, / h², Simultaneous equations Simplify to Seeking h. Therefore, h.

[0157] S7. Compare the first estimated collision time with the second estimated collision time, and perform multi-objective game arbitration in combination with a preset safety time threshold. Generate and output a full-process intelligent prevention and control strategy based on the arbitration result.

[0158] It should be noted that the purpose of this step is to determine, based on the time priority principle, which risk is more urgent and should be addressed first when the predicted trajectories of both thrombosis risk and bleeding risk are likely to reach their respective safety boundaries. A preset safety time threshold is established. The introduction of this feature serves as a safety time circuit breaker: if the collision time of one party is extremely long (approaching infinity), it indicates that the trajectory has degenerated into a static standstill or converged to a safe region, and the system determines that no intervention is needed in that direction.

[0159] If the first estimated collision time is less than the second estimated collision time, and the first estimated collision time is less than the preset safe time threshold, then the risk of venous thrombosis is determined to have dynamic priority, and a first prevention and control strategy for prioritizing the prevention of venous thrombosis is generated.

[0160] Specifically, the primary prevention and control strategy includes, but is not limited to: initiating drug-induced thrombosis prevention (such as subcutaneous injection of low molecular weight heparin), strengthening physical prevention measures (such as intermittent pneumatic compression pumps and elastic stockings), increasing the frequency of lower limb activities, and increasing the interval between D-dimer follow-up examinations.

[0161] Among them, the preset safety time threshold The typical range is 12 to 24 hours. This range is based on the time characteristics of clinical nursing rounds and the onset of interventions: preventive interventions for clinical VTE (such as low molecular weight heparin injection) usually take 4 to 6 hours to take effect, and the 12 to 24 hour safety window provides sufficient response and intervention time for the clinical team. If both collision times are greater than... The system outputs a command that "the current risk situation is stable, maintain the current prevention and control plan" without requiring intervention.

[0162] If the second estimated collision time is less than the first estimated collision time, and the second estimated collision time is less than the preset safe time threshold, then the bleeding risk is determined to have dynamic priority, and a second prevention and control strategy that prioritizes preventing bleeding is generated.

[0163] Specifically, the second prevention and control strategy includes, but is not limited to: reducing or suspending the dosage of anticoagulants, strengthening monitoring of bleeding sites, having hemostatic drugs (such as tranexamic acid) readily available, and avoiding invasive procedures.

[0164] It should be further noted that the system also has a built-in extreme value protection mechanism: if and All extremely short (all less than) And the difference between the two is less than the preset time resolution. (Typically, 0.5h to 1.0h is used to ensure that the difference in the order of the two collision times exceeds the computational noise introduced by the GPR fitting error and the sampling interval of the test index, thus making the arbitration conclusion of the time priority principle statistically reliable.) This indicates that the system is in a dual high-risk critical instability state, that is, the two risk dimensions almost simultaneously breach the boundary. At this time, simple priority ranking is insufficient to deal with the extreme situation. The system triggers the highest-level alarm mechanism, generates a dual-risk joint prevention and control strategy, marks it as an emergency state, and hands it over to the artificial clinical team for emergency intervention decision-making.

[0165] For example, in this embodiment, take h. Calculated in step S6. h, h. Due to and The system determined the risk of venous thrombosis to have dynamic priority and mapped it to generate the primary prevention strategy: immediate initiation of low molecular weight heparin prophylactic anticoagulation therapy, shortening the D-dimer retest interval to every 6 hours, and simultaneously initiating intermittent pneumatic pump physical prophylaxis. This strategy was issued 5.1 hours before the expected thrombus boundary was breached, allowing sufficient time for clinical intervention.

[0166] Exemplary system:

[0167] Please see Figure 2A fully intelligent prevention and control system for venous thromboembolism in hospitalized patients, comprising:

[0168] The data acquisition module is used to acquire static medical record characteristics of hospitalized patients and determine target test indicators;

[0169] The time series data determination module is used to determine a target time window based on the biological half-life of the target test index, and to obtain the measurement value of the target test index within the target time window to obtain target time series data.

[0170] The bipartite graph static decoupling module is used to construct a weighted bipartite graph of venous thrombosis features and bleeding features based on the static medical record features. Within the bipartite graph topological space, confounding factors are removed by cross-graph correlation calculation, and pure thrombosis baseline score and pure bleeding baseline score are extracted.

[0171] The time-series dynamic vectorization module is used to perform regression processing on the target time-series data to extract the first and second derivatives; and uses the first and second derivatives as time-series evolution features, and concatenates them with the pure thrombosis baseline score and the pure bleeding baseline score across dimensions to construct a dynamic feature vector.

[0172] The phase plane collision prediction module is used to dynamically generate adaptive thrombosis risk boundaries and adaptive bleeding contraindication boundaries that vary with the baseline scores in a two-dimensional phase plane, using the pure thrombosis baseline score and the pure bleeding baseline score as nonlinear driving parameters.

[0173] The prediction and extrapolation module is used to map the dynamic feature vector onto the two-dimensional phase plane to generate a short-range predicted trajectory, and to calculate the first predicted collision time when the short-range predicted trajectory breaks through the adaptive thrombosis risk boundary and the second predicted collision time when it breaks through the adaptive bleeding contraindication boundary.

[0174] The multi-objective game output module is used to compare the first predicted collision time with the second predicted collision time, and perform multi-objective game arbitration in combination with a preset safety time threshold. Based on the arbitration result, it generates and outputs a full-process intelligent prevention and control strategy.

[0175] Exemplary computer-readable medium:

[0176] Embodiments of this application may also be computer-readable storage media storing computer program instructions thereon, which, when executed by a processor, cause the processor to perform the steps described in the "Exemplary Methods" section above according to the various embodiments of this application.

[0177] Computer-readable storage media may take the form of any combination of one or more readable media. A readable medium may be a readable signal medium or a readable storage medium. A readable storage medium may, for example, include, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatuses, or devices, or any combination thereof. More specific examples of readable storage media (a non-exhaustive list) include: electrical connections having one or more wires, portable disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0178] The basic principles of this application have been described above with reference to specific embodiments. However, it should be noted that the advantages, benefits, and effects mentioned in this application are merely examples and not limitations, and should not be considered as essential features of each embodiment of this application. Furthermore, the specific details disclosed above are for illustrative and facilitative purposes only, and are not limitations. These details do not limit the application to the necessity of employing the aforementioned specific details for implementation.

[0179] The block diagrams of devices, apparatuses, devices, and systems involved in this application are merely illustrative examples and are not intended to require or imply that they must be connected, arranged, or configured in the manner shown in the block diagrams. As those skilled in the art will recognize, these devices, apparatuses, devices, and systems can be connected, arranged, and configured in any manner. Words such as “comprising,” “including,” “having,” etc., are open-ended terms meaning “including but not limited to,” and are used interchangeably with them. The terms “or” and “and” as used herein refer to the terms “and / or,” and are used interchangeably with them unless the context clearly indicates otherwise. The term “such as” as used herein refers to the phrase “such as but not limited to,” and is used interchangeably with it.

[0180] It should also be noted that in the apparatus, equipment, and methods of this application, the components or steps can be disassembled and / or recombined. These disassemblies and / or recombinations should be considered as equivalent solutions of this application.

[0181] The above description of the disclosed aspects is provided to enable any person skilled in the art to make or use this application. Various modifications to these aspects will be readily apparent to those skilled in the art, and the general principles defined herein can be applied to other aspects without departing from the scope of this application. Therefore, this application is not intended to be limited to the aspects shown herein, but rather to be accorded the widest scope consistent with the principles and novel features disclosed herein.

[0182] The above description has been given for purposes of illustration and description. Furthermore, this description is not intended to limit the embodiments of this application to the forms disclosed herein. Although numerous exemplary aspects and embodiments have been discussed above, those skilled in the art will recognize certain variations, modifications, alterations, additions, and sub-combinations thereof.

Claims

1. A method for intelligent prevention and control of venous thromboembolism in hospitalized patients throughout the whole process, characterized in that, Includes the following steps: Obtain static medical record characteristics of hospitalized patients and determine target test indicators; The target time window is determined based on the biological half-life of the target test index, and the measured value of the target test index within the target time window is obtained to obtain the target time series data. Based on the static medical record features, a weighted bipartite graph of venous thrombosis features and bleeding features is constructed. Within the bipartite graph topological space, confounding factors are removed by cross-graph correlation degree calculation, and pure thrombosis baseline score and pure bleeding baseline score are extracted. The target time series data is subjected to regression processing to extract the first and second derivatives; and the first and second derivatives are used as time series evolution features, which are then concatenated across dimensions with the pure thrombosis baseline score and the pure bleeding baseline score to construct a dynamic feature vector. In a two-dimensional phase plane, the pure thrombosis baseline score and the pure bleeding baseline score are used as nonlinear driving parameters to dynamically generate adaptive thrombosis risk boundaries and adaptive bleeding contraindication boundaries that vary with the baseline scores. The dynamic feature vector is mapped onto the two-dimensional phase plane to generate a short-range predicted trajectory. The first predicted collision time when the short-range predicted trajectory breaks through the adaptive thrombosis risk boundary and the second predicted collision time when it breaks through the adaptive bleeding contraindication boundary are calculated respectively. The first predicted collision time is compared with the second predicted collision time, and a multi-objective game arbitration is performed in combination with a preset safety time threshold. Based on the arbitration result, a full-process intelligent prevention and control strategy is generated and output.

2. The inpatient venous thromboembolism full-process intelligent prevention and control method according to claim 1, characterized in that, The process of determining the target time window based on the biological half-life of the target test index, and obtaining the measured value of the target test index within the target time window to obtain target time series data includes: An initial time window is calculated by multiplying the initial search constant by the biological half-life of the target test index, and the measured values ​​within the initial time window are obtained as the first time series data. The first time series data is input into a preset first Gaussian process regression model for fitting to obtain the fitting uncertainty; When the fitting uncertainty is greater than the preset quality threshold, the initial search constant is increased by a preset step size to expand the initial time window, and the measurement values ​​in the corresponding time window are re-acquired for fitting until the fitting uncertainty obtained by refitting is not greater than the preset quality threshold. The time window that satisfies the preset quality threshold is determined as the target time window. The measured values ​​of the target inspection index within the target time window are obtained to obtain the target time series data.

3. The intelligent prevention and control method for venous thromboembolism in hospitalized patients according to claim 1, characterized in that, The construction of a weighted bipartite graph based on the static medical record features, which combines venous thrombosis and bleeding features, and the calculation of confounding factors through cross-graph correlation within the bipartite graph topological space, includes: Map the static medical record features to feature nodes; Calculate the first contribution weight of each feature node to the risk of venous thrombosis and the second contribution weight to the risk of bleeding to construct a weighted bipartite graph; Based on the first contribution weight and the second contribution weight, calculate the cross-graph correlation ratio of each feature node; When the cross-graph correlation ratio is greater than a preset confounding threshold, the corresponding feature node is determined to be a confounding factor.

4. The intelligent prevention and control method for venous thromboembolism in hospitalized patients according to claim 3, characterized in that, The extraction of baseline scores for pure thrombosis and pure bleeding includes the following steps: The initial thrombosis score and initial bleeding score are calculated based on the static medical record features. For feature nodes identified as confounding factors, the pathological pathway sequences of venous thrombosis and hemorrhage are retrieved from the preset medical causal knowledge graph. Calculate the topological overlap between the pathological pathway sequences, and generate a dynamic stripping coefficient based on the topological overlap to characterize the independence weights; Using the dynamic stripping coefficient, the redundancy deduction scores of the confounding factors for thrombosis risk and bleeding risk are calculated respectively, and the corresponding redundancy deduction scores are subtracted from the initial thrombosis score and the initial bleeding score respectively to obtain the pure thrombosis baseline score and the pure bleeding baseline score.

5. The intelligent prevention and control method for venous thromboembolism in hospitalized patients according to claim 1, characterized in that, The construction of the dynamic feature vector includes the following steps: The target time series data is regressed and fitted with a continuous trend based on the second Gaussian process regression model to generate a continuous mean function that characterizes the change of the target test index over time. The posterior prediction variance output by the second Gaussian process regression model at the current time is extracted as the fitting confidence parameter. The continuous mean function is differentiated over time to extract the first and second derivatives at the current moment. The pure thrombosis baseline score, the first derivative, the second derivative, and the fitting confidence parameter are concatenated and spliced ​​to generate a thrombosis evolution sub-vector; The pure bleeding baseline score, the first derivative, the second derivative, and the fitting confidence parameter are concatenated and spliced ​​to generate the bleeding evolution sub-vector; The thrombosis evolution sub-vector and the bleeding evolution sub-vector are merged to generate a dynamic feature vector.

6. The intelligent prevention and control method for venous thromboembolism in hospitalized patients according to claim 1, characterized in that, The dynamic generation of adaptive thrombotic risk boundaries and adaptive bleeding contraindication boundaries that change with baseline scores includes the following steps: Establish a two-dimensional phase plane with the rate of hemorrhage worsening and the rate of thrombosis worsening as coordinate axes; The pure thrombosis baseline score is input into a first preset boundary generation function to generate a nonlinear adaptive thrombosis risk boundary that shrinks towards the origin of the coordinate system as the score increases. The pure bleeding baseline score is input into a second preset boundary generation function to generate a nonlinear adaptive bleeding taboo boundary that shrinks towards the origin of the coordinate system as the score increases.

7. The intelligent prevention and control method for venous thromboembolism in hospitalized patients according to claim 1, characterized in that, The calculation of the first predicted collision time when the short-range predicted trajectory breaks through the adaptive thrombus risk boundary and the second predicted collision time when it breaks through the adaptive bleeding contraindication boundary includes the following steps: The starting coordinates are the current state value in the dynamic feature vector. Using the first and second derivatives in the dynamic feature vector, a trajectory extrapolation equation with time parameters is constructed based on the second-order Taylor expansion, generating a short-range predicted trajectory with a curve shape in the two-dimensional phase plane. Solve the trajectory extrapolation equation and the first boundary equation corresponding to the adaptive thrombosis risk boundary simultaneously, and extract the positive minimum time solution as the first predicted collision time; Solve the trajectory extrapolation equation and the second boundary equation corresponding to the adaptive bleeding taboo boundary simultaneously, and extract the positive minimum time solution as the second predicted collision time.

8. The intelligent prevention and control method for venous thromboembolism in hospitalized patients according to claim 1, characterized in that, The process of generating a fully intelligent prevention and control strategy based on the arbitration result includes the following steps: If the first estimated collision time is less than the second estimated collision time, and the first estimated collision time is less than the preset safe time threshold, then the risk of venous thrombosis is determined to have dynamic priority, and a first prevention and control strategy for prioritizing the prevention of venous thrombosis is generated. If the second estimated collision time is less than the first estimated collision time, and the second estimated collision time is less than the preset safe time threshold, then the bleeding risk is determined to have dynamic priority, and a second prevention and control strategy that prioritizes preventing bleeding is generated.

9. A system for implementing the intelligent prevention and control method for venous thromboembolism in hospitalized patients according to any one of claims 1-8, characterized in that, include: The data acquisition module is used to acquire static medical record characteristics of hospitalized patients and determine target test indicators; The time series data determination module is used to determine a target time window based on the biological half-life of the target test index, and to obtain the measurement value of the target test index within the target time window to obtain target time series data. The bipartite graph static decoupling module is used to construct a weighted bipartite graph of venous thrombosis features and bleeding features based on the static medical record features. Within the bipartite graph topological space, confounding factors are removed by cross-graph correlation calculation, and pure thrombosis baseline score and pure bleeding baseline score are extracted. The time-series dynamic vectorization module is used to perform regression processing on the target time-series data to extract the first and second derivatives; and uses the first and second derivatives as time-series evolution features, and concatenates them with the pure thrombosis baseline score and the pure bleeding baseline score across dimensions to construct a dynamic feature vector. The phase plane collision prediction module is used to dynamically generate adaptive thrombosis risk boundaries and adaptive bleeding contraindication boundaries that vary with the baseline scores in a two-dimensional phase plane, using the pure thrombosis baseline score and the pure bleeding baseline score as nonlinear driving parameters. The prediction and extrapolation module is used to map the dynamic feature vector onto the two-dimensional phase plane to generate a short-range predicted trajectory, and to calculate the first predicted collision time when the short-range predicted trajectory breaks through the adaptive thrombosis risk boundary and the second predicted collision time when it breaks through the adaptive bleeding contraindication boundary. The multi-objective game output module is used to compare the first predicted collision time with the second predicted collision time, and perform multi-objective game arbitration in combination with a preset safety time threshold. Based on the arbitration result, it generates and outputs a full-process intelligent prevention and control strategy.

10. An electronic device comprising a memory and a processor, characterized in that: The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions, which, when executed by the processor, implement the steps of the method as described in any one of claims 1 to 8.

Images